TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images

Deep Learning-Powered CT-Less Multitracer Organ Segmentation From PET Images: A Solution for Unreliable CT Segmentation in PET/CT Imaging

PURPOSE

The common approach for organ segmentation in hybrid imaging relies on coregistered CT (CTAC) images. This method, however, presents several limitations in real clinical workflows where mismatch between PET and CT images are very common. Moreover, low-dose CTAC images have poor quality, thus challenging the segmentation task. Recent advances in CT-less PET imaging further highlight the necessity for an effective PET organ segmentation pipeline that does not rely on CT images. Therefore, the goal of this study was to develop a CT-less multitracer PET segmentation framework.

PATIENTS AND METHODS

We collected 2062 PET/CT images from multiple scanners. The patients were injected with either 18 F-FDG (1487) or 68 Ga-PSMA (575). PET/CT images with any kind of mismatch between PET and CT images were detected through visual assessment and excluded from our study. Multiple organs were delineated on CT components using previously trained in-house developed nnU-Net models. The segmentation masks were resampled to coregistered PET images and used to train 4 different deep learning models using different images as input, including noncorrected PET (PET-NC) and attenuation and scatter-corrected PET (PET-ASC) for 18 F-FDG (tasks 1 and 2, respectively using 22 organs) and PET-NC and PET-ASC for 68 Ga tracers (tasks 3 and 4, respectively, using 15 organs). The models' performance was evaluated in terms of Dice coefficient, Jaccard index, and segment volume difference.

RESULTS

The average Dice coefficient over all organs was 0.81 ± 0.15, 0.82 ± 0.14, 0.77 ± 0.17, and 0.79 ± 0.16 for tasks 1, 2, 3, and 4, respectively. PET-ASC models outperformed PET-NC models ( P < 0.05) for most of organs. The highest Dice values were achieved for the brain (0.93 to 0.96 in all 4 tasks), whereas the lowest values were achieved for small organs, such as the adrenal glands. The trained models showed robust performance on dynamic noisy images as well.

CONCLUSIONS

Deep learning models allow high-performance multiorgan segmentation for 2 popular PET tracers without the use of CT information. These models may tackle the limitations of using CT segmentation in PET/CT image quantification, kinetic modeling, radiomics analysis, dosimetry, or any other tasks that require organ segmentation masks.

SIRE: Scale-invariant, rotation-equivariant estimation of artery orientations using graph neural networks

The orientation of a blood vessel as visualized in 3D medical images is an important descriptor of its geometry that can be used for centerline extraction and subsequent segmentation, labeling, and visualization. Blood vessels appear at multiple scales and levels of tortuosity, and determining the exact orientation of a vessel is a challenging problem. Recent works have used 3D convolutional neural networks (CNNs) for this purpose, but CNNs are sensitive to variations in vessel size and orientation. We present SIRE: a scale-invariant rotation-equivariant estimator for local vessel orientation. SIRE is modular and has strongly generalizing properties due to symmetry preservations. SIRE consists of a gauge equivariant mesh CNN (GEM-CNN) that operates in parallel on multiple nested spherical meshes with different sizes. The features on each mesh are a projection of image intensities within the corresponding sphere. These features are intrinsic to the sphere and, in combination with the gauge equivariant properties of GEM-CNN, lead to SO(3) rotation equivariance. Approximate scale invariance is achieved by weight sharing and use of a symmetric maximum aggregation function to combine predictions at multiple scales. Hence, SIRE can be trained with arbitrarily oriented vessels with varying radii to generalize to vessels with a wide range of calibres and tortuosity. We demonstrate the efficacy of SIRE using three datasets containing vessels of varying scales; the vascular model repository (VMR), the ASOCA coronary artery set, and an in-house set of abdominal aortic aneurysms (AAAs). We embed SIRE in a centerline tracker which accurately tracks large calibre AAAs, regardless of the data SIRE is trained with. Moreover, a tracker can use SIRE to track small-calibre tortuous coronary arteries, even when trained only with large-calibre, non-tortuous AAAs. Additional experiments are performed to verify the rotational equivariant and scale invariant properties of SIRE. In conclusion, by incorporating SO(3) and scale symmetries, SIRE can be used to determine orientations of vessels outside of the training domain, offering a robust and data-efficient solution to geometric analysis of blood vessels in 3D medical images.

Automated proximal coronary artery calcium identification using artificial intelligence: advancing cardiovascular risk assessment

AIMS

Identification of proximal coronary artery calcium (CAC) may improve prediction of major adverse cardiac events (MACE) beyond the CAC score, particularly in patients with low CAC burden. We investigated whether the proximal CAC can be detected on gated cardiac CT and whether it provides prognostic significance with artificial intelligence (AI).

METHODS AND RESULTS

A total of 2016 asymptomatic adults with baseline CAC CT scans from a single site were followed up for MACE for 14 years. An AI algorithm to classify CAC into proximal or not was created using expert annotations of total and proximal CAC and AI-derived cardiac structures. The algorithm was evaluated for prognostic significance on AI-derived CAC segmentation. In 303 subjects with expert annotations, the classification of proximal vs. non-proximal CAC reached an area under receiver operating curve of 0.93 [95% confidence interval (CI) 0.91-0.95]. For prognostic evaluation, in an additional 588 subjects with mild AI-derived CAC scores (CAC score 1-99), the AI proximal involvement was associated with worse MACE-free survival (P = 0.008) and higher risk of MACE when adjusting for CAC score alone [hazard ratio (HR) 2.28, 95% CI 1.16-4.48, P = 0.02] or CAC score and clinical risk factors (HR 2.12, 95% CI 1.03-4.36, P = 0.04).

CONCLUSION

The AI algorithm could identify proximal CAC on CAC CT. The proximal location had modest prognostic significance in subjects with mild CAC scores. The AI identification of proximal CAC can be integrated into automatic CAC scoring and improves the risk prediction of CAC CT.

A deep learning method for total-body dynamic PET imaging with dual-time-window protocols

PURPOSE

Prolonged scanning durations are one of the primary barriers to the widespread clinical adoption of dynamic Positron Emission Tomography (PET). In this paper, we developed a deep learning algorithm that capable of predicting dynamic images from dual-time-window protocols, thereby shortening the scanning time.

METHODS

This study includes 70 patients (mean age ± standard deviation, 53.61 ± 13.53 years; 32 males) diagnosed with pulmonary nodules or breast nodules between 2022 to 2024. Each patient underwent a 65-min dynamic total-body [18F]FDG PET/CT scan. Acquisitions using early-stop protocols and dual-time-window protocols were simulated to reduce the scanning time. To predict the missing frames, we developed a bidirectional sequence-to-sequence model with attention mechanism (Bi-AT-Seq2Seq); and then compared the model with unidirectional or non-attentional models in terms of Mean Absolute Error (MAE), Bias, Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity (SSIM) of predicted frames. Furthermore, we reported the comparison of concordance correlation coefficient (CCC) of the kinetic parameters between the proposed method and traditional methods.

RESULTS

The Bi-AT-Seq2Seq significantly outperform unidirectional or non-attentional models in terms of MAE, Bias, PSNR, and SSIM. Using a dual-time-window protocol, which includes a 10-min early scan followed by a 5-min late scan, improves the four metrics of predicted dynamic images by 37.31%, 36.24%, 7.10%, and 0.014% respectively, compared to the early-stop protocol with a 15-min acquisition. The CCCs of tumor' kinetic parameters estimated with recovered full time-activity-curves (TACs) is higher than those with abbreviated TACs.

CONCLUSION

The proposed algorithm can accurately generate a complete dynamic acquisition (65 min) from dual-time-window protocols (10 + 5 min).

Planning transcatheter pulmonary valve implantation in the dysfunctional native RVOT: A semi-automated pipeline for dynamic analysis based on 4D-CT imaging

BACKGROUND AND OBJECTIVE

Dysfunction of the right ventricular outflow tract (RVOT) is a common long-term complication following surgical repair in patients with congenital heart disease. Transcatheter pulmonary valve implantation (TPVI) offers a viable alternative to surgical pulmonary valve replacement (SPVR) for treating pulmonary regurgitation but not all RVOT anatomies are suitable for TPVI. To identify a suitable landing zone (LZ) for TPVI, three-dimensional multiphase (4D) computed tomography (CT) is used to evaluate the size, shape, and dynamic behavior of the RVOT throughout the cardiac cycle. However, manually extracting measurements from multiplanar CT reformats is operator-dependent and time-consuming. Leveraging an optical-flow (OF) algorithm, we proposed a novel semi-automated pipeline for dynamic and comprehensive geometrical analysis of the RVOT anatomy.

METHODS

Upon 4D-CT availability, at a pre-defined reference time-point, the patient-specific anatomy is semi-automatically segmented to generate the corresponding three-dimensional surface, which is navigated through a graphical user interface to define the mid-section of the potential LZ. Based on the axial length of the intended device, the proximal and distal LZ cross-sections are automatically identified. An OF-based algorithm is used to track the three LZ cross-sections frame by frame throughout the cardiac cycle, taking RVOT out-of-plane motion into account to update RVOT contours on each cross-section and to elaborate LZ geometrical changes. Finally, LZ time-dependent geometrical features are quantified and extracted.

RESULTS

The pipeline was successfully applied to a retrospective cohort of patients, with OF-based tracking reporting excellent agreement (r2 = 0.99) compared to manual processing, with a bias < 1% for both LZ area and perimeter, while also significantly improving time efficiency. CT-derived measurements extracted from LZ mid-section were the most influential covariates affecting the likelihood of TPVI feasibility. Among these, the minimum perimeter outperformed all other geometric LZ parameters in classifying patients as suitable for either TPVI or SPVR and achieved the highest area under the curve of 0.99, with accuracy and precision of 0.93 and 0.92, respectively.

CONCLUSIONS

Dynamic OF-based quantification of key RVOT geometric parameters can enhance and expedite the selection process for TPVI candidates and guide optimal valve sizing during TPVI planning.

Application of machine learning methods for predicting esophageal variceal bleeding in patients with cirrhosis

OBJECTIVE

To develop and compare machine learning models based on CT morphology features, serum biomarkers, and basic physical conditions to predict esophageal variceal bleeding.

MATERIALS AND METHODS

Two hundred twenty-four cirrhotic patients with esophageal variceal bleeding and non-bleeding were included in the retrospective study. Clinical and serum biomarkers were used in our study. In addition, the open-access segmentation model was used to generate segmentation masks of the liver and spleen. Four machine learning models based on selected features are used for building prediction models, and the diagnostic performances of models were measured using the receiver operator characteristic analysis.

