A survey on deep learning in medical image analysis

Coronary artery segmentation framework based on three types of U-Net and voting ensembles

Coronary artery (CA) segmentation is critical for enabling disease diagnosis. However, the structural complexity and extensive branching of CAs may cause the segmentation outcomes of existing methods to exhibit discontinuities and considerable pseudo-CA regions. Therefore, we propose a voting-based ensemble segmentation framework based on three U-Net types to capture CA structural features from global and local perspectives. The lightweight U-Net performs direct segmentation on CAs, helping to eliminate interferences from small connected regions during segmentation and preserve global information. Patch-based and multi-slice U-Nets provide superior local partition information. Finally, a voting-based strategy is adopted to ensemble the segmentation results for the three models to obtain the final result. Our proposed segmentation framework performs well, attaining a Dice score of 82.31% on a large dataset.

ECn-MultiBSTM: multiclass epileptic seizure classification using electro cetacean optimized bidirectional long short-term memory model

Multiclass epileptic seizureclassification aims to identify and categorize different epileptic seizure types like a non-epileptic seizure, epileptic interictal seizure, and epileptic ictal seizurein individuals based on Electroencephalography (EEG) signal characteristics. Multi-class seizure classification requires recognizing various seizure forms and patterns, which can be challenging due to noise and high variability patterns in EEG signals. Existing models face limitations such as difficulty in handling the complex and dynamic nature of seizure patterns, poor generalization to unseen data, and sensitivity to noise and artifacts, all of which impact classification accuracy and reliability. To address these issues, the Electro Cetacean Optimization based Multi Bidirectional Long Short-Term Memory (ECn-MultiBSTM) model is proposed. The BiLSTM modelis utilized for feature extraction, which captures sequential data by processing data in both forward and backward directions. This bidirectional approach enables the model to identify subtle patterns that distinguish various seizure types with higher accuracy. The ECn-MultiBSTM model also incorporates advanced Electro Cetacean optimizationtechniques that enhance its ability to search efficiently for optimal solutions.Through dynamic social coordination and rapid search strategies, the model fine-tunes its hyperparameters, ensuring improved performance and adaptability.The proposed ECn-MultiBSTM model significantly enhances multiclassseizure classification performance, achieving impressive metrics of 95.84% accuracy, 95.30% precision, 95.54% F1-score,0.94% MCC, 95.79% sensitivity, and 95.88% specificity when evaluated on the CHB-MIT SCALP EEG dataset.

LiDSCUNet++: A lightweight depth separable convolutional UNet++ for vertebral column segmentation and spondylosis detection

Accurate computer-aided diagnosis systems rely on precise segmentation of the vertebral column to assist physicians in diagnosing various disorders. However, segmenting spinal disks and bones becomes challenging in the presence of abnormalities and complex anatomical structures. While Deep Convolutional Neural Networks (DCNNs) achieve remarkable results in medical image segmentation, their performance is limited by data insufficiency and the high computational complexity of existing solutions. This paper introduces LiDSCUNet++, a lightweight deep learning framework based on depthwise-separable and pointwise convolutions integrated with UNet++ for vertebral column segmentation. The model segments vertebral anomalies from dog radiographs, and the results are further processed by YOLOv8 for automated detection of Spondylosis Deformans. LiDSCUNet++ delivers comparable segmentation performance while significantly reducing trainable parameters, memory usage, energy consumption, and computational time, making it an efficient and practical solution for medical image analysis.

Predicting progression-free survival in sarcoma using MRI-based automatic segmentation models and radiomics nomograms: a preliminary multicenter study

OBJECTIVES

Some sarcomas are highly malignant, associated with high recurrence despite treatment. This multicenter study aimed to develop and validate a radiomics signature to estimate sarcoma progression-free survival (PFS).

MATERIALS AND METHODS

The study retrospectively enrolled 202 consecutive patients with pathologically diagnosed sarcoma, who had pre-treatment axial fat-suppressed T2-weighted images (FS-T2WI), and included them in the ROI-Net model for training. Among them, 120 patients were included in the radiomics analysis, all of whom had pre-treatment axial T1-weighted and transverse FS-T2WI images, and were randomly divided into a development group (n = 96) and a validation group (n = 24). In the development cohort, Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression was used to develop the radiomics features for PFS prediction. By combining significant clinical features with radiomics features, a nomogram was constructed using Cox regression.

RESULTS

The proposed ROI-Net framework achieved a Dice coefficient of 0.820 (0.791-0.848). The radiomics signature based on 21 features could distinguish high-risk patients with poor PFS. Univariate Cox analysis revealed that peritumoral edema, metastases, and the radiomics score were associated with poor PFS and were included in the construction of the nomogram. The Radiomics-T1WI-Clinical model exhibited the best performance, with AUC values of 0.947, 0.907, and 0.924 at 300 days, 600 days, and 900 days, respectively.

CONCLUSION

The proposed ROI-Net framework demonstrated high consistency between its segmentation results and expert annotations. The radiomics features and the combined nomogram have the potential to aid in predicting PFS for patients with sarcoma.

Segmentation of the nasopalatine canal and detection of canal furcation status with artificial intelligence on cone-beam computed tomography images

OBJECTIVES

The nasopalatine canal (NPC) is an anatomical formation with varying morphology. NPC can be visualized using the cone-beam computed tomography (CBCT). Also, CBCT has been used in many studies on artificial intelligence (AI). The "You only look once" (YOLO) is an AI framework that stands out with its speed. This study compared the observer and AI regarding the NPC segmentation and assessment of the NPC furcation status in CBCT images.

METHODS

In this study, axial sections of 200 CBCT images were used. These images were labeled and evaluated for the absence or presence of the NPC furcation. These images were then divided into three; 160 images were used as the training dataset, 20 as the validation dataset, and 20 as the test dataset. The training was performed by making 800 epochs using the YOLOv5x-seg model.

RESULTS

Sensitivity, Precision, F1 score, IoU, mAP, and AUC values were determined for NPC detection, segmentation, and classification of the YOLOv5x-seg model. The values were found to be 0.9680, 0.9953, 0.9815, 0.9636, 0.7930, and 0.8841, respectively, for the group with the absence of the NPC furcation; and 0.9827, 0.9975, 0.9900, 0.9803, 0.9637, and 0.9510, for the group with the presence of the NPC furcation.

CONCLUSIONS

Our results showed that even when the YOLOv5x-seg model is trained with the NPC furcation and fewer datasets, it achieves sufficient prediction accuracy. The segmentation feature of the YOLOv5 algorithm, which is based on an object detection algorithm, has achieved quite successful results despite its recent development.

Automated quantification of T1 and T2 relaxation times in liver mpMRI using deep learning: a sequence-adaptive approach

OBJECTIVES

To evaluate a deep learning sequence-adaptive liver multiparametric MRI (mpMRI) assessment with validation in different populations using total and segmental T1 and T2 relaxation time maps.

METHODS

A neural network was trained to label liver segmental parenchyma and its vessels on noncontrast T1-weighted gradient-echo Dixon in-phase acquisitions on 200 liver mpMRI examinations. Then, 120 unseen liver mpMRI examinations of patients with primary sclerosing cholangitis or healthy controls were assessed by coregistering the labels to noncontrast and contrast-enhanced T1 and T2 relaxation time maps for optimization and internal testing. The algorithm was externally tested in a segmental and total liver analysis of previously unseen 65 patients with biopsy-proven liver fibrosis and 25 healthy volunteers. Measured relaxation times were compared to manual measurements using intraclass correlation coefficient (ICC) and Wilcoxon test.

RESULTS

Comparison of manual and deep learning-generated segmental areas on different T1 and T2 maps was excellent for segmental (ICC = 0.95 ± 0.1; p < 0.001) and total liver assessment (0.97 ± 0.02, p < 0.001). The resulting median of the differences between automated and manual measurements among all testing populations and liver segments was 1.8 ms for noncontrast T1 (median 835 versus 842 ms), 2.0 ms for contrast-enhanced T1 (median 518 versus 519 ms), and 0.3 ms for T2 (median 37 versus 37 ms).

CONCLUSION

Automated quantification of liver mpMRI is highly effective across different patient populations, offering excellent reliability for total and segmental T1 and T2 maps. Its scalable, sequence-adaptive design could foster comprehensive clinical decision-making.

RELEVANCE STATEMENT

The proposed automated, sequence-adaptive algorithm for total and segmental analysis of liver mpMRI may be co-registered to any combination of parametric sequences, enabling comprehensive quantitative analysis of liver mpMRI without sequence-specific training.

KEY POINTS

A deep learning-based algorithm automatically quantified segmental T1 and T2 relaxation times in liver mpMRI. The two-step approach of segmentation and co-registration allowed to assess arbitrary sequences. The algorithm demonstrated high reliability with manual reader quantification. No additional sequence-specific training is required to assess other parametric sequences. The DL algorithm has the potential to enhance individual liver phenotyping.

A strategy for the automatic diagnostic pipeline towards feature-based models: a primer with pleural invasion prediction from preoperative PET/CT images

BACKGROUND

This study aims to explore the feasibility to automate the application process of nomograms in clinical medicine, demonstrated through the task of preoperative pleural invasion prediction in non-small cell lung cancer patients using PET/CT imaging.

RESULTS

The automatic pipeline involves multimodal segmentation, feature extraction, and model prediction. It is validated on a cohort of 1116 patients from two medical centers. The performance of the feature-based diagnostic model outperformed both the radiomics model and individual machine learning models. The segmentation models for CT and PET images achieved mean dice similarity coefficients of 0.85 and 0.89, respectively, and the segmented lung contours showed high consistency with the actual contours. The automatic diagnostic system achieved an accuracy of 0.87 in the internal test set and 0.82 in the external test set, demonstrating comparable overall diagnostic performance to the human-based diagnostic model. In comparative analysis, the automatic diagnostic system showed superior performance relative to other segmentation and diagnostic pipelines.

CONCLUSIONS

The proposed automatic diagnostic system provides an interpretable, automated solution for predicting pleural invasion in non-small cell lung cancer.

Deep learning without borders: recent advances in ultrasound image classification for liver diseases diagnosis

INTRODUCTION

Liver diseases are among the top global health burdens. Recently, there has been an increasing significance of diagnostics without discomfort to the patient; among them, ultrasound is the most used. Deep learning, in particular convolutional neural networks, has revolutionized the classification of liver diseases by automatically performing some specific analyses of difficult images.

AREAS COVERED

This review summarizes the progress that has been made in deep learning techniques for the classification of liver diseases using ultrasound imaging. It evaluates various models from CNNs to their hybrid versions, such as CNN-Transformer, for detecting fatty liver, fibrosis, and liver cancer, among others. Several challenges in the generalization of data and models across a different clinical environment are also discussed.

EXPERT OPINION

Deep learning has great prospects for automatic diagnosis of liver diseases. Most of the models have performed with high accuracy in different clinical studies. Despite this promise, challenges relating to generalization have remained. Future hardware developments and access to quality clinical data continue to further improve the performance of these models and ensure their vital role in the diagnosis of liver diseases.

Quantum-enhanced intelligent system for personalized adaptive radiotherapy dose estimation

This research introduces a novel quantum-enhanced intelligent system tailored for personalized adaptive radiotherapy dose estimation. The system efficiently models radiation transport and predicts patient-specific dose distributions by integrating quantum algorithms, deep learning, and Monte Carlo simulations. Quantum-enhanced Monte Carlo simulations, employing algorithms such as Harrow-Hassidim-Lloyd (HHL) and Variational Quantum Eigensolver (VQE), achieve computational speedups of 8-15 times compared to classical methods while maintaining high accuracy. The deep learning architecture leverages convolutional and recurrent neural networks to capture complex anatomical and dosimetric patterns. Validation on simulated datasets demonstrates a 50-70% reduction in mean absolute error and 2-3% improvements in gamma index metrics compared to conventional approaches. Dose-volume histogram analysis further highlights enhanced Dice coefficients and reduced Hausdorff distances. These advancements underscore the potential for precise, efficient, and clinically relevant dose estimations, paving the way for improved outcomes in personalized adaptive radiotherapy.

Potential Use of a New Energy Vision (NEV) Camera for Diagnostic Support of Carpal Tunnel Syndrome: Development of a Decision-Making Algorithm to Differentiate Carpal Tunnel-Affected Hands from Controls

Introduction: Carpal Tunnel Syndrome (CTS) is a prevalent neuropathy requiring accurate, non-invasive diagnostics to minimize patient burden. This study evaluates the New Energy Vision (NEV) camera, an RGB-based multispectral imaging tool, to detect CTS through skin texture and color analysis, developing a machine learning algorithm to distinguish CTS-affected hands from controls. Methods: A two-part observational study included 103 participants (50 controls, 53 CTS patients) in Part 1, using NEV camera images to train a Support Vector Machine (SVM) classifier. Part 2 compared median nerve-damaged (MED) and ulnar nerve-normal (ULN) palm areas in 32 CTS patients. Validations included nerve conduction tests (NCT), Semmes-Weinstein monofilament testing (SWMT), and Boston Carpal Tunnel Questionnaire (BCTQ). Results: The SVM classifier achieved 93.33% accuracy (confusion matrix: [[14, 1], [1, 14]]), with 81.79% cross-validation accuracy. Part 2 identified significant differences (p < 0.05) in color proportions (e.g., red_proportion) and Haralick texture features between MED and ULN areas, corroborated by BCTQ and SWMT. Conclusions: The NEV camera, leveraging multispectral imaging, offers a promising non-invasive CTS diagnostic tool using detection of nerve-related skin changes. Further validation is needed for clinical adoption.

Accuracy of a deep neural network for automated pulmonary embolism detection on dedicated CT pulmonary angiograms

PURPOSE

To assess the performance of a Deep Neural Network (DNN)-based prototype algorithm for automated PE detection on CTPA scans.

METHODS

Patients who had previously undergone CTPA with three different systems (SOMATOM Force, go.Top, and Definition AS; Siemens Healthineers, Forchheim, Germany) because of suspected PE from September 2022 to January 2023 were retrospectively enrolled in this study (n = 1,000, 58.8 % women). For detailed evaluation, all PE were divided into three location-based subgroups: central arteries, lobar branches, and peripheral regions. Clinical reports served as ground truth. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were determined to evaluate the performance of DNN-based PE detection.

RESULTS

Cases were excluded due to incomplete data (n = 32), inconclusive report (n = 17), insufficient contrast detected in the pulmonary trunk (n = 40), or failure of the preprocessing algorithms (n = 8). Therefore, the final cohort included 903 cases with a PE prevalence of 12 % (n = 110). The model achieved a sensitivity, specificity, PPV, and NPV of 84.6, 95.1, 70.5, and 97.8 %, respectively, and delivered an overall accuracy of 93.8 %. Among the false positive cases (n = 39), common sources of error included lung masses, pneumonia, and contrast flow artifacts. Common sources of false negatives (n = 17) included chronic and subsegmental PEs.

CONCLUSION

The proposed DNN-based algorithm provides excellent performance for the detection of PE, suggesting its potential utility to support radiologists in clinical reading and exam prioritization.

Deep learning-driven multi-class classification of brain strokes using computed tomography: A step towards enhanced diagnostic precision

OBJECTIVE

To develop and validate deep learning models leveraging CT imaging for the prediction and classification of brain stroke conditions, with the potential to enhance accuracy and support clinical decision-making.

MATERIALS AND METHODS

This retrospective, bi-center study included data from 250 patients, with a dataset of 8186 CT images collected from 2017 to 2022. Two AI models were developed using the Expanded ResNet101 deep learning framework as a two-step model. Model performance was evaluated using confusion matrices, supplemented by external validation with an independent dataset. External validation was conducted by an expert and two external members. Overall accuracy, confidence intervals, Cohen's Kappa value, and McNemar's test P-values were calculated.

RESULTS

A total of 8186 CT images were incorporated, with 6386 images used for the training and 900 datasets for testing and validation in Model 01. Further, 1619 CT images were used for training and 600 datasets for testing and validation in Model 02. The average accuracy, precision, and F1 score for both models were assessed: Model 01 achieved 99.6 %, 99.4 %, and 99.6 % respectively, whereas Model 02 achieved 99.2 %, 98.8 %, and 99.1 %. The external validation accuracies were 78.6 % (95 % CI: 0.73,0.83; P < 0.001) and 60.2 % (95 % CI: 0.48,0.70; P < 0.001) for Models 01 and 02 respectively, as evaluated by the expert.

CONCLUSION

Deep learning models demonstrated high accuracy, precision, and F1 scores in predicting outcomes for brain stroke patients. With larger cohort and diverse radiologic mimics, these models could support clinicians in prognosis and decision-making.

Mitosis detection and classification for breast cancer diagnosis: What we know and what is next

Breast cancer is the second most deadly malignancy in women, behind lung cancer. Despite significant improvements in medical research, breast cancer is still accurately diagnosed with histological analysis. During this procedure, pathologists examine a physical sample for the presence of mitotic cells, or dividing cells. However, the high resolution of histopathology images and the difficulty of manually detecting tiny mitotic nuclei make it particularly challenging to differentiate mitotic cells from other types of cells. Numerous studies have addressed the detection and classification of mitosis, owing to increasing capacity and developments in automated approaches. The combination of machine learning and deep learning techniques has greatly revolutionized the process of identifying mitotic cells by offering automated, precise, and efficient solutions. In the last ten years, several pioneering methods have been presented, advancing towards practical applications in clinical settings. Unlike other forms of cancer, breast cancer and gliomas are categorized according to the number of mitotic divisions. Numerous papers have been published on techniques for identifying mitosis due to easy access to datasets and open competitions. Convolutional neural networks and other deep learning architectures can precisely identify mitotic cells, significantly decreasing the amount of labor that pathologists must perform. This article examines the techniques used over the past decade to identify and classify mitotic cells in histologically stained breast cancer hematoxylin and eosin images. Furthermore, we examine the benefits of current research techniques and predict forthcoming developments in the investigation of breast cancer mitosis, specifically highlighting machine learning and deep learning.