RESULTS

Two hundred twenty-four cirrhosis patients with esophageal varices, including 112 patients with bleeding (mean age 52.8 ± 11.5 years, range 18-80 years) and 112 patients with non-bleeding (mean age 57.3 ± 10.5 years, range 34-85 years). The two groups showed significant differences in standardized spleen volume, fibrinogen, alanine aminotransferase, aspartate aminotransferase, D-dimer, platelet, and age. The ratio of the training set to the test set was 8:2 in our research, and the 5-fold cross-validation was used in the research. The AUCs of linear regression, random forest, support vector machine, and adaptive boosting were, respectively, 0.742, 0.854, 0.719, and 0.821 in the training set. For the test set, the AUCs of models were, respectively, 0.763, 0.818, 0.648, and 0.804.

CONCLUSIONS

Our study used CT morphological measurements, serum biomarkers, and age to build machine learning models, and the random forest and adaptive boosting had potential added value in predictive model construction.

KEY POINTS

Question Esophageal variceal bleeding is an intractable complication of liver cirrhosis. Early prediction and prevention of esophageal variceal bleeding is important for patients with liver cirrhosis. Findings It was feasible and clinically meaningful to construct machine learning models based on CT morphology features, serum biomarkers, and physical conditions to predict variceal bleeding. Clinical relevance Our study may provide a promising tool with which clinicians can conduct therapeutic decisions on fewer invasive procedures for the prediction of esophageal variceal bleeding.

Post-Training Network Compression for 3D Medical Image Segmentation: Reducing Computational Efforts via Tucker Decomposition

Purpose To investigate whether the computational effort of three-dimensional CT-based multiorgan segmentation with TotalSegmentator can be reduced via Tucker decomposition-based network compression. Materials and Methods In this retrospective study, Tucker decomposition was applied to the convolutional kernels of the TotalSegmentator model, an nnU-Net model trained on a comprehensive CT dataset for automatic segmentation of 117 anatomic structures. The proposed approach reduced the floating-point operations and memory required during inference, offering an adjustable trade-off between computational efficiency and segmentation quality. This study used the publicly available TotalSegmentator dataset containing 1228 segmented CT scans and a test subset of 89 CT scans and used various downsampling factors to explore the relationship between model size, inference speed, and segmentation accuracy. Segmentation performance was evaluated using the Dice score. Results The application of Tucker decomposition to the TotalSegmentator model substantially reduced the model parameters and floating-point operations across various compression ratios, with limited loss in segmentation accuracy. Up to 88.17% of the model's parameters were removed, with no evidence of differences in performance compared with the original model for 113 of 117 classes after fine-tuning. Practical benefits varied across different graphics processing unit architectures, with more distinct speedups on less powerful hardware. Conclusion The study demonstrated that post hoc network compression via Tucker decomposition presents a viable strategy for reducing the computational demand of medical image segmentation models without substantially impacting model accuracy. Keywords: Deep Learning, Segmentation, Network Compression, Convolution, Tucker Decomposition Supplemental material is available for this article. © RSNA, 2025.

Opportunistic assessment of steatotic liver disease in lung cancer screening eligible individuals

BACKGROUND

Steatotic liver disease (SLD) is a potentially reversible condition but often goes unnoticed with the risk for end-stage liver disease.

PURPOSE

To opportunistically estimate SLD on lung screening chest computed tomography (CT) and investigate its prognostic value in heavy smokers participating in the National Lung Screening Trial (NLST).

MATERIAL AND METHODS

We used a deep learning model to segment the liver on non-contrast-enhanced chest CT scans of 19,774 NLST participants (age 61.4 ± 5.0 years; 41.2% female) at baseline and on the 1-year follow-up scan if no cancer was detected. SLD was defined as hepatic fat fraction (HFF) ≥5% derived from Hounsfield unit measures of the segmented liver. Participants with SLD were categorized as lean (body mass index [BMI] < 25 kg/m2) and overweight (BMI ≥ 25 kg/m2). The primary outcome was all-cause mortality. Cox proportional hazard regression assessed the association between (1) SLD and mortality at baseline and (2) the association between a change in HFF and mortality within 1 year.

RESULTS

There were 5.1% (1000/19,760) all-cause deaths over a median follow-up of 6 (range, 0.8-6) years. At baseline, SLD was associated with increased mortality in lean but not in overweight/obese participants as compared to participants without SLD (hazard ratio [HR] adjusted for risk factors: 1.93 [95% confidence interval 1.52-2.45]; p = 0.001). Individuals with an increase in HFF within 1 year had a significantly worse outcome than participants with stable HFF (HR adjusted for risk factors: 1.29 [1.01-1.65]; p = 0.04).

CONCLUSION

SLD is an independent predictor for long-term mortality in heavy smokers beyond known clinical risk factors.

Deep learning for autosegmentation for radiotherapy treatment planning: State-of-the-art and novel perspectives

The rapid development of artificial intelligence (AI) has gained importance, with many tools already entering our daily lives. The medical field of radiation oncology is also subject to this development, with AI entering all steps of the patient journey. In this review article, we summarize contemporary AI techniques and explore the clinical applications of AI-based automated segmentation models in radiotherapy planning, focusing on delineation of organs at risk (OARs), the gross tumor volume (GTV), and the clinical target volume (CTV). Emphasizing the need for precise and individualized plans, we review various commercial and freeware segmentation tools and also state-of-the-art approaches. Through our own findings and based on the literature, we demonstrate improved efficiency and consistency as well as time savings in different clinical scenarios. Despite challenges in clinical implementation such as domain shifts, the potential benefits for personalized treatment planning are substantial. The integration of mathematical tumor growth models and AI-based tumor detection further enhances the possibilities for refining target volumes. As advancements continue, the prospect of one-stop-shop segmentation and radiotherapy planning represents an exciting frontier in radiotherapy, potentially enabling fast treatment with enhanced precision and individualization.

SPINEPS-automatic whole spine segmentation of T2-weighted MR images using a two-phase approach to multi-class semantic and instance segmentation

OBJECTIVES

Introducing SPINEPS, a deep learning method for semantic and instance segmentation of 14 spinal structures (ten vertebra substructures, intervertebral discs, spinal cord, spinal canal, and sacrum) in whole-body sagittal T2-weighted turbo spin echo images.

MATERIAL AND METHODS

This local ethics committee-approved study utilized a public dataset (train/test 179/39 subjects, 137 female), a German National Cohort (NAKO) subset (train/test 1412/65 subjects, mean age 53, 694 female), and an in-house dataset (test 10 subjects, mean age 70, 5 female). SPINEPS is a semantic segmentation model, followed by a sliding window approach utilizing a second model to create instance masks from the semantic ones. Segmentation evaluation metrics included the Dice score and average symmetrical surface distance (ASSD). Statistical significance was assessed using the Wilcoxon signed-rank test.

RESULTS

On the public dataset, SPINEPS outperformed a nnUNet baseline on every structure and metric (e.g., an average over vertebra instances: dice 0.933 vs 0.911, p < 0.001, ASSD 0.21 vs 0.435, p < 0.001). SPINEPS trained on automated annotations of the NAKO achieves an average global Dice score of 0.918 on the combined NAKO and in-house test split. Adding the training data from the public dataset outperforms this (average instance-wise Dice score over the vertebra substructures 0.803 vs 0.778, average global Dice score 0.931 vs 0.918).

CONCLUSION

SPINEPS offers segmentation of 14 spinal structures in T2w sagittal images. It provides a semantic mask and an instance mask separating the vertebrae and intervertebral discs. This is the first publicly available algorithm to enable this segmentation.

KEY POINTS

Question No publicly available automatic approach can yield semantic and instance segmentation masks for the whole spine (including posterior elements) in T2-weighted sagittal TSE images. Findings Segmenting semantically first and then instance-wise outperforms a baseline trained directly on instance segmentation. The developed model produces high-resolution MRI segmentations for the whole spine. Clinical relevance This study introduces an automatic approach to whole spine segmentation, including posterior elements, in arbitrary fields of view T2w sagittal MR images, enabling easy biomarker extraction, automatic localization of pathologies and degenerative diseases, and quantifying analyses as downstream research.

NURBS curve shape prior-guided multiscale attention network for automatic segmentation of the inferior alveolar nerve

Accurate segmentation of the inferior alveolar nerve (IAN) within Cone-Beam Computed Tomography (CBCT) images is critical for the precise planning of oral and maxillofacial surgeries, especially to avoid IAN damage. Existing methods often fail due to the low contrast of the IAN and the presence of artifacts, which can cause segmentation discontinuities. To address these challenges, this paper proposes a novel approach that employs Non-Uniform Rational B-Spline (NURBS) curve shape priors into a multiscale attention network for the automatic segmentation of the IAN. Firstly, an automatic method for generating non-uniform rational B-spline (NURBS) shape prior is proposed and introduced into the segmentation network, which significantly enhancing the continuity and accuracy of IAN segmentation. Then a multiscale attention segmentation network, incorporating a dilation selective attention module is developed, to improve the network's feature extraction capacity. The proposed approach is validated on both in-house and public datasets, showcasing superior performance compared to established benchmarks, achieving 80.29±11.04% dice coefficient (Dice) and 68.14±12.06% intersection of union (IoU), the 95% Hausdorff distance (95HD) reaches 1.61±6.14 mm and mean surface distance (MSD) reaches 0.64±2.16 mm on private dataset. On public dataset, the Dice reaches 80.69±4.93%, IoU reaches 67.86±6.73%, 95HD reaches 1.04±0.95 mm, and MSD reaches 0.42±0.34 mm. Compared to state-of-the-art networks, the proposed approach out-performs in both voxel accuracy and surface distance. It offers significant potential to improve doctors' efficiency in segmentation tasks and holds promise for applications in dental surgery planning. The source codes are available at https://github.com/SJTUjsl/NURBS_IAN.git.

Automatic medical imaging segmentation via self-supervising large-scale convolutional neural networks

PURPOSE

This study aims to develop a robust, large-scale deep learning model for medical image segmentation, leveraging self-supervised learning to overcome the limitations of supervised learning and data variability in clinical settings.

METHODS AND MATERIALS

We curated a substantial multi-center CT dataset for self-supervised pre-training using masked image modeling with sparse submanifold convolution. We designed a series of Sparse Submanifold U-Nets (SS-UNets) of varying sizes and performed self-supervised pre-training. We fine-tuned the SS-UNets on the TotalSegmentator dataset. The evaluation encompassed robustness tests on four unseen datasets and transferability assessments on three additional datasets.

RESULTS

Our SS-UNets exhibited superior performance in comparison to state-of-the-art self-supervised methods, demonstrating higher Dice Similarity Coefficient (DSC) and Surface Dice Coefficient (SDC) metrics. SS-UNet-B achieved 84.3 % DSC and 88.0 % SDC in TotalSegmentator. We further demonstrated the scalability of our networks, with segmentation performance increasing with model size, demonstrated from 58 million to 1.4 billion parameters:4.6 % DSC and 3.2 % SDC improvement in TotalSegmentator from SS-UNet-B to SS-UNet-H.