BCT-Net: semantic-guided breast cancer segmentation on BUS

Accurately and swiftly segmenting breast tumors is significant for cancer diagnosis and treatment. Ultrasound imaging stands as one of the widely employed methods in clinical practice. However, due to challenges such as low contrast, blurred boundaries, and prevalent shadows in ultrasound images, tumor segmentation remains a daunting task. In this study, we propose BCT-Net, a network amalgamating CNN and transformer components for breast tumor segmentation. BCT-Net integrates a dual-level attention mechanism to capture more features and redefines the skip connection module. We introduce the utilization of a classification task as an auxiliary task to impart additional semantic information to the segmentation network, employing supervised contrastive learning. A hybrid objective loss function is proposed, which combines pixel-wise cross-entropy, binary cross-entropy, and supervised contrastive learning loss. Experimental results demonstrate that BCT-Net achieves high precision, with Pre and DSC indices of 86.12% and 88.70%, respectively. Experiments conducted on the BUSI dataset of breast ultrasound images manifest that this approach exhibits high accuracy in breast tumor segmentation.

The integration of artificial intelligence into clinical medicine: Trends, challenges, and future directions

BACKGROUND AND OBJECTIVES

AI has emerged as a transformative force in clinical medicine, changing the diagnosis, treatment, and management of patients. Tools have been derived for working with ML, DL, and NLP algorithms to analyze large complex medical datasets with unprecedented accuracy and speed, thereby improving diagnostic precision, treatment personalization, and patient care outcomes. For example, CNNs have dramatically improved the accuracy of medical imaging diagnoses, and NLP algorithms have greatly helped extract insights from unstructured data, including EHRs. However, there are still numerous challenges that face AI integration into clinical workflows, including data privacy, algorithmic bias, ethical dilemmas, and problems with the interpretability of "black-box" AI models. These barriers have thus far prevented the widespread application of AI in health care, and its possible trends, obstacles, and future implications are necessary to be systematically explored. The purpose of this paper is, therefore, to assess the current trends in AI applications in clinical medicine, identify those obstacles that are hindering adoption, and identify possible future directions. This research hopes to synthesize evidence from other peer-reviewed articles to provide a more comprehensive understanding of the role that AI plays to advance clinical practices, improve patient outcomes, or enhance decision-making.

METHODS

A systematic review was done according to the PRISMA guidelines to explore the integration of Artificial Intelligence in clinical medicine, including trends, challenges, and future directions. PubMed, Cochrane Library, Web of Science, and Scopus databases were searched for peer-reviewed articles from 2014 to 2024 with keywords such as "Artificial Intelligence in Medicine," "AI in Clinical Practice," "Machine Learning in Healthcare," and "Ethical Implications of AI in Medicine." Studies focusing on AI application in diagnostics, treatment planning, and patient care reporting measurable clinical outcomes were included. Non-clinical AI applications and articles published before 2014 were excluded. Selected studies were screened for relevance, and then their quality was critically appraised to synthesize data reliably and rigorously.

RESULTS

This systematic review includes the findings of 8 studies that pointed out the transformational role of AI in clinical medicine. AI tools, such as CNNs, had diagnostic accuracy more than the traditional methods, particularly in radiology and pathology. Predictive models efficiently supported risk stratification, early disease detection, and personalized medicine. Despite these improvements, significant hurdles, including data privacy, algorithmic bias, and resistance from clinicians regarding the "black-box" nature of AI, had yet to be surmounted. XAI has emerged as an attractive solution that offers the promise to enhance interpretability and trust. As a whole, AI appeared promising in enhancing diagnostics, treatment personalization, and clinical workflows by dealing with systemic inefficiencies.

CONCLUSION

The transformation potential of AI in clinical medicine can transform diagnostics, treatment strategies, and efficiency. Overcoming obstacles such as concerns about data privacy, the danger of algorithmic bias, and difficulties with interpretability may pave the way for broader use and facilitate improvement in patient outcomes while transforming clinical workflows to bring sustainability into healthcare delivery.

A unified approach to medical image segmentation by leveraging mixed supervision and self and transfer learning (MIST)

Medical image segmentation is important for quantitative disease diagnosis and treatment but relies on accurate pixel-wise labels, which are costly, time-consuming, and require domain expertise. This work introduces MIST (MIxed supervision, Self, and Transfer learning) to reduce manual labeling in medical image segmentation. A small set of cases was manually annotated ("strong labels"), while the rest used automated, less accurate labels ("weak labels"). Both label types trained a dual-branch network with a shared encoder and two decoders. Self-training iteratively refined weak labels, and transfer learning reduced computational costs by freezing the encoder and fine-tuning the decoders. Applied to segmenting muscle, subcutaneous, and visceral adipose tissue, MIST used only 100 manually labeled slices from 20 CT scans to generate accurate labels for all slices of 102 internal scans, which were then used to train a 3D nnU-Net model. Using MIST to update weak labels significantly improved nnU-Net segmentation accuracy compared to training directly on strong and weak labels. Dice similarity coefficient (DSC) increased for muscle (89.2 ± 4.3% to 93.2 ± 2.1%), subcutaneous (75.1 ± 14.4% to 94.2 ± 2.8%), and visceral adipose tissue (66.6 ± 16.4% to 77.1 ± 19.0% ) on an internal dataset (p<.05). DSC improved for muscle (80.5 ± 6.9% to 86.6 ± 3.9%) and subcutaneous adipose tissue (61.8 ± 12.5% to 82.7 ± 11.1%) on an external dataset (p<.05). MIST reduced the annotation burden by 99%, enabling efficient, accurate pixel-wise labeling for medical image segmentation. Code is available at https://github.com/rsummers11/NIH_CADLab_Body_Composition.

Single-shot T2 mapping via multi-slice information sharing based on switching modulation patterns multiple overlapping-echo detachment imaging

BACKGROUND

Quantitative magnetic resonance imaging (qMRI) offers reliable biomarkers in clinic. Nevertheless, most qMRI methods are time-consuming and sensitive to motion. Single-shot multiple overlapping-echo detachment (MOLED) magnetic resonance imaging can deliver robust T2 mapping in about 100 ms with high motion tolerance. However, its spatial resolution is relatively low due to the limitations of signal-to-noise ratio (SNR) and echo-train length. At the mean time, the number of echoes with different evolution times collected is usually limited, which is not conducive to T2 mapping in high accuracy.

PURPOSE

To propose a novel method to improve the spatial resolution and quantification accuracy of single-shot MOLED T2 mapping.

METHODS

A new method called switching modulation patterns multiple overlapping-echo detachment imaging (SWP-MOLED) was designed for multi-slice information sharing via switching the k-space modulation pattern of MOLED imaging. In the SWP-MOLED pulse sequence, three different k-space modulation patterns were devised, making the 12 main echoes of any three adjacent slices symmetrically and uniformly distributed around their k-space centers to obtain diverse contrast weighting information. A multi-slice fusion three-dimensional spatial attention context-guided U-Net was trained with 3000/7000 synthetic data with geometric/brain patterns to efficiently learn the mapping relationship between SWP-MOLED signals and T2 maps. Experiments on numerical human brains, a phantom containing MnCl2 solutions with different concentrations, three healthy volunteers, and three patients diagnosed with meningioma or glioblastoma were performed. The effectiveness of the new method was quantitatively assessed using the structure similarity index measure (SSIM) and root mean square error (RMSE). Multiple statistical analyses were utilized to evaluate the accuracy and significance of the method, including linear regression, Bland-Altman analysis, Mann-Whitney test, Wilcoxon signed rank test, and Friedman test with Bonferroni correction, with the p-value significance level of 0.05.

RESULTS

The results from numerical human brain (The average SSIM of the reconstructed T2 maps was 0.9742/0.9782/0.9826 for MOLED/MS-MOLED/SWP-MOLED) and phantom (The slope of linear fitting of the predicted T2 values vs. reference values was 0.9934/9942/0.9972 for MOLED/MS-MOLED/SWP-MOLED) demonstrated that more accurate T2 maps were delivered by the proposed method, closely resembling the reference maps. From the Friedman test performed on the results of the test data set after the multi-comparison correction, we found that the pairwise performance differences among different reconstruction networks were all statistically significant (p < 0.001). In healthy human brain experiments, the comparison of SWP-MOLED reconstruction with reference measurements indicated no significant difference (p = 0.4504). SWP-MOLED was quite repeatable (average coefficient of variation [CV] = 4.17%) and was not corrupted by motion (average CV = 7.49%). Moreover, the proposed method exhibited clearer lesion contours in clinical cases, demonstrating the potential of the proposed method for clinical applications.

CONCLUSIONS

SWP-MOLED can efficiently exploit the structural similarity and parameter-weighted information diversity of adjacent slices to improve the spatial resolution and quantification accuracy of MOLED T2 mapping. It also exhibits excellent motion robustness. This technique would extend the application of MOLED imaging.