CONCLUSIONS

We demonstrate the efficacy of self-supervised learning for medical image segmentation in the CT, MRI and PET domains. Our approach significantly reduces reliance on extensively labeled data, mitigates risks of overfitting, and enhances model generalizability. Future applications may allow accurate segmentation of organs and lesions across several imaging domains, potentially streamlining cancer detection and radiotherapy treatment planning.

Automatic segmentation of cardiac structures can change the way we evaluate dose limits for radiotherapy in the left breast

PURPOSE

Radiotherapy is a crucial part of breast cancer treatment. Precision in dose assessment is essential to minimize side effects. Traditionally, anatomical structures are delineated manually, a time-consuming process subject to variability. automatic segmentation, including methods based on multiple atlases and deep learning, offers a promising alternative. For the radiotherapy treatment of the left breast, the RTOG 1005 protocol highlights the importance of cardiac delineation and the need to minimize cardiac exposure to radiation. Our study aims to evaluate dose distribution in auto-segmented substructures and establish models to correlate them with dose in the cardiac area.

METHODS AND MATERIALS

Anatomical structures were auto-segmented using TotalSegmentator and Limbus AI. The relationship between the volume of the cardiac area and of organs at risk was assessed using log-linear regressions.

RESULTS

The mean dose distribution was considerable for LAD (left anterior descending coronary artery), heart, and left ventricle. The volumetric distribution of organs at risk is evaluated for specific RTOG 1005 isodoses. We highlight the greater variability in the absolute volumetric evaluation. Log-linear regression models are presented to estimate dose constraint parameters. We highlight a greater number of highly correlated comparisons for absolute dose-volume assessment.

CONCLUSIONS

Dose-volume assessment protocols in patients with left breast cancer often neglect cardiac substructures. However, automatic tools can overcome these technical difficulties. In this study, we correlated the dose in the cardiac area with the doses in specific substructures and suggested limits for planning evaluation. Our data also indicates that statistical models could be applied in the assessment of those substructures where an automatic segmentation tool is not available. Our data also shows a benefit in reporting absolute dose-volume thresholds for future cause-effect assessments.

The Role of AI in the Evaluation of Neuroendocrine Tumors: Current State of the Art

Advancements in Artificial Intelligence (AI) are driving a paradigm shift in the field of medical diagnostics, integrating new developments into various aspects of the clinical workflow. Neuroendocrine neoplasms are a diverse and heterogeneous group of tumors that pose significant diagnostic and management challenges due to their variable clinical presentations and biological behavior. Innovative approaches are essential to overcome these challenges and improve the current standard of care. AI-driven applications, particularly in imaging workflows, hold promise for enhancing tumor detection, classification, and grading by leveraging advanced radiomics and deep learning techniques. This article reviews the current and emerging applications of AI computer vision in the care of neuroendocrine neoplasms, focusing on its integration into imaging workflows, diagnostics, prognostic modeling, and therapeutic planning.

MedShapeNet - a large-scale dataset of 3D medical shapes for computer vision

OBJECTIVES

The shape is commonly used to describe the objects. State-of-the-art algorithms in medical imaging are predominantly diverging from computer vision, where voxel grids, meshes, point clouds, and implicit surface models are used. This is seen from the growing popularity of ShapeNet (51,300 models) and Princeton ModelNet (127,915 models). However, a large collection of anatomical shapes (e.g., bones, organs, vessels) and 3D models of surgical instruments is missing.

METHODS

We present MedShapeNet to translate data-driven vision algorithms to medical applications and to adapt state-of-the-art vision algorithms to medical problems. As a unique feature, we directly model the majority of shapes on the imaging data of real patients. We present use cases in classifying brain tumors, skull reconstructions, multi-class anatomy completion, education, and 3D printing.

RESULTS

By now, MedShapeNet includes 23 datasets with more than 100,000 shapes that are paired with annotations (ground truth). Our data is freely accessible via a web interface and a Python application programming interface and can be used for discriminative, reconstructive, and variational benchmarks as well as various applications in virtual, augmented, or mixed reality, and 3D printing.

CONCLUSIONS

MedShapeNet contains medical shapes from anatomy and surgical instruments and will continue to collect data for benchmarks and applications. The project page is: https://medshapenet.ikim.nrw/.

Development and validation of a prediction model for cardiac events in patients with hepatocellular carcinoma undergoing stereotactic body radiation therapy

PURPOSE

To develop and validate a prediction model for major adverse cardiac events (MACE) in hepatocellular carcinoma (HCC) patients treated with stereotactic body radiation therapy (SBRT).

METHODS AND MATERIALS

We retrospectively identified 1893 HCC patients who received SBRT at two institutions, with one serving as the development cohort (n=1473) and the other as the validation cohort (n=420). MACE was defined as any cardiac event classified as grade 3 or higher, according to the Common Terminology Criteria for Adverse Events, version 5.0. We evaluated 15 clinical and 88 dosimetric parameters using bootstrapped forward selection and area under the curve (AUC) to identify significant predictors for MACE. Based on these factors, we constructed the Cardiac Event Index (CEI) model, categorizing patients into distinct risk groups. Model performance was assessed for discrimination, efficiency, and calibration.

RESULTS

The occurrence rate of MACE was 5.8% in the development cohort and 6.7% in the validation cohort. Five parameters were selected for predicting MACE and were incorporated into the CEI model using the following equation: CEI = age score + hypertension + current smoking + (2 × history of cardiac disease) + (0.05 × heart-V5 [%]), which yielded an AUC of 0.770 for MACE and 0.750 for coronary artery disease. The CEI model stratified patients into low-, intermediate-, and high-risk groups that had MACE incidence rates of 0.4%, 4.9%, and 22.8%, respectively. The impact of heart-V5 on MACE was minimal in low- and intermediate-risk groups but pronounced in the high-risk group. In the validation cohort, the CEI model yielded an AUC of 0.809 for MACE and 0.793 for coronary artery disease.

CONCLUSIONS

The CEI model demonstrated robust performance in predicting MACE, revealing the significant influence of clinical factors and the minimal impact of SBRT. This model can inform evidence-based decisions regarding cardiac dose optimization in SBRT planning.

Current State of Evidence for Use of MRI in LI-RADS

The American College of Radiology Liver Imaging Reporting and Data System (LI-RADS) is the preeminent framework for classification and risk stratification of liver observations on imaging in patients at high risk for hepatocellular carcinoma. In this review, the pathogenesis of hepatocellular carcinoma and the use of MRI in LI-RADS is discussed, including specifically the LI-RADS diagnostic algorithm, its components, and its reproducibility with reference to the latest supporting evidence. The LI-RADS treatment response algorithms are reviewed, including the more recent radiation treatment response algorithm. The application of artificial intelligence, points of controversy, LI-RADS relative to other liver imaging systems, and possible future directions are explored. After reading this article, the reader will have an understanding of the foundation and application of LI-RADS as well as possible future directions.

Robust and generalizable artificial intelligence for multi-organ segmentation in ultra-low-dose total-body PET imaging: a multi-center and cross-tracer study

PURPOSE

Positron Emission Tomography (PET) is a powerful molecular imaging tool that visualizes radiotracer distribution to reveal physiological processes. Recent advances in total-body PET have enabled low-dose, CT-free imaging; however, accurate organ segmentation using PET-only data remains challenging. This study develops and validates a deep learning model for multi-organ PET segmentation across varied imaging conditions and tracers, addressing critical needs for fully PET-based quantitative analysis.

MATERIALS AND METHODS

This retrospective study employed a 3D deep learning-based model for automated multi-organ segmentation on PET images acquired under diverse conditions, including low-dose and non-attenuation-corrected scans. Using a dataset of 798 patients from multiple centers with varied tracers, model robustness and generalizability were evaluated via multi-center and cross-tracer tests. Ground-truth labels for 23 organs were generated from CT images, and segmentation accuracy was assessed using the Dice similarity coefficient (DSC).

RESULTS

In the multi-center dataset from four different institutions, our model achieved average DSC values of 0.834, 0.825, 0.819, and 0.816 across varying dose reduction factors and correction conditions for FDG PET images. In the cross-tracer dataset, the model reached average DSC values of 0.737, 0.573, 0.830, 0.661, and 0.708 for DOTATATE, FAPI, FDG, Grazytracer, and PSMA, respectively.

CONCLUSION

The proposed model demonstrated effective, fully PET-based multi-organ segmentation across a range of imaging conditions, centers, and tracers, achieving high robustness and generalizability. These findings underscore the model's potential to enhance clinical diagnostic workflows by supporting ultra-low dose PET imaging.

CLINICAL TRIAL NUMBER

Not applicable. This is a retrospective study based on collected data, which has been approved by the Research Ethics Committee of Ruijin Hospital affiliated to Shanghai Jiao Tong University School of Medicine.

Moving Beyond CT Body Composition Analysis: Using Style Transfer for Bringing CT-Based Fully-Automated Body Composition Analysis to T2-Weighted MRI Sequences

OBJECTIVES

Deep learning for body composition analysis (BCA) is gaining traction in clinical research, offering rapid and automated ways to measure body features like muscle or fat volume. However, most current methods prioritize computed tomography (CT) over magnetic resonance imaging (MRI). This study presents a deep learning approach for automatic BCA using MR T2-weighted sequences.

METHODS

Initial BCA segmentations (10 body regions and 4 body parts) were generated by mapping CT segmentations from body and organ analysis (BOA) model to synthetic MR images created using an in-house trained CycleGAN. In total, 30 synthetic data pairs were used to train an initial nnU-Net V2 in 3D, and this preliminary model was then applied to segment 120 real T2-weighted MRI sequences from 120 patients (46% female) with a median age of 56 (interquartile range, 17.75), generating early segmentation proposals. These proposals were refined by human annotators, and nnU-Net V2 2D and 3D models were trained using 5-fold cross-validation on this optimized dataset of real MR images. Performance was evaluated using Sørensen-Dice, Surface Dice, and Hausdorff Distance metrics including 95% confidence intervals for cross-validation and ensemble models.

RESULTS

The 3D ensemble segmentation model achieved the highest Dice scores for the body region classes: bone 0.926 (95% confidence interval [CI], 0.914-0.937), muscle 0.968 (95% CI, 0.961-0.975), subcutaneous fat 0.98 (95% CI, 0.971-0.986), nervous system 0.973 (95% CI, 0.965-0.98), thoracic cavity 0.978 (95% CI, 0.969-0.984), abdominal cavity 0.989 (95% CI, 0.986-0.991), mediastinum 0.92 (95% CI, 0.901-0.936), pericardium 0.945 (95% CI, 0.924-0.96), brain 0.966 (95% CI, 0.927-0.989), and glands 0.905 (95% CI, 0.886-0.921). Furthermore, body part 2D ensemble model reached the highest Dice scores for all labels: arms 0.952 (95% CI, 0.937-0.965), head + neck 0.965 (95% CI, 0.953-0.976), legs 0.978 (95% CI, 0.968-0.988), and torso 0.99 (95% CI, 0.988-0.991). The overall average Dice across body parts (2D = 0.971, 3D = 0.969, P = ns) and body regions (2D = 0.935, 3D = 0.955, P < 0.001) ensemble models indicates stable performance across all classes.