Real-time brain tumour diagnoses using a novel lightweight deep learning model

Brain tumours continue to be a primary cause of worldwide death, highlighting the critical need for effective and accurate diagnostic tools. This article presents MK-YOLOv8, an innovative lightweight deep learning framework developed for the real-time detection and categorization of brain tumours from MRI images. Based on the YOLOv8 architecture, the proposed model incorporates Ghost Convolution, the C3Ghost module, and the SPPELAN module to improve feature extraction and substantially decrease computational complexity. An x-small object detection layer has been added, supporting precise detection of small and x-small tumours, which is crucial for early diagnosis. Trained on the Figshare Brain Tumour (FBT) dataset comprising (3,064) MRI images, MK-YOLOv8 achieved a mean Average Precision (mAP) of 99.1% at IoU (0.50) and 88.4% at IoU (0.50-0.95), outperforming YOLOv8 (98% and 78.8%, respectively). Glioma recall improved by 26%, underscoring the enhanced sensitivity to challenging tumour types. With a computational footprint of only 96.9 GFLOPs (representing 37.5% of YOYOLOv8x'sFLOPs) and utilizing 12.6 million parameters, a mere 18.5% of YOYOLOv8's parameters, MK-YOLOv8 delivers high efficiency with reduced resource demands. Also, it trained on the Br35H dataset (801 images) to guarantee the model's robustness and generalization; it achieved a mAP of 98.6% at IoU (0.50). The suggested model operates at 62 frames per second (FPS) and is suited for real-time clinical processes. These developments establish MK-YOLOv8 as an innovative framework, overcoming challenges in tiny tumour identification and providing a generalizable, adaptable, and precise detection approach for brain tumour diagnostics in clinical settings.

ChatGPT-4's Accuracy in Estimating Thyroid Nodule Features and Cancer Risk From Ultrasound Images

OBJECTIVE

To evaluate the performance of GPT-4 and GPT-4o in accurately identifying features and categories from thyroid nodule ultrasound images following the American College of Radiology Thyroid Imaging Reporting and Data System (TI-RADS).

METHODS

This comparative validation study, conducted between October 2023 and May 2024, utilized 202 thyroid ultrasound images sourced from 3 open-access databases. Both complete and cropped versions of each image were independently evaluated by expert radiologists to establish a reference standard for TI-RADS features and categories. GPT-4 and GPT-4o were prompted to analyze each image, and their generated TI-RADS outputs were compared to the reference standard.

RESULTS

GPT-4 demonstrated high specificity but low sensitivity when assessing complete thyroid ultrasound images across most TI-RADS categories, resulting in mixed overall accuracy. For low-risk nodules (benign), GPT-4 achieved 25.0% sensitivity, 99.5% specificity, and 93.6% accuracy. In contrast, in the higher risk moderately suspicious category GPT-4 showed 75% sensitivity, 22.2% specificity, and 42.1% accuracy. While GPT-4 effectively identified features like smooth margins (73% vs 65% the reference standard), it struggled to identify other features like isoechoic echogenicity (5% vs 46%), and echogenic foci (3% vs 27%). The assessment of cropped images using both GPT-4 and GPT-4o followed similar patterns, though with slight deviations indicating a decrease in performance compared to GPT-4's assessment of complete images.

CONCLUSION

While GPT-4 and GPT-4o models show potential for improving the efficiency of thyroid nodule triage, their performance remains suboptimal, particularly in higher-risk categories. Further refinement and validation of these models are necessary before clinical implementation.

Generative adversarial networks in medical image reconstruction: A systematic literature review

PURPOSE

Recent advancements in generative adversarial networks (GANs) have demonstrated substantial potential in medical image processing. Despite this progress, reconstructing images from incomplete data remains a challenge, impacting image quality. This systematic literature review explores the use of GANs in enhancing and reconstructing medical imaging data.

METHOD

A document survey of computing literature was conducted using the ACM Digital Library to identify relevant articles from journals and conference proceedings using keyword combinations, such as "generative adversarial networks or generative adversarial network," "medical image or medical imaging," and "image reconstruction."

RESULTS

Across the reviewed articles, there were 122 datasets used in 175 instances, 89 top metrics employed 335 times, 10 different tasks with a total count of 173, 31 distinct organs featured in 119 instances, and 18 modalities utilized in 121 instances, collectively depicting significant utilization of GANs in medical imaging. The adaptability and efficacy of GANs were showcased across diverse medical tasks, organs, and modalities, utilizing top public as well as private/synthetic datasets for disease diagnosis, including the identification of conditions like cancer in different anatomical regions. The study emphasized GAN's increasing integration and adaptability in diverse radiology modalities, showcasing their transformative impact on diagnostic techniques, including cross-modality tasks. The intricate interplay between network size, batch size, and loss function refinement significantly impacts GAN's performance, although challenges in training persist.

CONCLUSIONS

The study underscores GANs as dynamic tools shaping medical imaging, contributing significantly to image quality, training methodologies, and overall medical advancements, positioning them as substantial components driving medical advancements.

Artificial intelligence in pediatric osteopenia diagnosis: evaluating deep network classification and model interpretability using wrist X-rays

Osteopenia is a bone disorder that causes low bone density and affects millions of people worldwide. Diagnosis of this condition is commonly achieved through clinical assessment of bone mineral density (BMD). State of the art machine learning (ML) techniques, such as convolutional neural networks (CNNs) and transformer models, have gained increasing popularity in medicine. In this work, we employ six deep networks for osteopenia vs. healthy bone classification using X-ray imaging from the pediatric wrist dataset GRAZPEDWRI-DX. We apply two explainable AI techniques to analyze and interpret visual explanations for network decisions. Experimental results show that deep networks are able to effectively learn osteopenic and healthy bone features, achieving high classification accuracy rates. Among the six evaluated networks, DenseNet201 with transfer learning yielded the top classification accuracy at 95.2 %. Furthermore, visual explanations of CNN decisions provide valuable insight into the blackbox inner workings and present interpretable results. Our evaluation of deep network classification results highlights their capability to accurately differentiate between osteopenic and healthy bones in pediatric wrist X-rays. The combination of high classification accuracy and interpretable visual explanations underscores the promise of incorporating machine learning techniques into clinical workflows for the early and accurate diagnosis of osteopenia.

Non-parametric Bayesian deep learning approach for whole-body low-dose PET reconstruction and uncertainty assessment

Positron emission tomography (PET) imaging plays a pivotal role in oncology for the early detection of metastatic tumors and response to therapy assessment due to its high sensitivity compared to anatomical imaging modalities. The balance between image quality and radiation exposure is critical, as reducing the administered dose results in a lower signal-to-noise ratio (SNR) and information loss, which may significantly affect clinical diagnosis. Deep learning (DL) algorithms have recently made significant progress in low-dose (LD) PET reconstruction. Nevertheless, a successful clinical application requires a thorough evaluation of uncertainty to ensure informed clinical judgment. We propose NPB-LDPET, a DL-based non-parametric Bayesian framework for LD PET reconstruction and uncertainty assessment. Our framework utilizes an Adam optimizer with stochastic gradient Langevin dynamics (SGLD) to sample from the underlying posterior distribution. We employed the Ultra-low-dose PET Challenge dataset to assess our framework's performance relative to the Monte Carlo dropout benchmark. We evaluated global reconstruction accuracy utilizing SSIM, PSNR, and NRMSE, local lesion conspicuity using mean absolute error (MAE) and local contrast, and the clinical relevance of uncertainty maps employing correlation between the uncertainty measures and the dose reduction factor (DRF). Our NPB-LDPET reconstruction method exhibits a significantly superior global reconstruction accuracy for various DRFs (paired t-test, p < 0.0001 , N=10, 631). Moreover, we demonstrate a 21% reduction in MAE (573.54 vs. 723.70, paired t-test, p < 0.0001 , N=28) and an 8.3% improvement in local lesion contrast (2.077 vs. 1.916, paired t-test, p < 0.0001 , N=28). Furthermore, our framework exhibits a stronger correlation between the predicted uncertainty 95th percentile score and the DRF (r 2 = 0.9174 vs.r 2 = 0.6144 , N=10, 631). The proposed framework has the potential to improve clinical decision-making for LD PET imaging by providing a more accurate and informative reconstruction while reducing radiation exposure.

Artificial intelligence for early detection of diabetes mellitus complications via retinal imaging

Background

Diabetes mellitus (DM) increases the risk of vascular complications, and retinal vasculature imaging serves as a valuable indicator of both microvascular and macrovascular health. Moreover, artificial intelligence (AI)-enabled systems developed for high-throughput detection of diabetic retinopathy (DR) using digitized retinal images have become clinically adopted. This study reviews AI applications using retinal images for DM-related complications, highlighting advancements beyond DR screening, diagnosis, and prognosis, and addresses implementation challenges, such as ethics, data privacy, equitable access, and explainability.

Methods

We conducted a thorough literature search across several databases, including PubMed, Scopus, and Web of Science, focusing on studies involving diabetes, the retina, and artificial intelligence. We reviewed the original research based on their methodology, AI algorithms, data processing techniques, and validation procedures to ensure a detailed analysis of AI applications in diabetic retinal imaging.

Results

Retinal images can be used to diagnose DM complications including DR, neuropathy, nephropathy, and atherosclerotic cardiovascular disease, as well as to predict the risk of cardiovascular events. Beyond DR screening, AI integration also offers significant potential to address the challenges in the comprehensive care of patients with DM.

Conclusion

With the ability to evaluate the patient's health status in relation to DM complications as well as risk prognostication of future cardiovascular complications, AI-assisted retinal image analysis has the potential to become a central tool for modern personalized medicine in patients with DM.

Histogram matching-enhanced adversarial learning for unsupervised domain adaptation in medical image segmentation

BACKGROUND

Unsupervised domain adaptation (UDA) seeks to mitigate the performance degradation of deep neural networks when applied to new, unlabeled domains by leveraging knowledge from source domains. In medical image segmentation, prevailing UDA techniques often utilize adversarial learning to address domain shifts for cross-modality adaptation. Current research on adversarial learning tends to adopt increasingly complex models and loss functions, making the training process highly intricate and less stable/robust. Furthermore, most methods primarily focused on segmentation accuracy while neglecting the associated confidence levels and uncertainties.

PURPOSE

To develop a simple yet effective UDA method based on histogram matching-enhanced adversarial learning (HMeAL-UDA), and provide comprehensive uncertainty estimations of the model predictions.

METHODS

Aiming to bridge the domain gap while reducing the model complexity, we developed a novel adversarial learning approach to align multi-modality features. The method, termed HMeAL-UDA, integrates a plug-and-play histogram matching strategy to mitigate domain-specific image style biases across modalities. We employed adversarial learning to constrain the model in the prediction space, enabling it to focus on domain-invariant features during segmentation. Moreover, we quantified the model's prediction confidence using Monte Carlo (MC) dropouts to assess two voxel-level uncertainty estimates of the segmentation results, which were subsequently aggregated into a volume-level uncertainty score, providing an overall measure of the model's reliability. The proposed method was evaluated on three public datasets (Combined Healthy Abdominal Organ Segmentation [CHAOS], Beyond the Cranial Vault [BTCV], and Abdominal Multi-Organ Segmentation Challenge [AMOS]) and one in-house clinical dataset (UTSW). We used 30 MRI scans (20 from the CHAOS dataset and 10 from the in-house dataset) and 30 CT scans from the BTCV dataset for UDA-based, cross-modality liver segmentation. Additionally, 240 CT scans and 60 MRI scans from the AMOS dataset were utilized for cross-modality multi-organ segmentation. The training and testing sets for each modality were split with ratios of approximately 4:1-3:1.

RESULTS

Extensive experiments on cross-modality medical image segmentation demonstrated the superiority of HMeAL-UDA over two state-of-the-art approaches. HMeAL-UDA achieved a mean (± s.d.) Dice similarity coefficient (DSC) of 91.34% ± 1.23% and an HD95 of 6.18 ± 2.93 mm for cross-modality (from CT to MRI) adaptation of abdominal multi-organ segmentation, and a DSC of 87.13% ± 3.67% with an HD95 of 2.48 ± 1.56 mm for segmentation adaptation in the opposite direction (MRI to CT). The results are approaching or even outperforming those of supervised methods trained with "ground-truth" labels in the target domain. In addition, we provide a comprehensive assessment of the model's uncertainty, which can help with the understanding of segmentation reliability to guide clinical decisions.

CONCLUSION

HMeAL-UDA provides a powerful segmentation tool to address cross-modality domain shifts, with the potential to generalize to other deep learning applications in medical imaging.

Deep learning-driven modality imputation and subregion segmentation to enhance high-grade glioma grading

PURPOSE

This study aims to develop a deep learning framework that leverages modality imputation and subregion segmentation to improve grading accuracy in high-grade gliomas.

MATERIALS AND METHODS

A retrospective analysis was conducted using data from 1,251 patients in the BraTS2021 dataset as the main cohort and 181 clinical cases collected from a medical center between April 2013 and June 2018 (51 years ± 17; 104 males) as the external test set. We propose a PatchGAN-based modality imputation network with an Aggregated Residual Transformer (ART) module combining Transformer self-attention and CNN feature extraction via residual links, paired with a U-Net variant for segmentation. Generative accuracy used PSNR and SSIM for modality conversions, while segmentation performance was measured with DSC and HD95 across necrotic core (NCR), edema (ED), and enhancing tumor (ET) regions. Senior radiologists conducted a comprehensive Likert-based assessment, with diagnostic accuracy evaluated by AUC. Statistical analysis was performed using the Wilcoxon signed-rank test and the DeLong test.

RESULTS

The best source-target modality pairs for imputation were T1 to T1ce and T1ce to T2 (p < 0.001). In subregion segmentation, the overall DSC was 0.878 and HD95 was 19.491, with the ET region showing the highest segmentation accuracy (DSC: 0.877, HD95: 12.149). Clinical validation revealed an improvement in grading accuracy by the senior radiologist, with the AUC increasing from 0.718 to 0.913 (P < 0.001) when using the combined imputation and segmentation models.

CONCLUSION

The proposed deep learning framework improves high-grade glioma grading by modality imputation and segmentation, aiding the senior radiologist and offering potential to advance clinical decision-making.

XLLC-Net: A lightweight and explainable CNN for accurate lung cancer classification using histopathological images

Lung cancer imaging plays a crucial role in early diagnosis and treatment, where machine learning and deep learning have significantly advanced the accuracy and efficiency of disease classification. This study introduces the Explainable and Lightweight Lung Cancer Net (XLLC-Net), a streamlined convolutional neural network designed for classifying lung cancer from histopathological images. Using the LC25000 dataset, which includes three lung cancer classes and two colon cancer classes, we focused solely on the three lung cancer classes for this study. XLLC-Net effectively discerns complex disease patterns within these classes. The model consists of four convolutional layers and contains merely 3 million parameters, considerably reducing its computational footprint compared to existing deep learning models. This compact architecture facilitates efficient training, completing each epoch in just 60 seconds. Remarkably, XLLC-Net achieves a classification accuracy of 99.62% [Formula: see text] 0.16%, with precision, recall, and F1 score of 99.33% [Formula: see text] 0.30%, 99.67% [Formula: see text] 0.30%, and 99.70% [Formula: see text] 0.30%, respectively. Furthermore, the integration of Explainable AI techniques, such as Saliency Map and GRAD-CAM, enhances the interpretability of the model, offering clear visual insights into its decision-making process. Our results underscore the potential of lightweight DL models in medical imaging, providing high accuracy and rapid training while ensuring model transparency and reliability.

A generalist deep-learning volume segmentation tool for volume electron microscopy of biological samples

We present the Volume Segmentation Tool (VST), a deep learning software tool that implements volumetric image segmentation in volume electron microscopy image stack data from a wide range of biological sample types. VST automates the handling of data preprocessing, data augmentation, and network building, as well as the configuration for model training, while adapting to the specific dataset. We have tried to make VST more accessible by designing it to operate entirely on local hardware and have provided a browser-based interface with additional features for visualizations of the networks and augmented datasets. VST can utilise contour map prediction to support instance segmentation on top of semantic segmentation. Through examples from various resin-embedded sample derived transmission electron microscopy and scanning electron microscopy datasets, we demonstrate that VST achieves state of the art performance compared to existing approaches.

Large Scale MRI Collection and Segmentation of Cirrhotic Liver

Liver cirrhosis represents the end stage of chronic liver disease, characterized by extensive fibrosis and nodular regeneration that significantly increases mortality risk. While magnetic resonance imaging (MRI) offers a non-invasive assessment, accurately segmenting cirrhotic livers presents substantial challenges due to morphological alterations and heterogeneous signal characteristics. Deep learning approaches show promise for automating these tasks, but progress has been limited by the absence of large-scale, annotated datasets. Here, we present CirrMRI600+, the first comprehensive dataset comprising 628 high-resolution abdominal MRI scans (310 T1-weighted and 318 T2-weighted sequences, totaling nearly 40,000 annotated slices) with expert-validated segmentation labels for cirrhotic livers. The dataset includes demographic information, clinical parameters, and histopathological validation where available. Additionally, we provide benchmark results from 11 state-of-the-art deep learning experiments to establish performance standards. CirrMRI600+ enables the development and validation of advanced computational methods for cirrhotic liver analysis, potentially accelerating progress toward automated Cirrhosis visual staging and personalized treatment planning.

AI in Neurology: Everything, Everywhere, all at Once PART 2: Speech, Sentience, Scruples, and Service

Artificial intelligence (AI) applications are finding use in real-world neurological settings. Whereas part 1 of this 3-part review series focused on the birth of AI and its foundational principles, this part 2 review shifts gears to explore more practical aspects of neurological care. The review details how large language models, generative AI, and robotics are supporting diagnostic accuracy, patient interaction, and treatment personalization. Special attention is given to ethical and philosophical facets of AI that nonetheless impact practical aspects of care and patient safety, such as accountability for AI-driven decisions and the "black box" nature of many algorithms. We will discuss whether AI systems can develop sentience, and the implications for human-AI collaboration. By examining human-robot interactions in neurology, this part 2 review highlights the profound impact AI could have on patient care and, as covered in the ensuing part 3, on global health care delivery and data analytics, while maintaining ethical oversight and human control. ANN NEUROL 2025.

KNN algorithm for accurate identification of IFP lesions in the knee joint: a multimodal MRI study

Knee-related disorders represent a major global health concern and are a leading cause of pain and mobility impairment, particularly in older adults. In clinical medicine, the precise identification and classification of knee joint diseases are essential for early diagnosis and effective treatment. This study presents a novel approach for identifying infrapatellar fat pad (IFP) lesions using the K-Nearest Neighbor (KNN) algorithm in combination with multimodal Magnetic Resonance Imaging (MRI) techniques, specifically mDxion-Quant (mDQ) and T2 mapping (T2m). These imaging methods provide quantitative parameters such as fat fraction (FF), T2*, and T2 values. A set of derived features was constructed through feature engineering to better capture variations within the IFP. These features were used to train the KNN model for classifying knee joint conditions. The proposed method achieved classification accuracies of 94.736% and 92.857% on the training and testing datasets, respectively, outperforming the CNN-Class8 benchmark. This technique holds substantial clinical potential for the early detection of knee joint pathologies, monitoring disease progression, and evaluating post-surgical outcomes.