CONCLUSIONS

The presented approach facilitates efficient and automated extraction of BCA parameters from T2-weighted MRI sequences, providing precise and detailed body composition information across various regions and body parts.

Detection and correction of translational motion in SPECT with exponential data consistency conditions

Objective.Rigid patient motion can cause artifacts in single photon emission computed tomography (SPECT) images, compromising the diagnosis and treatment planning. Exponential data consistency conditions (eDCCs) are mathematical equations describing the redundancy of exponential SPECT measurements. It has been recently shown that eDCCs can be used to detect patient motion in SPECT projections. This study aimed at developing a fully data-driven method based on eDCCs to estimate and correct for translational motion in SPECT.Approach.If all activity is encompassed inside a convex regionKof constant attenuation, eDCCs can be derived from SPECT projections and can be used to verify the pairwise consistency of these projections. Our method assumes a single patient translation between two detector gantry positions. The proposed method estimates both the three-dimensional shift and the motion index, i.e. the index of the first gantry position after motion occurred. The estimation minimizes the eDCCs between the subset of projections before the motion index and the subset of motion-corrected projections after the motion index.Results.We evaluated the proposed method using Monte Carlo simulated and experimental data of a NEMA IEC phantom and simulated projections of a liver patient. The method's robustness was assessed by applying various motion vectors and motion indices. Motion detection and correction with eDCCs were sensitive to movements above 3 mm. The accuracy of the estimation was below the 2.39 mm pixel spacing with good precision in all studied cases. The proposed method led to a significant improvement in the quality of reconstructed SPECT images. The activity recovery coefficient relative to the SPECT image without motion was above 90% on average over the six spheres of the NEMA IEC phantom and 97% for the liver patient. For example, for a(2,2,2)cm translation in the middle of the liver acquisition, the activity recovery coefficient was improved from 35% (non-corrected projections) to 99% (motion-corrected projections).Significance.The study proposed and demonstrated the good performance of translational motion detection and correction with eDCCs in SPECT acquisition data.

Comparative analysis of intestinal tumor segmentation in PET CT scans using organ based and whole body deep learning

BACKGROUND

18-Fluoro-deoxyglucose positron emission tomography/computed tomography (FDG-PET/CT) is a valuable imaging tool widely used in the management of cancer patients. Deep learning models excel at segmenting highly metabolic tumors but face challenges in regions with complex anatomy and normal cell uptake, such as the gastro-intestinal tract. Despite these challenges, it remains important to achieve accurate segmentation of gastro-intestinal tumors.

METHODS

Here, we present an international multicenter comparative study between a novel organ-focused approach and a whole-body training method to evaluate the effectiveness of training data homogeneity in accurately identifying gastro-intestinal tumors. In the organ-focused method, the training data is limited to cases with intestinal tumors which makes the network trained with more homogeneous data and with stronger presence of intestinal tumor signals. The whole body approach extracts the intestinal tumors from the results of a model trained on the whole-body scans. Both approaches were trained using diffuse large B cell (DLBCL) patients from a large multi-center clinical trial (NCT01287741).

RESULTS

We report an improved mean(±std) Dice score of 0.78(±0.21) for the organ-based approach on the hold-out set, compared to 0.63(±0.30) for the whole-body approach, with the p-value of less than 0.0001. At the lesion level, the proposed organ-based approach also shows increased precision, recall, and F1-score. An independent trial was used to evaluate the generalizability of the proposed method to non-Hodgkin's lymphoma (NHL) patients with follicular lymphoma (FL).

CONCLUSION

Given the variability in structure and metabolism across tissues in the body, our quantitative findings suggest organ-focused training enhances intestinal tumor segmentation by leveraging tissue homogeneity in the training data, contrasting with the whole-body training approach, which, by its very nature, is a more heterogeneous data set.

Unified total body CT image with multiple organ specific windowings: validating improved diagnostic accuracy and speed in trauma cases

Total-body CT scans are useful in saving trauma patients; however, interpreting numerous images with varied window settings slows injury detection. We developed an algorithm for "unified total-body CT image with multiple organ-specific windowings (Uni-CT)", and assessing its impact on physician accuracy and speed in trauma CT interpretation. From November 7, 2008, to June 19, 2020, 40 cases of total-body CT images for blunt trauma with multiple injuries, were collected from the emergency department of Osaka General Medical Center and randomly divided into two groups. In half of the cases, the Uni-CT algorithm using semantic segmentation assigned visibility-friendly window settings to each organ. Four physicians with varying levels of experience interpreted 20 cases using the algorithm and 20 cases in conventional settings. The performance was analyzed based on the accuracy, sensitivity, specificity of the target findings, and diagnosis speed. In the proposal and conventional groups, patients had an average of 2.6 and 2.5 targeting findings, mean ages of 51.8 and 57.7 years, and male proportions of 60% and 45%, respectively. The agreement rate for physicians' diagnoses was κ = 0.70. Average accuracy, sensitivity, and specificity of target findings were 84.8%, 74.3%, 96.9% and 85.5%, 81.2%, 91.5%, respectively, with no significant differences. Diagnostic speed per case averaged 71.9 and 110.4 s in each group (p < 0.05). The Uni-CT algorithm improved the diagnostic speed of total-body CT for trauma, maintaining accuracy comparable to that of conventional methods.

Automatic Segmentation of Cardiovascular Structures on Chest CT Data Sets: An Update of the TotalSegmentator

INTRODUCTION

Quantitative analysis is an important factor in radiological routine. Recently the TotalSegmentator was released, a free-to-use segmentation tool with over 104 structures included. Our aim was to add missing and enhance previously included cardiovascular (CV) structures to potentially help find new insights into diseases such as aortic aneurysms in future studies. The TotalSegmentator data set with 1613 CT scans (mean age 63.6 ± 15.9 (SD); 675 female), was used. CT scans were selected from clinical routine including various protocols and pathologies. The data set was split in training (1472), validation (57) and testing (84). Segmentations were performed in dedicated imaging software using an iterative approach for training to reduce segmentation workload. Eleven structures were added, and segmentations of six structures were enhanced. The Dice similarity score (DICE) and the Normalized surface distance (NSD) were calculated on an internal and external data set. The external validation was performed on the Dongyang data set. The Mann Whitney U test was performed to evaluate the performance increase on the previously included structures.

RESULTS

Median DICE [IQR] and NSD [IQR] were 0.967 [0.020] and 1.000 [0.000], respectively. DICE (p < 0.001) and NSD (p < 0.001) significantly increased for 5/6 structures. On evaluation using the external data set, DICE and NSD were 0.970 [0.020] and 1.000 [0.000].

CONCLUSION

Accurate segmentations and enhanced segmentations of previously included CV structures were successfully implemented. This suggests further usage in research studies while still running on conventional computers with or without a dedicated graphics processing unit.

Enhancing U-Net-based Pseudo-CT generation from MRI using CT-guided bone segmentation for radiation treatment planning in head & neck cancer patients

Objective.This study investigates the effects of various training protocols on enhancing the precision of MRI-only Pseudo-CT generation for radiation treatment planning and adaptation in head & neck cancer patients. It specifically tackles the challenge of differentiating bone from air, a limitation that frequently results in substantial deviations in the representation of bony structures on Pseudo-CT images.Approach.The study included 25 patients, utilizing pre-treatment MRI-CT image pairs. Five cases were randomly selected for testing, with the remaining 20 used for model training and validation. A 3D U-Net deep learning model was employed, trained on patches of size 643with an overlap of 323. MRI scans were acquired using the Dixon gradient echo (GRE) technique, and various contrasts were explored to improve Pseudo-CT accuracy, including in-phase, water-only, and combined water-only and fat-only images. Additionally, bone extraction from the fat-only image was integrated as an additional channel to better capture bone structures on Pseudo-CTs. The evaluation involved both image quality and dosimetric metrics.Main results.The generated Pseudo-CTs were compared with their corresponding registered target CTs. The mean absolute error (MAE) and peak signal-to-noise ratio (PSNR) for the base model using combined water-only and fat-only images were 19.20 ± 5.30 HU and 57.24 ± 1.44 dB, respectively. Following the integration of an additional channel using a CT-guided bone segmentation, the model's performance improved, achieving MAE and PSNR of 18.32 ± 5.51 HU and 57.82 ± 1.31 dB, respectively. The measured results are statistically significant, with ap-value<0.05. The dosimetric assessment confirmed that radiation treatment planning on Pseudo-CT achieved accuracy comparable to conventional CT.Significance.This study demonstrates improved accuracy in bone representation on Pseudo-CTs achieved through a combination of water-only, fat-only and extracted bone images; thus, enhancing feasibility of MRI-based simulation for radiation treatment planning.

Artificial Intelligence Model for Detection of Colorectal Cancer on Routine Abdominopelvic CT Examinations: A Training and External-Testing Study

Background: Radiologists are prone to missing some colorectal cancers (CRCs) on routine abdominopelvic CT examinations that are in fact detectable on the images. Objective: To develop an artificial intelligence (AI) model to detect CRC on routine abdominopelvic CT examinations, performed without bowel preparation. Methods: This retrospective study included 3945 patients (2275 men, 1670 women; mean age, 62 years): a training set of 2662 patients from Severance Hospital with CRC who underwent routine contrast-enhanced abdominopelvic CT before treatment between January 2010 and December 2014; and internal (841 patients from Severance Hospital) and external (442 patients from Gangnam Severance Hospital) test sets of patients who underwent routine contrast-enhanced abdominopelvic CT for any indication and colonoscopy within a 2-month interval between January 2018 and June 2018. A radiologist, accessing colonoscopy reports, determined which CRCs were visible on CT and placed bounding boxes around lesions on all slices showing CRC, serving as the reference standard. A contemporary transformer-based object detection network was adapted and trained to create an AI model (https://github.com/boktae7/colorectaltumor) to automatically detect CT-visible CRC on unprocessed DICOM slices. AI performance was evaluated using alternative free-response ROC analysis, per-lesion sensitivity, and per-patient specificity; performance in the external test set was compared to that of two radiologist readers. Clinical radiology reports were also reviewed. Results: In the internal (93 CT-visible CRCs in 92 patients) and external (26 CT-visible CRCs in 26 patients) test sets, AI had AUC of 0.867 and 0.808, sensitivity of 79.6% and 80.8%, and specificity of 91.2% and 90.9%, respectively. In the external test set, the two radiologists had sensitivities of 73.1% and 80.8% (p=.74 and p>.99 vs AI) and specificities of 98.3% and 98.6% (both p<.001 vs AI); AI correctly detected five of nine CRCs missed by at least one reader. The clinical radiology reports raised suspicion for 75.9% of CRCs in the external test set. Conclusion: The findings demonstrate the AI model's utility for automated detection of CRC on routine abdominopelvic CT examinations. Clinical Impact: The AI model could help reduce the frequency of missed CRCs on routine examinations performed for reasons unrelated to CRC detection.