The Role of Hydrogen Sulfide in the Localization and Structural-Functional Organization of p53 Following Traumatic Brain Injury: Development of a YOLO Model for Detection and Quantification of Apoptotic Nuclei

Traumatic brain injury (TBI) triggers a cascade of molecular and cellular disturbances, including apoptosis, inflammation, and destabilization of neuronal connections. The transcription factor p53 plays a pivotal role in regulating cell fate following brain injury by initiating pro-apoptotic signaling cascades. Hydrogen sulfide (H2S) may significantly contribute to the regulation of p53. Using scanning laser confocal microscopy, we found that after TBI, p53 accumulates extensively in the damaged cerebral cortex, showing distinct subcellular localization in neurons and astrocytes. In neurons, p53 predominantly localizes to the cytoplasm, suggesting involvement in mitochondria-dependent apoptosis, whereas in astrocytes, p53 is found in both the nucleus and cytoplasm, indicating possible activation of transcription-dependent apoptotic pathways. Quantitative analysis confirmed a correlation between p53 localization and morphological signs of cell death, as revealed by Sytox Green and Hoechst nuclear staining. Modulating H2S levels exerted a marked influence on p53 expression and distribution. Administration of the H2S donor sodium thiosulfate (Na2S2O3) reduced the overall number of p53-positive cells, decreased nuclear localization, and lowered the level of apoptosis. Conversely, inhibition of H2S synthesis using aminooxyacetic acid (AOAA) led to enhanced p53 expression, increased numbers of cells exhibiting nuclear fragmentation, and a more pronounced apoptotic response. These findings highlight a neuroprotective role for H2S, likely mediated through the suppression of p53-dependent cell death pathways. To improve analytical accuracy, we developed a YOLO-based deep-learning model for the automated detection of fragmented nuclei. Additionally, evolutionary and molecular dynamics analysis revealed a high degree of p53 conservation among vertebrates and indicated that, although H2S does not form stable complexes with p53, it may modulate its conformational dynamics.

EnsembleEdgeFusion: advancing semantic segmentation in microvascular decompression imaging with innovative ensemble techniques

Semantic segmentation involves an imminent part in the investigation of medical images, particularly in the domain of microvascular decompression, where publicly available datasets are scarce, and expert annotation is demanding. In response to this challenge, this study presents a meticulously curated dataset comprising 2003 RGB microvascular decompression images, each intricately paired with annotated masks. Extensive data preprocessing and augmentation strategies were employed to fortify the training dataset, enhancing the robustness of proposed deep learning model. Numerous up-to-date semantic segmentation approaches, including DeepLabv3+, U-Net, DilatedFastFCN with JPU, DANet, and a custom Vanilla architecture, were trained and evaluated using diverse performance metrics. Among these models, DeepLabv3 + emerged as a strong contender, notably excelling in F1 score. Innovatively, ensemble techniques, such as stacking and bagging, were introduced to further elevate segmentation performance. Bagging, notably with the Naïve Bayes approach, exhibited significant improvements, underscoring the potential of ensemble methods in medical image segmentation. The proposed EnsembleEdgeFusion technique exhibited superior loss reduction during training compared to DeepLabv3 + and achieved maximum Mean Intersection over Union (MIoU) scores of 77.73%, surpassing other models. Category-wise analysis affirmed its superiority in accurately delineating various categories within the test dataset.

Recognizing artery segments on carotid ultrasonography using embedding concatenation of deep image and vision-language models

Objective.Evaluating large artery atherosclerosis is critical for predicting and preventing ischemic strokes. Ultrasonographic assessment of the carotid arteries is the preferred first-line examination due to its ease of use, noninvasive, and absence of radiation exposure. This study proposed an automated classification model for the common carotid artery (CCA), carotid bulb, internal carotid artery (ICA), and external carotid artery (ECA) to enhance the quantification of carotid artery examinations.Approach. A total of 2943 B-mode ultrasound images (CCA: 1563; bulb: 611; ICA: 476; ECA: 293) from 288 patients were collected. Three distinct sets of embedding features were extracted from artificial intelligence networks including pre-trained DenseNet201, vision transformer, and echo contrastive language-image pre-training models using deep learning architectures for pattern recognition. These features were then combined in a support vector machine classifier to interpret the anatomical structures in B-mode images.Main results. After ten-fold cross-validation, the model achieved an accuracy of 82.3%, which was significantly better than using individual feature sets, with ap-value of <0.001.Significance.The proposed model could make carotid artery examinations more accurate and consistent with the achieved classification accuracy. The source code is available athttps://github.com/buddykeywordw/Artery-Segments-Recognition.

Parallel VMamba and Attention-Based Pneumonia Severity Prediction from CXRs: A Robust Model with Segmented Lung Replacement Augmentation

Background/Objectives: Rapid and accurate assessment of lung diseases, like pneumonia, is critical for effective clinical decision-making, particularly during pandemics when disease progression can be severe. Early diagnosis plays a crucial role in preventing complications, necessitating the development of fast and efficient AI-based models for automated severity assessment. Methods: In this study, we introduce a novel approach that leverages VMamba, a state-of-the-art vision model based on the VisualStateSpace (VSS) framework and 2D-Selective-Scan (SS2D) spatial scanning, to enhance lung severity prediction. Integrated in a parallel multi-image regions approach, VMamba effectively captures global and local contextual features through structured state-space modeling, improving feature representation and robustness in medical image analysis. Additionally, we integrate a segmented lung replacement augmentation strategy to enhance data diversity and improve model generalization. The proposed method is trained on the RALO and COVID-19 datasets and compared against state-of-the-art models. Results: Experimental results demonstrate that our approach achieves superior performance, outperforming existing techniques in prediction accuracy and robustness. Key evaluation metrics, including Mean Absolute Error (MAE) and Pearson Correlation (PC), confirm the model's effectiveness, while the incorporation of segmented lung replacement augmentation further enhances adaptability to diverse lung conditions. Conclusions: These findings highlight the potential of our method for reliable and immediate clinical applications in lung infection assessment.

Three-Blind Validation Strategy of Deep Learning Models for Image Segmentation

Image segmentation plays a central role in computer vision applications such as medical imaging, industrial inspection, and environmental monitoring. However, evaluating segmentation performance can be particularly challenging when ground truth is not clearly defined, as is often the case in tasks involving subjective interpretation. These challenges are amplified by inter- and intra-observer variability, which complicates the use of human annotations as a reliable reference. To address this, we propose a novel validation framework-referred to as the three-blind validation strategy-that enables rigorous assessment of segmentation models in contexts where subjectivity and label variability are significant. The core idea is to have a third independent expert, blind to the labeler identities, assess a shuffled set of segmentations produced by multiple human annotators and/or automated models. This allows for the unbiased evaluation of model performance and helps uncover patterns of disagreement that may indicate systematic issues with either human or machine annotations. The primary objective of this study is to introduce and demonstrate this validation strategy as a generalizable framework for robust model evaluation in subjective segmentation tasks. We illustrate its practical implementation in a mammography use case involving dense tissue segmentation while emphasizing its potential applicability to a broad range of segmentation scenarios.

Automatic transformer-based grading of multiple retinal inflammatory signs in uveitis on fluorescein angiography

BACKGROUND

Grading fluorescein angiography (FA) for uveitis is complex, often leading to the oversight of retinal inflammation in clinical studies. This study aims to develop an automated method for grading retinal inflammation.

METHODS

Patients from Jules-Gonin Eye Hospital with active or resolved uveitis who underwent FA between 2018 and 2021 were included. FAs were acquired using a standardized protocol, anonymized, and annotated following the Angiography Scoring for Uveitis Working Group criteria, for four inflammatory signs of the posterior pole. Intergrader agreement was assessed by four independent graders. Four deep learning transformer models were developed, and performance was evaluated using the Ordinal Classification Index, accuracy, F1 scores, and Kappa scores. Saliency analysis was employed to visualize model predictions.

FINDINGS

A total of 543 patients (1042 eyes, 40987 images) were included in the study. The models closely matched expert graders in detecting vascular leakage (F1-score = 0·87, 1-OCI = 0·89), capillary leakage (F1-score = 0·86, 1-OCI = 0·89), macular edema (F1-score = 0·82, 1-OCI = 0·86), and optic disc hyperfluorescence (F1-score = 0·72, 1-OCI = 0·85). Saliency analysis confirmed that the models focused on relevant retinal structures. The mean intergrader agreement across all inflammatory signs was F1-score = 0·79 and 1-OCI = 0·83.

INTERPRETATION

We developed a vision transformer-based model for the automatic grading of retinal inflammation in uveitis, utilizing the largest dataset of FAs in uveitis to date. This approach provides significant clinical benefits for the evaluation of uveitis and paves the way for future advancements, including the identification of novel biomarkers through the integration of clinical data and other modalities.

Development and validation of an artificial intelligence-based pipeline for predicting oral epithelial dysplasia malignant transformation

BACKGROUND

Oral epithelial dysplasia (OED) is a potentially malignant histopathological diagnosis given to lesions of the oral cavity that are at risk of progression to malignancy. Manual grading of OED is subject to substantial variability and does not reliably predict prognosis, potentially resulting in sub-optimal treatment decisions.

METHOD

We developed a Transformer-based artificial intelligence (AI) pipeline for the prediction of malignant transformation from whole-slide images (WSIs) of Haematoxylin and Eosin (H&E) stained OED tissue slides, named ODYN (Oral Dysplasia Network). ODYN can simultaneously classify OED and assign a predictive score (ODYN-score) to quantify the risk of malignant transformation. The model was trained on a large cohort using three different scanners (Sheffield, 358 OED WSIs, 105 control WSIs) and externally validated on cases from three independent centres (Birmingham and Belfast, UK, and Piracicaba, Brazil; 108 OED WSIs).

RESULTS

Model testing yielded an F1-score of 0.96 for classification of dysplastic vs non-dysplastic slides, and an AUROC of 0.73 for malignancy prediction, gaining comparable results to clinical grading systems.

CONCLUSIONS

With further large-scale prospective validation, ODYN promises to offer an objective and reliable solution for assessing OED cases, ultimately improving early detection and treatment of oral cancer.

Expert-guided StyleGAN2 image generation elevates AI diagnostic accuracy for maxillary sinus lesions

BACKGROUND

The progress of artificial intelligence (AI) research in dental medicine is hindered by data acquisition challenges and imbalanced distributions. These problems are especially apparent when planning to develop AI-based diagnostic or analytic tools for various lesions, such as maxillary sinus lesions (MSL) including mucosal thickening and polypoid lesions. Traditional unsupervised generative models struggle to simultaneously control the image realism, diversity, and lesion-type specificity. This study establishes an expert-guided framework to overcome these limitations to elevate AI-based diagnostic accuracy.

METHODS

A StyleGAN2 framework was developed for generating clinically relevant MSL images (such as mucosal thickening and polypoid lesion) under expert control. The generated images were then integrated into training datasets to evaluate their effect on ResNet50's diagnostic performance.

RESULTS

Here we show: 1) Both lesion subtypes achieve satisfactory fidelity metrics, with structural similarity indices (SSIM > 0.996) and maximum mean discrepancy values (MMD < 0.032), and clinical validation scores close to those of real images; 2) Integrating baseline datasets with synthetic images significantly enhances diagnostic accuracy for both internal and external test sets, particularly improving area under the precision-recall curve (AUPRC) by approximately 8% and 14% for mucosal thickening and polypoid lesions in the internal test set, respectively.

CONCLUSIONS

The StyleGAN2-based image generation tool effectively addressed data scarcity and imbalance through high-quality MSL image synthesis, consequently boosting diagnostic model performance. This work not only facilitates AI-assisted preoperative assessment for maxillary sinus lift procedures but also establishes a methodological framework for overcoming data limitations in medical image analysis.

Development and validation of ultrasound-based radiomics deep learning model to identify bone erosion in rheumatoid arthritis

OBJECTIVE

To develop and validate a deep learning radiomics fusion model (DLR) based on ultrasound (US) images to identify bone erosion in rheumatoid arthritis (RA) patients.

METHODS

A total of 432 patients with RA at two institutions were collected. Three hundred twelve patients from center 1 were randomly divided into a training set (N = 218) and an internal test set (N = 94) in a 7:3 ratio; meanwhile, 124 patients from center 2 were as an external test set. Radiomics (Rad) and deep learning (DL) features were extracted based on hand-crafted radiomics and deep transfer learning networks. The least absolute shrinkage and selection operator regression was employed to establish DLR fusion feature from the Rad and DL features. Subsequently, 10 machine learning algorithms were used to construct models and the final optimal model was selected. The performance of models was evaluated using receiver operating characteristic (ROC) and decision curve analysis (DCA). The diagnostic efficacy of sonographers was compared with and without the assistance of the optimal model.

RESULTS

LR was chosen as the optimal algorithm for model construction account for superior performance (Rad/DL/DLR: area under the curve [AUC] = 0.906/0.974/0.979) in the training set. In the internal test set, DLR_LR as the final model had the highest AUC (AUC = 0.966), which was also validated in the external test set (AUC = 0.932). With the aid of DLR_LR model, the overall performance of both junior and senior sonographers improved significantly (P < 0.05), and there was no significant difference between the junior sonographer with DLR_LR model assistance and the senior sonographer without assistance (P > 0.05).

CONCLUSION

DLR model based on US images is the best performer and is expected to become an important tool for identifying bone erosion in RA patients. Key Points • DLR model based on US images is the best performer in identifying BE in RA patients. • DLR model may assist the sonographers to improve the accuracy of BE evaluations.

Advancements in Hematologic Malignancy Detection: A Comprehensive Survey of Methodologies and Emerging Trends

The investigation and diagnosis of hematologic malignancy using blood cell image analysis are major and emerging subjects that lie at the intersection of artificial intelligence and medical research. This survey systematically examines the state-of-the-art in blood cancer detection through image-based analysis, aimed at identifying the most effective computational strategies and highlighting emerging trends. This review focuses on three principal objectives, namely, to categorize and compare traditional machine learning (ML), deep learning (DL), and hybrid learning approaches; to evaluate performance metrics such as accuracy, precision, recall, and area under the ROC curve; and to identify methodological gaps and propose directions for future research. Methodologically, we organize the literature by categorizing the malignancy types-leukemia, lymphoma, and multiple myeloma-and particularizing the preprocessing steps, feature extraction techniques, network architectures, and ensemble strategies employed. For ML methods, we discuss classical classifiers including support vector machines and random forests; for DL, we analyze convolutional neural networks (e.g., AlexNet, VGG, and ResNet) and transformer-based models; and for hybrid systems, we examine combinations of CNNs with attention mechanisms or traditional classifiers. Our synthesis reveals that DL models consistently outperform ML baselines, achieving classification accuracies above 95% in benchmark datasets, with hybrid models pushing peak accuracy to 99.7%. However, challenges remain in data scarcity, class imbalance, and generalizability to clinical settings. We conclude by recommending the integration of multimodal data, semisupervised learning, and rigorous external validation to advance toward deployable diagnostic tools. This survey also provides a comprehensive roadmap for researchers and clinicians striving to harness AI for reliable hematologic cancer detection.

Transfer Learning and Multi-Feature Fusion-Based Deep Learning Model for Idiopathic Macular Hole Diagnosis and Grading from Optical Coherence Tomography Images

Background

Idiopathic macular hole is an ophthalmic disease that seriously affects vision, and its early diagnosis and treatment have important clinical significance to reduce the occurrence of blindness. At present, OCT is the gold standard for diagnosing this disease, but its application is limited due to the need for professional ophthalmologist to diagnose it. The introduction of artificial intelligence will break this situation and make its diagnosis efficient, and how to build an effective predictive model is the key to the problem, and more clinical trials are still needed to verify it.

Objective

This study aims to evaluate the role of deep learning systems in Idiopathic Macular Hole diagnosis, grading, and prediction.

Methods

A single-center, retrospective study used binocular OCT images from IMH patients at the First Affiliated Hospital of Nanchang University (November 2019 - January 2023). A deep learning algorithm, including traditional omics, Resnet101, and fusion models incorporating multi-feature fusion and transfer learning, was developed. Model performance was evaluated using accuracy and AUC. Logistic regression analyzed clinical factors, and a nomogram predicted surgical risk. Analysis was conducted with SPSS 22.0 and R 3.6.3. P < 0.05 was statistically significant.

Results

Among 229 OCT images, the traditional omics, Resnet101, and fusion models achieved accuracies of 93%, 94%, and 95%, respectively, in the training set. In the test set, the fusion model and Resnet101 correctly identified 39 images, while the traditional omics model identified 35. The nomogram had a C-index of 0.996, with macular hole diameter most strongly associated with surgical risk.

Conclusion

The deep learning system with transfer learning and multi-feature fusion effectively diagnoses and grades IMH from OCT images.

What are the perceptions of AI in radiology among UK medical students and junior doctors?

BackgroundThe integration of artificial intelligence (AI) in radiology has the potential to improve diagnostic accuracy and efficiency. Medical students and junior doctors will likely use AI more frequently in the future, making their perceptions essential for identifying educational gaps.PurposeTo explore the perceptions of UK medical students and junior doctors regarding AI in radiology.Material and MethodsA cross-sectional survey was distributed across UK medical schools and foundation programs. A total of 250 responses were analyzed using descriptive statistics and non-parametric tests, focusing on career impact, clinical effectiveness, educational development, and ethical concerns.ResultsMost respondents (55.2%) were undeterred by career uncertainties related to AI, with 64% confident that AI would not replace radiologists. Up to 80.6% supported AI's clinical benefits, and 63.2% endorsed its educational integration. However, there were concerns about job displacement and insufficient AI training. Medical students were more worried about job security than junior doctors, while those committed to radiology were less apprehensive and viewed AI as complementary.ConclusionEducational programs and regulatory frameworks are essential to facilitate AI integration in radiology. Addressing concerns about job displacement and improving AI education will be key to preparing future radiologists for technological advancements.