Automatic bone marrow segmentation for precise [177Lu]Lu-PSMA-617 dosimetry

BACKGROUND

Bone marrow (BM) is the dominant dose-limiting organ in [177Lu]Lu-PSMA-617 therapy for patients with metastasized castration resistant prostate cancer, where BM dosimetry is challenging due to segmentation.

PURPOSE

We aim to develop an automatic image-based segmentation method on peri-therapeutic sequential [177Lu]Lu-PSMA-617 images for personalized BM dosimetry.

METHODS

Quantitative SPECT/CT imaging at 2, 20, 40 and 60 (n = 14)/200 (n = 16) h post [177Lu]Lu-PSMA-617 injection were analyzed for 10 patients with 30 treatment cycles. X-means clustering was applied on the deep learning-based segmented lumbar spines CT images to classify the BM region. A single threshold method, two empirical segmentation methods (one sphere and five spheres), and gold standard manual segmentation were also implemented. The Dice similarity coefficient between BM masks of the X-means clustering and single threshold method was calculated as compared to the gold standard. BM mean absorbed dose (Dmean) was obtained for different segmentation methods. Absolute errors and Bland-Altman analysis were also evaluated for BM Dmean derived from evaluated segmentation methods compared with the gold standard. Wilcoxon signed-rank test was performed for statistical evaluation. BM Dmean was correlated with the change of platelets and white blood cells (WBC) pre- and post-treatment using Pearson correlation analysis.

RESULTS

In 30 cycles of 10 patients, the average Dice is 0.76 ± 0.18 for the X-means clustering method, as compared to 0.61 ± 0.19 for the single threshold method. The gold standard yields mean BM Dmean of 0.46 ± 0.69 Gy. The X-means clustering method exhibits significantly (p < 0.01) lower mean absolute BM Dmean (25.34 ± 64.48%) errors, followed by the single threshold (32.46 ± 69.49%), one sphere (50.53 ± 35.40%), and five spheres (72.73 ± 115.97%) methods. Bland-Altman analysis reveals that the X-means clustering method has a smaller Dmean difference (0.0330 Gy) compared to the single threshold (0.0512 Gy), one sphere (-0.1903 Gy), and five spheres (0.2108 Gy) methods. Stronger correlations (r ≤ -0.65) are found between platelets/WBC changes and BM Dmean from both the gold standard and X-means clustering methods than other methods.

CONCLUSIONS

X-means clustering is feasible to segment the BM based on the CT images of peri-therapy SPECT/CT and shows advantages compared with the single threshold and empirical sphere segmentation methods.

Automated detection and quantification of aortic calcification in coronary CT angiography using deep learning: A comparative study of manual and automated scoring methods

BACKGROUND

Aortic calcification, often incidentally detected during coronary artery calcium (CAC) scans, is underutilized in cardiovascular risk assessments due to manual quantification challenges. This study evaluates a deep learning model for automating aortic calcification detection and quantification in coronary CT angiography (CTA) images. We validate against manual assessments and compare the association of manual and automated assessments with incident major adverse cardiovascular events (MACE).

METHODS

A deep learning algorithm was applied to CAC scans from 670 participants in the CORE320 and CORE64 studies. Aortic calcification in the aortic root, ascending, and descending aorta was quantified manually and automatically. Concordance correlation coefficients (CCC) assessed agreement, and Cox regression and ROC analyses evaluated association with incident MACE.

RESULTS

Automated scoring demonstrated high concordance with manual methods (CCC: 0.926-0.992), supporting its reliability in assessing aortic calcifications. ROC analysis revealed that the automated method was as effective as the manual technique in predicting MACE (p ​> ​0.05).

CONCLUSION

Automated aortic calcification scoring is a reliable alternative to manual methods, offering consistency and efficiency in the analysis of incidental findings on CAC scans.

Systematic Review: AI Applications in Liver Imaging with a Focus on Segmentation and Detection

The liver is a frequent focus in radiology due to its diverse pathology, and artificial intelligence (AI) could improve diagnosis and management. This systematic review aimed to assess and categorize research studies on AI applications in liver radiology from 2018 to 2024, classifying them according to areas of interest (AOIs), AI task and imaging modality used. We excluded reviews and non-liver and non-radiology studies. Using the PRISMA guidelines, we identified 6680 articles from the PubMed/Medline, Scopus and Web of Science databases; 1232 were found to be eligible. A further analysis of a subgroup of 329 studies focused on detection and/or segmentation tasks was performed. Liver lesions were the main AOI and CT was the most popular modality, while classification was the predominant AI task. Most detection and/or segmentation studies (48.02%) used only public datasets, and 27.65% used only one public dataset. Code sharing was practiced by 10.94% of these articles. This review highlights the predominance of classification tasks, especially applied to liver lesion imaging, most often using CT imaging. Detection and/or segmentation tasks relied mostly on public datasets, while external testing and code sharing were lacking. Future research should explore multi-task models and improve dataset availability to enhance AI's clinical impact in liver imaging.

Real world federated learning with a knowledge distilled transformer for cardiac CT imaging

Federated learning is a renowned technique for utilizing decentralized data while preserving privacy. However, real-world applications often face challenges like partially labeled datasets, where only a few locations have certain expert annotations, leaving large portions of unlabeled data unused. Leveraging these could enhance transformer architectures' ability in regimes with small and diversely annotated sets. We conduct the largest federated cardiac CT analysis to date (n = 8, 104) in a real-world setting across eight hospitals. Our two-step semi-supervised strategy distills knowledge from task-specific CNNs into a transformer. First, CNNs predict on unlabeled data per label type and then the transformer learns from these predictions with label-specific heads. This improves predictive accuracy and enables simultaneous learning of all partial labels across the federation, and outperforms UNet-based models in generalizability on downstream tasks. Code and model weights are made openly available for leveraging future cardiac CT analysis.

Optimizing MR-based attenuation correction in hybrid PET/MR using deep learning: validation with a flatbed insert and consistent patient positioning

PURPOSE

To address the challenges of verifying MR-based attenuation correction (MRAC) in PET/MR due to CT positional mismatches and alignment issues, this study utilized a flatbed insert and arms-down positioning during PET/CT scans to achieve precise MR-CT matching for accurate MRAC evaluation.

METHODS

A validation dataset of 21 patients underwent whole-body [18F]FDG PET/CT followed by [18F]FDG PET/MR. A flatbed insert ensured consistent positioning, allowing direct comparison of four MRAC methods-four-tissue and five-tissue models with discrete and continuous μ-maps-against CT-based attenuation correction (CTAC). A deep learning-based framework, trained on a dataset of 300 patients, was used to generate synthesized-CTs from MR images, forming the basis for all MRAC methods. Quantitative analyses were conducted at the whole-body, region of interest, and lesion levels, with lesion-distance analysis evaluating the impact of bone proximity on standardized uptake value (SUV) quantification.

RESULTS

Distinct differences were observed among MRAC methods in spine and femur regions. Joint histogram analysis showed MRAC-4 (continuous μ-map) closely aligned with CTAC. Lesion-distance analysis revealed MRAC-4 minimized bone-induced SUV interference (r = 0.01, p = 0.8643). However, tissues prone to bone segmentation interference, such as the spine and liver, exhibited greater SUV variability and lower reproducibility in MRAC-4 compared to MRAC-2 (2D bone segmentation, discrete μ-map) and MRAC-3 (3D bone segmentation, discrete μ-map).

CONCLUSION

Using a flatbed insert, this study validated MRAC with high precision. Continuous μ-value MRAC method (MRAC-4) demonstrated superior accuracy and minimized bone-related SUV errors but faced challenges in reproducibility, particularly in bone-rich regions.

Optimizing transformer-based network via advanced decoder design for medical image segmentation

U-Net is widely used in medical image segmentation due to its simple and flexible architecture design. To address the challenges of scale and complexity in medical tasks, several variants of U-Net have been proposed. In particular, methods based on Vision Transformer (ViT), represented by Swin UNETR, have gained widespread attention in recent years. However, these improvements often focus on the encoder, overlooking the crucial role of the decoder in optimizing segmentation details. This design imbalance limits the potential for further enhancing segmentation performance. To address this issue, we analyze the roles of various decoder components, including upsampling method, skip connection, and feature extraction module, as well as the shortcomings of existing methods. Consequently, we propose Swin DER (i.e.,SwinUNETRDecoderEnhanced andRefined), by specifically optimizing the design of these three components. Swin DER performs upsampling using learnable interpolation algorithm called offset coordinate neighborhood weighted up sampling (Onsampling) and replaces traditional skip connection with spatial-channel parallel attention gate (SCP AG). Additionally, Swin DER introduces deformable convolution along with attention mechanism in the feature extraction module of the decoder. Our model design achieves excellent results, surpassing other state-of-the-art methods on both the Synapse dataset and the MSD brain tumor segmentation task. Code is available at:https://github.com/WillBeanYang/Swin-DER.

Automatic cervical lymph nodes detection and segmentation in heterogeneous computed tomography images using deep transfer learning

To develop a deep learning model using transfer learning for automatic detection and segmentation of neck lymph nodes (LNs) in computed tomography (CT) images, the study included 11,013 annotated LNs with a short-axis diameter ≥ 3 mm from 626 head and neck cancer patients across four hospitals. The nnUNet model was used as a baseline, pre-trained on a large-scale head and neck dataset, and then fine-tuned with 4,729 LNs from hospital A for detection and segmentation. Validation was conducted on an internal testing cohort (ITC A) and three external testing cohorts (ETCs B, C, and D), with 1684 and 4600 LNs, respectively. Detection was evaluated via sensitivity, positive predictive value (PPV), and false positive rate per case (FP/vol), while segmentation was assessed using the Dice similarity coefficient (DSC) and Hausdorff distance (HD95). For detection, the sensitivity, PPV, and FP/vol in ITC A were 54.6%, 69.0%, and 3.4, respectively. In ETCs, the sensitivity ranged from 45.7% at 3.9 FP/vol to 63.5% at 5.8 FP/vol. Segmentation achieved a mean DSC of 0.72 in ITC A and 0.72 to 0.74 in ETCs, as well as a mean HD95 of 3.78 mm in ITC A and 2.73 mm to 2.85 mm in ETCs. No significant sensitivity difference was found between contrast-enhanced and unenhanced CT images (p = 0.502) or repeated CT images (p = 0.815) during adaptive radiotherapy. The model's segmentation accuracy was comparable to that of experienced oncologists. The model shows promise in automatically detecting and segmenting neck LNs in CT images, potentially reducing oncologists' segmentation workload.