Enhancing Craniomaxillofacial Surgeries with Artificial Intelligence Technologies

Artificial intelligence (AI) can be applied in multiple subspecialties in craniomaxillofacial (CMF) surgeries. This article overviews AI fundamentals focusing on classification, object detection, and segmentation-core tasks used in CMF applications. The article then explores the development and integration of AI in dentoalveolar surgery, implantology, traumatology, oncology, craniofacial surgery, and orthognathic and feminization surgery. It highlights AI-driven advancements in diagnosis, pre-operative planning, intra-operative assistance, post-operative management, and outcome prediction. Finally, the challenges in AI adoption are discussed, including data limitations, algorithm validation, and clinical integration.

Transforming urban waste collection inventory: AI-Based container classification and Re-Identification

This work lays the groundwork for creating an automated system for the inventory of urban waste elements. Our primary contribution is the development of, to the best of our knowledge, the first re-identification system for urban waste elements that uses Artificial Intelligence and Computer Vision, incorporating information from a classification module and geolocation context to enhance post-processing performance. This re-identification system helps to create and update inventories by determining if a new image matches an existing element in the inventory based on visual similarity or, if not, by adding it as a new identity (new class or new identity of an existing class). Such a system could be highly valuable to local authorities and waste management companies, offering improved facility maintenance, geolocation, and additional applications. This work also addresses the dynamic nature of urban environments and waste management elements by exploring Continual Learning strategies to adapt pretrained systems to new settings with different urban elements. Experimental results show that the proposed system operates effectively across various container types and city layouts. These findings were validated through testing in two different Spanish locations, a "City" and a "Campus", differing in size, illumination conditions, seasons, urban design and container appearance. For the final re-identification system, the baseline system achieves 53.18 mAP (mean Average Precision) in the simple scenario, compared to 21.54 mAP in the complex scenario, with additional challenging unseen variability. Incorporating the proposed post-processing techniques significantly improved results, reaching 74.14 mAP and 71.75 mAP in the simple and complex scenario respectively.

Private Data Incrementalization: Data-Centric Model Development for Clinical Liver Segmentation

Machine Learning models, more specifically Artificial Neural Networks, are transforming medical imaging by enabling precise liver segmentation, a crucial task for diagnosing and treating liver diseases. However, these models often face challenges in adapting to diverse clinical data sources as differences in dataset volume, resolution, and origin impact generalization and performance. This study introduces a Private Data Incrementalization, a data-centric approach to enhance the adaptability of Artificial Neural Networks by progressively exposing them to varied clinical data. As the target of this study is not to propose a new image segmentation model, the existing medical imaging segmentation models-including U-Net, ResUNet++, Fully Convolutional Network, and a modified algorithm based on the Conditional Bernoulli Diffusion Model-are used. The study evaluates these four models using a curated private dataset of computed tomography scans from Coimbra University Hospital, supplemented by two public datasets, 3D-IRCADb01 and CHAOS. The Private Data Incrementalization method systematically increases the volume and diversity of training data, simulating real-world conditions where models must handle varied imaging contexts. Pre-processing and post-processing stages, incremental training, and performance evaluations reveal that structured exposure to diverse datasets improves segmentation performance, with ResUNet++ achieving the highest accuracy (0.9972) and Dice Similarity Coefficient (0.9449), and the best Average Symmetric Surface Distance (0.0053 mm), demonstrating the importance of dataset diversity and volume for segmentation models' robustness and generalization. Private Data Incrementalization thus offers a scalable strategy for building resilient segmentation models, ultimately benefiting clinical workflows, patient care, and healthcare resource management by addressing the variability inherent in clinical imaging data.

Automated whole-breast ultrasound tumor diagnosis using attention-inception network

PURPOSE

Automated Whole-Breast Ultrasound (ABUS) has been widely used as an important tool in breast cancer diagnosis due to the ability of this technique to provide complete three-dimensional (3D) images of breasts. To eliminate the risk of misdiagnosis, computer-aided diagnosis (CADx) systems have been proposed to assist radiologists. Convolutional neural networks (CNNs), renowned for the automatic feature extraction capabilities, have developed rapidly in medical image analysis, and this study proposes a CADx system based on 3D CNN for ABUS.

MATERIALS AND METHODS

This study used a private dataset collected at Sun Yat-Sen University Cancer Center (SYSUCC) from 396 breast tumor patients. First, the tumor volume of interest (VOI) was extracted and resized, and then the tumor was enhanced by histogram equalization. Second, a 3D U-Net++ was employed to segment the tumor mask. Finally, the VOI, the enhanced VOI, and the corresponding tumor mask were fed into a 3D Attention-Inception network to classify the tumor as benign or malignant.

RESULTS

The experiment results indicate an accuracy of 89.4%, a sensitivity of 91.2%, a specificity of 87.6%, and an area under the receiver operating characteristic curve (AUC) of 0.9262, which suggests that the proposed CADx system for ABUS images rivals the performance of experienced radiologists in tumor diagnosis tasks.

CONCLUSION

This study proposes a CADx system consisting of a 3D U-Net++ tumor segmentation model and a 3D attention inception neural network tumor classification model for diagnosis in ABUS images. The results indicate that the proposed CADx system is effective and efficient in tumor diagnosis tasks.

A fusocelular skin dataset with whole slide images for deep learning models

Cutaneous spindle cell (CSC) lesions encompass a spectrum from benign to malignant neoplasms, often posing significant diagnostic challenges. Computer-aided diagnosis systems offer a promising solution to make pathologists' decisions objective and faster. These systems usually require large-scale datasets with curated labels for effective training; however, manual annotation is time-consuming and expensive. To overcome this challenge, crowdsourcing has emerged as a popular and valuable strategy to scale up the labeling process by distributing the effort among different non-expert annotators. This work introduces AI4SkIN, the first public dataset Whole Slide Images (WSIs) for CSC neoplasms, annotated using an innovative crowdsourcing protocol. AI4SkIN dataset contains 641 Hematoxylin and Eosin stained WSIs with multiclass labels from both expert and trainee pathologists. The dataset improves CSC neoplasm diagnosis using advanced machine learning and crowdsourcing based on Gaussian Processes, showing that models trained on non-expert labels perform comparably to those using expert labels. In conclusion, we illustrate that AI4SkIN provides a good resource for developing and validating methods for multiclass CSC neoplasm classification.

Prediction and Stage Classification of Pressure Ulcers in Intensive Care Patients by Machine Learning

Background/Objective: Pressure ulcers are a serious clinical problem associated with high morbidity, mortality and healthcare costs, especially in intensive care unit (ICU) patients. Existing risk assessment tools, such as the Braden Score, are often inadequate in ICU patients and have poor discriminatory power between classes. This increases the need for more sensitive, predictive and integrative systems. The aim of this study was to classify pressure ulcer stages (Stages I-IV) with high accuracy using machine learning algorithms using demographic, clinical and laboratory data of ICU patients and to evaluate the model performance at a level that can be integrated into clinical decision support systems. Methods: A total of 200 patients hospitalized in the ICU were included in the study. Using demographic, clinical and laboratory data of the patients, six different machine learning algorithms (SVM, KNN, ANN, Decision Tree, Naive Bayes and Discriminant Analysis) were used for classification. The models were evaluated using confusion matrices, ROC-AUC analyses and metrics such as class-based sensitivity and error rate. Results: SVM, KNN and ANN models showed the highest success in classifying pressure ulcer stages, achieving 99% overall accuracy and excellent performance with AUC = 1.00. Variables such as Braden score, albumin and CRP levels contributed significantly to model performance. ROC curves showed that the models provided strong discrimination between classes. Key predictors of pressure ulcer severity included prolonged ICU stay (p < 0.001), low albumin (Stage I: 3.4 ± 0.5 g/dL vs. Stage IV: 2.4 ± 0.8 g/dL; p < 0.001) and high CRP (Stage I: 28 mg/L vs. Stage IV: 142 mg/L; p < 0.001). Conclusions: This study shows that machine learning algorithms offer high accuracy and generalization potential in pressure ulcer classification. In particular, the effectiveness of algorithms such as SVM, ANN and KNN in detecting early-stage ulcers is promising in terms of integration into clinical decision support systems. In future studies, the clinical validity of the model should be increased with multicenter datasets and visual-data-based hybrid models.

A multi-layered defense against adversarial attacks in brain tumor classification using ensemble adversarial training and feature squeezing

Deep learning, particularly convolutional neural networks (CNNs), has proven valuable for brain tumor classification, aiding diagnostic and therapeutic decisions in medical imaging. Despite their accuracy, these models are vulnerable to adversarial attacks, compromising their reliability in clinical settings. In this research, we utilized a VGG16-based CNN model to classify brain tumors, achieving 96% accuracy on clean magnetic resonance imaging (MRI) data. To assess robustness, we exposed the model to Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks, which reduced accuracy to 32% and 13%, respectively. We then applied a multi-layered defense strategy, including adversarial training with FGSM and PGD examples and feature squeezing techniques such as bit-depth reduction and Gaussian blurring. This approach improved model resilience, achieving 54% accuracy on FGSM and 47% on PGD adversarial examples. Our results highlight the importance of proactive defense strategies for maintaining the reliability of AI in medical imaging under adversarial conditions.