A Machine Learning Model Based on Radiomic Features as a Tool to Identify Active Giant Cell Arteritis on [18F]FDG-PET Images During Follow-Up

Objective: To investigate the feasibility of a machine learning (ML) model based on radiomic features to identify active giant cell arteritis (GCA) in the aorta and differentiate it from atherosclerosis in follow-up [18F]FDG-PET/CT images for therapy monitoring. Methods: To train the ML model, 64 [18F]FDG-PET scans of 34 patients with proven GCA and 34 control subjects with type 2 diabetes mellitus were retrospectively included. The aorta was delineated into the ascending, arch, descending, and abdominal aorta. From each segment, 95 features were extracted. All segments were randomly split into a training/validation (n = 192; 80%) and test set (n = 46; 20%). In total, 441 ML models were trained, using combinations of seven feature selection methods, seven classifiers, and nine different numbers of features. The performance was assessed by area under the curve (AUC). The best performing ML model was compared to the clinical report of nuclear medicine physicians in 19 follow-up scans (7 active GCA, 12 inactive GCA). For explainability, an occlusion map was created to illustrate the important regions of the aorta for the decision of the ML model. Results: The ten-feature model with ANOVA as the feature selector and random forest classifier demonstrated the highest performance (AUC = 0.92 ± 0.01). Compared with the clinical report, this model showed a higher PPV (0.83 vs. 0.80), NPV (0.85 vs. 0.79), and accuracy (0.84 vs. 0.79) in the detection of active GCA in follow-up scans. Conclusions: The current radiomics ML model was able to identify active GCA and differentiate GCA from atherosclerosis in follow-up [18F]FDG-PET/CT scans. This demonstrates the potential of the ML model as a monitoring tool in challenging [18F]FDG-PET scans of GCA patients.

Multi-modal dataset creation for federated learning with DICOM-structured reports

Purpose Federated training is often challenging on heterogeneous datasets due to divergent data storage options, inconsistent naming schemes, varied annotation procedures, and disparities in label quality. This is particularly evident in the emerging multi-modal learning paradigms, where dataset harmonization including a uniform data representation and filtering options are of paramount importance.Methods DICOM-structured reports enable the standardized linkage of arbitrary information beyond the imaging domain and can be used within Python deep learning pipelines with highdicom. Building on this, we developed an open platform for data integration with interactive filtering capabilities, thereby simplifying the process of creation of patient cohorts over several sites with consistent multi-modal data.Results In this study, we extend our prior work by showing its applicability to more and divergent data types, as well as streamlining datasets for federated training within an established consortium of eight university hospitals in Germany. We prove its concurrent filtering ability by creating harmonized multi-modal datasets across all locations for predicting the outcome after minimally invasive heart valve replacement. The data include imaging and waveform data (i.e., computed tomography images, electrocardiography scans) as well as annotations (i.e., calcification segmentations, and pointsets), and metadata (i.e., prostheses and pacemaker dependency).Conclusion Structured reports bridge the traditional gap between imaging systems and information systems. Utilizing the inherent DICOM reference system arbitrary data types can be queried concurrently to create meaningful cohorts for multi-centric data analysis. The graphical interface as well as example structured report templates are available at https://github.com/Cardio-AI/fl-multi-modal-dataset-creation .

Machine Learning-Based CT Radiomics Model to Predict the Risk of Hip Fragility Fracture

RATIONALE AND OBJECTIVES

This research aimed to develop a combined model based on proximal femur attenuation values and radiomics features at routine CT to predict hip fragility fracture using machine learning methods.

METHOD

A total of 254 patients (training cohort, n=132; test cohort 1, n=56；test cohort 2, n=66) who underwent hip or pelvic CT scans were included. Three different machine learning methods were used to build the Support Vector Machine (SVM) model, Logistic Regression (LR) model and Random Forest (RF) model respectively. The method that exhibited the best performance in the training cohort and test cohort 1 was selected to represent the radiomics model for subsequent studies. The mean CT Hounsfield unit of three-dimensional CT images at the proximal femur was extracted to construct the mean CTHU model. Multivariate logistic regression was performed using mean CT Hounsfield unit together with radiomics features, and the combined model was subsequently developed with a visualized nomogram.

RESULTS

Among the radiomics models based on three machine learning methods, the LR model showed the best performance in the training cohort (AUC=0.875, 95% CI=0.806-0.926) and in the test cohort 1 (AUC=0.851, 95% CI=0.730-0.932). Compared to the mean CT model and the LR model, the combined model showed superior discriminatory power in the training cohort (AUC=0.934, 95% CI=0.895-0.972), the test cohort 1 (AUC=0.893, 95% CI=0.812-0.974) and the test cohort 2 (AUC=0.851, 95% CI=0.742-0.927).

CONCLUSION

The combined model, based on the mean CT Hounsfield unit of the proximal femur and radiomics features, can provide an accurate quantitative imaging basis for individualized risk prediction of hip fragility fracture.

18F-FDG-PET/CT Uptake by Noncancerous Lung as a Predictor of Interstitial Lung Disease Induced by Immune Checkpoint Inhibitors

RATIONALE AND OBJECTIVES

Immune checkpoint inhibitors (ICIs) have improved lung cancer prognosis; however, ICI-related interstitial lung disease (ILD) is fatal and difficult to predict. Herein, we hypothesized that pre-existing lung inflammation on radiological imaging can be a potential risk factor for ILD onset. Therefore, we investigated the association between high uptake in noncancerous lung (NCL) on 18F- FDG-PET/CT and ICI-ILD in lung cancer.

METHODS

Patients with primary lung cancer who underwent FDG-PET/CT within three months prior to ICI therapy were retrospectively included. Artificial intelligence was utilized for extracting the NCL regions (background lung) from the lung contralateral to the primary tumor. FDG uptake by the NCL was assessed via the SUVmax (NCL-SUVmax), SUVmean (NCL-SUVmean), and total glycolytic activity (NCL-TGA)defined as NCL-SUVmean×NCL volume [mL]. NCL-SUVmean and NCL-TGA were calculated using the following four SUV thresholds: 0.5, 1.0, 1.5, and 2.0.

RESULTS

Of the 165 patients, 28 (17.0%) developed ILD. Univariate analysis showed that high values of NCL-SUVmax, NCL-SUVmean2.0 (SUV threshold=2.0), and NCL-TGA1.0 (SUV threshold=1.0) were significantly associated with ILD onset (all p = 0.003). Multivariate analysis adjusted for age, tumor FDG uptake, and pre-existing interstitial lung abnormalities revealed that a high NCL-TGA1.0 (≥149.45) was independently associated with ILD onset (odds ratio, 6.588; p = 0.002). Two-year cumulative incidence of ILD was significantly higher in the high NCL-TGA1.0 group than in the low group (58.4% vs. 14.4%; p < 0.001).

CONCLUSION

High uptake of NCL on FDG-PET/CT is correlated with ICI-ILD development, which could serve as a risk stratification tool before ICI therapy in primary lung cancer.

G-SET-DCL: a guided sequential episodic training with dual contrastive learning approach for colon segmentation

PURPOSE

This article introduces a novel deep learning approach to substantially improve the accuracy of colon segmentation even with limited data annotation, which enhances the overall effectiveness of the CT colonography pipeline in clinical settings.

METHODS

The proposed approach integrates 3D contextual information via guided sequential episodic training in which a query CT slice is segmented by exploiting its previous labeled CT slice (i.e., support). Segmentation starts by detecting the rectum using a Markov Random Field-based algorithm. Then, supervised sequential episodic training is applied to the remaining slices, while contrastive learning is employed to enhance feature discriminability, thereby improving segmentation accuracy.

RESULTS

The proposed method, evaluated on 98 abdominal scans of prepped patients, achieved a Dice coefficient of 97.3% and a polyp information preservation accuracy of 98.28%. Statistical analysis, including 95% confidence intervals, underscores the method's robustness and reliability. Clinically, this high level of accuracy is vital for ensuring the preservation of critical polyp details, which are essential for accurate automatic diagnostic evaluation. The proposed method performs reliably in scenarios with limited annotated data. This is demonstrated by achieving a Dice coefficient of 97.15% when the model was trained on a smaller number of annotated CT scans (e.g., 10 scans) than the testing dataset (e.g., 88 scans).

CONCLUSIONS

The proposed sequential segmentation approach achieves promising results in colon segmentation. A key strength of the method is its ability to generalize effectively, even with limited annotated datasets-a common challenge in medical imaging.

Multisensory Extended Reality Applications Offer Benefits for Volumetric Biomedical Image Analysis in Research and Medicine

3D data from high-resolution volumetric imaging is a central resource for diagnosis and treatment in modern medicine. While the fast development of AI enhances imaging and analysis, commonly used visualization methods lag far behind. Recent research used extended reality (XR) for perceiving 3D images with visual depth perception and touch but used restrictive haptic devices. While unrestricted touch benefits volumetric data examination, implementing natural haptic interaction with XR is challenging. The research question is whether a multisensory XR application with intuitive haptic interaction adds value and should be pursued. In a study, 24 experts for biomedical images in research and medicine explored 3D medical shapes with 3 applications: a multisensory virtual reality (VR) prototype using haptic gloves, a simple VR prototype using controllers, and a standard PC application. Results of standardized questionnaires showed no significant differences between all application types regarding usability and no significant difference between both VR applications regarding presence. Participants agreed to statements that VR visualizations provide better depth information, using the hands instead of controllers simplifies data exploration, the multisensory VR prototype allows intuitive data exploration, and it is beneficial over traditional data examination methods. While most participants mentioned manual interaction as the best aspect, they also found it the most improvable. We conclude that a multisensory XR application with improved manual interaction adds value for volumetric biomedical data examination. We will proceed with our open-source research project ISH3DE (Intuitive Stereoptic Haptic 3D Data Exploration) to serve medical education, therapeutic decisions, surgery preparations, or research data analysis.

Development and validation of fully automated robust deep learning models for multi-organ segmentation from whole-body CT images

PURPOSE

This study aimed to develop a deep-learning framework to generate multi-organ masks from CT images in adult and pediatric patients.

METHODS

A dataset consisting of 4082 CT images and ground-truth manual segmentation from various databases, including 300 pediatric cases, were collected. In strategy#1, the manual segmentation masks provided by public databases were split into training (90%) and testing (10% of each database named subset #1) cohort. The training set was used to train multiple nnU-Net networks in five-fold cross-validation (CV) for 26 separate organs. In the next step, the trained models from strategy #1 were used to generate missing organs for the entire dataset. This generated data was then used to train a multi-organ nnU-Net segmentation model in a five-fold CV (strategy#2). Models' performance were evaluated in terms of Dice coefficient (DSC) and other well-established image segmentation metrics.

RESULTS

The lowest CV DSC for strategy#1 was 0.804 ± 0.094 for adrenal glands while average DSC > 0.90 were achieved for 17/26 organs. The lowest DSC for strategy#2 (0.833 ± 0.177) was obtained for the pancreas, whereas DSC > 0.90 was achieved for 13/19 of the organs. For all mutual organs included in subset #1 and subset #2, our model outperformed the TotalSegmentator models in both strategies. In addition, our models outperformed the TotalSegmentator models on subset #3.

CONCLUSIONS

Our model was trained on images with significant variability from different databases, producing acceptable results on both pediatric and adult cases, making it well-suited for implementation in clinical setting.

Automated CT Measurement of Total Kidney Volume for Predicting Renal Function Decline after 177Lu Prostate-specific Membrane Antigen-I&T Radioligand Therapy

Background Lutetium 177 (177Lu) prostate-specific membrane antigen (PSMA) radioligand therapy is a novel treatment option for metastatic castration-resistant prostate cancer. Evidence suggests nephrotoxicity is a delayed adverse effect in a considerable proportion of patients. Purpose To identify predictive markers for clinically significant deterioration of renal function in patients undergoing 177Lu-PSMA-I&T radioligand therapy. Materials and Methods This retrospective study analyzed patients who underwent at least four cycles of 177Lu-PSMA-I&T therapy between December 2015 and May 2022. Total kidney volume (TKV) at 3 and 6 months after treatment was extracted from CT images using TotalSegmentator, a deep learning segmentation model based on the nnU-Net framework. A decline in estimated glomerular filtration rate (eGFR) of 30% or greater was defined as clinically significant, indicating a higher risk of end-stage renal disease. Two-sided t tests and Mann-Whitney U tests were used to compare baseline nephrotoxic risk factors, changes in eGFR and TKV, prior treatments, and the number of 177Lu-PSMA-I&T cycles between patients with and without clinically significant eGFR decline at 12 months. Threshold values to differentiate between these two patient groups were identified using receiver operating characteristic curve analysis and the Youden index. Results A total of 121 patients (mean age, 76 years ± 7 [SD]) who underwent four or more cycles of 177Lu-PSMA-I&T therapy with 12 months of follow-up were included. A 10% or greater decrease in TKV at 6 months predicted 30% or greater eGFR decline at 12 months (area under the receiver operating characteristic curve, 0.90 [95% CI: 0.85, 0.96]; P < .001), surpassing other parameters. Baseline risk factors (ρ = 0.01; P = .88), prior treatments (ρ = -0.06; P = .50), and number of 177Lu-PSMA-I&T cycles (ρ = 0.08; P = .36) did not correlate with relative eGFR percentage decrease at 12 months. Conclusion Automated TKV assessment on standard-of-care CT images predicted deterioration of renal function 12 months after 177Lu-PSMA-I&T therapy initiation in metastatic castration-resistant prostate cancer. Its better performance than early relative eGFR change highlights its potential as a noninvasive marker when treatment decisions are pending. © RSNA, 2025 Supplemental material is available for this article.

A machine learning-based pipeline for multi-organ/tissue patient-specific radiation dosimetry in CT

OBJECTIVES

To develop a machine learning-based pipeline for multi-organ/tissue personalized radiation dosimetry in CT.

MATERIALS AND METHODS

For the study, 95 chest CT scans and 85 abdominal CT scans were collected retrospectively. For each CT scan, a personalized Monte Carlo (MC) simulation was carried out. The produced 3D dose distributions and the respective CT examinations were utilized for the development of organ/tissue-specific dose prediction deep neural networks (DNNs). A pipeline that integrates a robust open-source organ segmentation tool with the dose prediction DNNs was developed for the automatic estimation of radiation doses for 30 organs/tissues including sub-volumes of the heart and lungs. The accuracy and time efficiency of the presented methodology was assessed. Statistical analysis (t-tests) was conducted to determine if the differences between the ground truth organ/tissue radiation dose estimates and the respective dose predictions were significant.

RESULTS

The lowest median percentage differences between MC-derived organ/tissue doses and DNN dose predictions were observed for the lung vessels (4.3%), small bowel (4.7%), pulmonary artery (4.7%), and colon (5.2%), while the highest differences were observed for the right lung's upper lobe (13.3%), spleen (13.1%), pancreas (12.1%), and stomach (11.6%). Statistical analysis showed that the differences were not significant (p-value > 0.18). Furthermore, the mean inference time, regarding the validation cohort, of the developed methodology was 77.0 ± 11.0 s.

CONCLUSION

The proposed workflow enables fast and accurate organ/tissue radiation dose estimations. The developed algorithms and dose prediction DNNs are publicly available ( https://github.com/eltzanis/multi-structure-CT-dosimetry ).

CLINICAL RELEVANCE STATEMENT

The accuracy and time efficiency of the developed pipeline compose a useful tool for personalized dosimetry in CT. By adopting the proposed workflow, institutions can utilize an automated pipeline for patient-specific dosimetry in CT.

KEY POINTS

Personalized dosimetry is ideal, but is time-consuming. The proposed pipeline composes a tool for facilitating patient-specific CT dosimetry in routine clinical practice. The developed workflow integrates a robust open-source segmentation tool with organ/tissue-specific dose prediction neural networks.

A survey on deep learning in medical image registration: New technologies, uncertainty, evaluation metrics, and beyond

Deep learning technologies have dramatically reshaped the field of medical image registration over the past decade. The initial developments, such as regression-based and U-Net-based networks, established the foundation for deep learning in image registration. Subsequent progress has been made in various aspects of deep learning-based registration, including similarity measures, deformation regularizations, network architectures, and uncertainty estimation. These advancements have not only enriched the field of image registration but have also facilitated its application in a wide range of tasks, including atlas construction, multi-atlas segmentation, motion estimation, and 2D-3D registration. In this paper, we present a comprehensive overview of the most recent advancements in deep learning-based image registration. We begin with a concise introduction to the core concepts of deep learning-based image registration. Then, we delve into innovative network architectures, loss functions specific to registration, and methods for estimating registration uncertainty. Additionally, this paper explores appropriate evaluation metrics for assessing the performance of deep learning models in registration tasks. Finally, we highlight the practical applications of these novel techniques in medical imaging and discuss the future prospects of deep learning-based image registration.

Comparison of Radiation Dose and Image Quality in Pediatric Abdominopelvic Photon-Counting Versus Energy-Integrating Detector CT

OBJECTIVE

Adoption of abdominal photon counting detector CT (PCD-CT) into clinical pediatric CT practice requires evidence that it provides diagnostic images at acceptable radiation doses. Thus, this study aimed to compare radiation dose and image quality of PCD-CT and conventional energy-integrating detector CT (EID-CT) in pediatric abdominopelvic CT.

MATERIALS AND METHODS

This institutional review board-approved retrospective study included 147 children (median age 8.5 y; 80 boys, 67 girls) who underwent clinically indicated contrast-enhanced abdominopelvic PCD-CT between October 1, 2022 and April 30, 2023 and 147 children (median age 8.5 y; 74 boys, 73 girls) who underwent EID-CT between July 1, 2021 and January 1, 2022. Patients in the 2 groups were matched by age and effective diameter. Radiation dose parameters (CT dose index volume, CTDIvol; dose length product, DLP; size-specific dose estimate, SSDE) were recorded. In a subset of 25 matched pairs, subjective image quality was assessed on a scale of 1 to 4 (1=highest quality), and liver attenuation, dose-normalized noise, and contrast-to-noise ratio (CNR) were measured. Groups were compared using parametric and/or nonparametric testing.

RESULTS

Among the 147 matched pairs, there were no significant differences in sex (P=0.576), age (P=0.084), or diameter (P=0.668). PCD-CT showed significantly lower median CTDIvol, DLP, and SSDE (1.6 mGy, 63.8 mGy-cm, 3.1 mGy) compared with EID-CT (3.7 mGy, 155.3 mGy-cm, 6.0 mGy) (P<0.001). In the subset of 25 patients, PCD-CT and EID-CT showed no significant difference in overall image quality for reader 1 (1.0 vs. 1.0, P=0.781) or reader 2 (1.0 vs. 1.0, P=0.817), or artifacts for reader 1 (1.0 vs. 1.0, P=0.688) or reader 2 (1.0 vs. 1.0, P=0.219). After normalizing for radiation dose, image noise was significantly lower with PCD-CT (P<0.001), while CNR in the liver (P=0.244) and portal vein (P=0.079) were comparable to EID-CT.

CONCLUSION

Abdominopelvic PCD-CT in children significantly reduces radiation dose while maintaining subjective image quality, and accounting for dose levels, has the potential to lower image noise and achieve comparable CNR to EID-CT. These data expand understanding of the capabilities of PCD-CT and support its routine use in children.

CompositIA: an open-source automated quantification tool for body composition scores from thoraco-abdominal CT scans

BACKGROUND

Body composition scores allow for quantifying the volume and physical properties of specific tissues. However, their manual calculation is time-consuming and prone to human error. This study aims to develop and validate CompositIA, an automated, open-source pipeline for quantifying body composition scores from thoraco-abdominal computed tomography (CT) scans.

METHODS

A retrospective dataset of 205 contrast-enhanced thoraco-abdominal CT examinations was used for training, while 54 scans from a publicly available dataset were used for independent testing. Two radiology residents performed manual segmentation, identifying the centers of the L1 and L3 vertebrae and segmenting the corresponding axial slices. MultiResUNet was used to identify CT slices intersecting the L1 and L3 vertebrae, and its performance was evaluated using the mean absolute error (MAE). Two U-nets were used to segment the axial slices, with performance evaluated through the volumetric Dice similarity coefficient (vDSC). CompositIA's performance in quantifying body composition indices was assessed using mean percentage relative error (PRE), regression, and Bland-Altman analyses.

RESULTS

On the independent dataset, CompositIA achieved a MAE of about 5 mm in detecting slices intersecting the L1 and L3 vertebrae, with a MAE < 10 mm in at least 85% of cases and a vDSC greater than 0.85 in segmenting axial slices. Regression and Bland-Altman analyses demonstrated a strong linear relationship and good agreement between automated and manual scores (p values < 0.001 for all indices), with mean PREs ranging from 5.13% to 15.18%.

CONCLUSION

CompositIA facilitated the automated quantification of body composition scores, achieving high precision in independent testing.

RELEVANCE STATEMENT

CompositIA is an automated, open-source pipeline for quantifying body composition indices from CT scans, simplifying clinical assessments, and expanding their applicability.

KEY POINTS

Manual body composition assessment from CTs is time-consuming and prone to errors. CompositIA was trained on 205 CT scans and tested on 54 scans. CompositIA demonstrated mean percentage relative errors under 15% compared to manual indices. CompositIA simplifies body composition assessment through an artificial intelligence-driven and open-source pipeline.

Generative artificial intelligence enables the generation of bone scintigraphy images and improves generalization of deep learning models in data-constrained environments

PURPOSE

Advancements of deep learning in medical imaging are often constrained by the limited availability of large, annotated datasets, resulting in underperforming models when deployed under real-world conditions. This study investigated a generative artificial intelligence (AI) approach to create synthetic medical images taking the example of bone scintigraphy scans, to increase the data diversity of small-scale datasets for more effective model training and improved generalization.

METHODS

We trained a generative model on 99mTc-bone scintigraphy scans from 9,170 patients in one center to generate high-quality and fully anonymized annotated scans of patients representing two distinct disease patterns: abnormal uptake indicative of (i) bone metastases and (ii) cardiac uptake indicative of cardiac amyloidosis. A blinded reader study was performed to assess the clinical validity and quality of the generated data. We investigated the added value of the generated data by augmenting an independent small single-center dataset with synthetic data and by training a deep learning model to detect abnormal uptake in a downstream classification task. We tested this model on 7,472 scans from 6,448 patients across four external sites in a cross-tracer and cross-scanner setting and associated the resulting model predictions with clinical outcomes.

RESULTS

The clinical value and high quality of the synthetic imaging data were confirmed by four readers, who were unable to distinguish synthetic scans from real scans (average accuracy: 0.48% [95% CI 0.46-0.51]), disagreeing in 239 (60%) of 400 cases (Fleiss' kappa: 0.18). Adding synthetic data to the training set improved model performance by a mean (± SD) of 33(± 10)% AUC (p < 0.0001) for detecting abnormal uptake indicative of bone metastases and by 5(± 4)% AUC (p < 0.0001) for detecting uptake indicative of cardiac amyloidosis across both internal and external testing cohorts, compared to models without synthetic training data. Patients with predicted abnormal uptake had adverse clinical outcomes (log-rank: p < 0.0001).

CONCLUSIONS

Generative AI enables the targeted generation of bone scintigraphy images representing different clinical conditions. Our findings point to the potential of synthetic data to overcome challenges in data sharing and in developing reliable and prognostic deep learning models in data-limited environments.

OPPORTUNISTIC ASSESSMENT OF CARDIOVASCULAR RISK USING AI-DERIVED STRUCTURAL AORTIC AND CARDIAC PHENOTYPES FROM NON-CONTRAST CHEST COMPUTED TOMOGRAPHY

Background

Primary prevention of cardiovascular disease relies on accurate risk assessment using scores such as the Pooled Cohort Equations (PCE) and PREVENT. However, necessary input variables for these scores are often unavailable in the electronic health record (EHR), and information from routinely collected data (e.g., non-contrast chest CT) may further improve performance. Here, we test whether a risk prediction model based on structural features of the heart and aorta from chest CT has added value to existing clinical algorithms for predicting major adverse cardiovascular events (MACE).

Methods

We developed a LASSO model to predict fatal MACE over 12 years of follow-up using structural radiomics features describing cardiac chamber and aorta segmentations from 13,437 lung cancer screening chest CTs from the National Lung Screening Trial. We compared this radiomics model to the PCE and PREVENT scores in an external testing set of 4,303 individuals who had a chest CT at a Mass General Brigham site and had no history of diabetes, prior MACE, or statin treatment. Discrimination for incident MACE was assessed using the concordance index. We used a binary threshold to determine MACE rates in patients who were statin-eligible or ineligible by the PCE/PREVENT scores (≥7.5% risk) or the radiomics score (≥5.0% risk). Results were stratified by whether all variables were available to calculate the PCE or PREVENT scores.

Results

In the external testing set (n = 4,303; mean age 61.5 ± 9.3 years; 47.1% male), 8.0% had incident MACE over a median 5.1 years of follow-up. The radiomics risk score significantly improved discrimination beyond the PCE (c-index 0.653 vs. 0.567, p < 0.001) and performed similarly in individuals who were missing inputs. Those statin-eligible by both the radiomics and PCE scores had a 2.6-fold higher incidence of MACE than those eligible by the PCE score alone (29.5 [20.5, 39.1] vs. 11.2 [8.0, 14.4] events per 1,000 person-years among PCE-eligible individuals). In patients missing inputs, incident MACE rates were 1.8-fold higher in those statin-eligible by the radiomics score than those statin-ineligible (29.5 [21.9, 37.6] vs. 16.7 [14.3, 19.0] events per 1000 person-years). Similar results were found when comparing to the PREVENT score. Left ventricular volume and short axis length were most predictive of myocardial infarction, while left atrial sphericity and surface-to-volume ratio were most predictive of stroke.

Conclusions

Based on a single chest CT, a cardiac shape-based risk prediction model predicted cardiovascular events beyond clinical algorithms and demonstrated similar performance in patients who were missing inputs to standard cardiovascular risk calculators. Patients at high-risk by the radiomics score may benefit from intensified primary prevention (e.g., statin prescription).

Multilevel support-assisted prototype optimization network for few-shot medical segmentation of lung lesions

Medical image annotation is scarce and costly. Few-shot segmentation has been widely used in medical image from only a few annotated examples. However, its research on lesion segmentation for lung diseases is still limited, especially for pulmonary aspergillosis. Lesion areas usually have complex shapes and blurred edges. Lesion segmentation requires more attention to deal with the diversity and uncertainty of lesions. To address this challenge, we propose MSPO-Net, a multilevel support-assisted prototype optimization network designed for few-shot lesion segmentation in computerized tomography (CT) images of lung diseases. MSPO-Net learns lesion prototypes from low-level to high-level features. Self-attention threshold learning strategy can focus on the global information and obtain an optimal threshold for CT images. Our model refines prototypes through a support-assisted prototype optimization module, adaptively enhancing their representativeness for the diversity of lesions and adapting to the unseen lesions better. In clinical examinations, CT is more practical than X-rays. To ensure the quality of our work, we have established a small-scale CT image dataset for three lung diseases and annotated by experienced doctors. Experiments demonstrate that MSPO-Net can improve segmentation performance and robustness of lung disease lesion. MSPO-Net achieves state-of-the-art performance in both single and unseen lung disease segmentation, indicating its potentiality to reduce doctors' workload and improve diagnostic accuracy. This research has certain clinical significance. Code is available at https://github.com/Tian-Yuan-ty/MSPO-Net .

End-to-end prediction of clinical outcomes in head and neck squamous cell carcinoma with foundation model-based multiple instance learning

Background

Foundation models (FMs) show promise in medical AI by learning flexible features from large datasets, potentially surpassing handcrafted radiomics. Outcome prediction of head and neck squamous cell carcinomas (HNSCC) with FMs using routine imaging remains unexplored.

Purpose

To evaluate end-to-end FM-based multiple instance learning (MIL) for 2-year overall survival (OS), locoregional control (LRC), and freedom from distant metastasis (FFDM) prediction and risk group stratification using pretreatment CT scans in HNSCC.

Materials and Methods

We analyzed data of 2485 patients from three retrospective HNSCC cohorts (RADCURE, HN1, HN-PET-CT), treated between 2004 and 2017 with available pre-treatment CTs and primary gross tumor volume (GTVp) segmentations. The RADCURE cohort was split into training (n=1464) and test (N=606), with HN1 (n=131) and HN-PET-CT (n=284) as additional test cohorts. FM-based MIL models (2D, multiview and 3D) for 2-year endpoint prediction and risk stratification wre evaluated based on area under the receiver operator curve (AUROC) and Kaplan-Meier (KM) with hazard ratios (HR), compared with radiomics and assessed for multimodal enhancement with clinical baselines.

Results

2D MIL models achieved 2-year test AUROCs of 0.75-0.84 (OS), 0.66-0.75 (LRC) and 0.71-0.78 (FFDM), outperforming multiview and 3D MIL (AUROCs: 0.50-0.77, p≥0.15) and comparable or superior to radiomics (AUROCs: 0.64-0.74, p≥0.012). Significant stratification was observed (HRs: 2.14-4.77, p≤0.039). Multimodal enhancement of 2-year OS/FFDM (AUROCs: 0.82-0.87, p≤0.018) was observed for patients without human papilloma virus positive (HPV+) tumors.

Conclusion

FM-based MIL demonstrates promise in HNSCC risk prediction, showing similar or superior performance to radiomics and enhancing clinical baselines in non-HPV+ patients.

Usefulness of Radiomics and Kidney Volume Based on Non-Enhanced Computed Tomography in Chronic Kidney Disease: Initial Report

INTRODUCTION

Chronic kidney disease (CKD) is classified according to the estimated glomerular filtration rate (eGFR), but kidney volume (KV) can also provide meaningful information. Very few radiomics (RDX) studies on CKD have utilized computed tomography (CT). This study aimed to determine whether non-enhanced computed tomography (NECT)-based RDX can be useful in evaluation of patients with CKD and to compare it with KV.

METHODS

The NECT scans of 64 subjects with impaired kidney function (defined as <60 mL/min/1.73 m2) and 60 controls with normal kidney function were retrospectively analyzed. Kidney segmentations, volume measurements, and RDX features extraction were performed. Machine-learning models using RDX were constructed to classify the kidneys as having structural markers of impaired or normal function.

RESULTS

The median KV in the impaired kidney function group was 114.83 mL vs. 159.43 mL (p < 0.001) in the control group. There was a statistically significant strong positive correlation between KV and eGFR (rs = 0.579, p < 0.001) and a strong negative correlation between KV and serum creatinine level (rs = -0.514, p < 0.001). The KV-based models achieved the best area under the curve (AUC) of 0.746, whereas the RDX-based models achieved the best AUC of 0.878.

CONCLUSIONS

RDX can be useful in identifying patients with impaired kidney function on NECT. RDX-based models outperformed KV-based models. RDX has the potential to identify patients with a higher risk of CKD based on imaging, which, as we believe, can indirectly support clinical decision-making.

Automated Segmentation of Trunk Musculature with a Deep CNN Trained from Sparse Annotations in Radiation Therapy Patients with Metastatic Spine Disease

Purpose

Given the high prevalence of vertebral fractures post-radiotherapy in patients with metastatic spine disease, accurate and rapid muscle segmentation could support efforts to quantify muscular changes due to disease or treatment and enable biomechanical modeling for assessments of vertebral loading to improve personalized evaluation of vertebral fracture risk. This study presents a deep-learning approach for segmenting the complete volume of the trunk muscles from clinical CT images trained using sparsely annotated data.

Materials and Methods

we extracted 2,009 axial CT images at the midpoint of each vertebral level (T4 to L4) from clinical CT of 148 cancer patients. The key extensor and flexor muscles (up to 8 muscles per side) were manually contoured and labeled per image in the thoracic and lumbar regions. We first trained a 2D nnU-Net deep-learning model on these labels to segment key extensor and flexor muscles. Using these sparse annotations per spine, we trained the model to segment each muscle's entire 3D volume.

Results

The proposed method achieved comparable performance to manual segmentations, as assessed by expert radiologists, with a mean Dice score above 0.769. Significantly, the model drastically reduced segmentation time, from 4.3-6.5 hours for manual segmentation of 14 single axial CT images to approximately 1 minute for segmenting the complete thoracic-abdominal 3D volume.

Conclusion

The approach demonstrates high potential for automating 3D muscle segmentation, significantly reducing the manual intervention required for generating musculoskeletal models, and could be instrumental in enhancing clinical decision-making and patient care in radiation oncology.