A survey on deep learning in medical image analysis

An MRI-based deep transfer learning radiomics nomogram for predicting meningioma grade

The aim of this study was to establish a nomogram based on clinical, radiomics, and deep transfer learning (DTL) features to predict meningioma grade. Three hundred forty meningiomas from one hospital composed the training set, and 102 meningiomas from another hospital composed the test set. The enhanced T1 WI images were used for analysis. The clinical, radiomics and DTL features were selected to construct the model. Radiomics and DTL scores were calculated. The deep transfer learning radiomics (DTLR) nomogram was developed on the basis of selected clinical features, radiomics scores and DTL scores. Receiver operating characteristic (ROC) curves and decision curve analysis (DCA) curves were drawn. The clinical features of sex, shape, indistinct margin and peritumoral edema were selected and used to construct the clinical model. Thirty-two radiomics features and 28 DTL features were selected for model construction. The clinical model had an AUC of 0.788. (95% CI: 0.6996-0.8756), with an accuracy of 0.745, a sensitivity of 0.941, and a specificity of 0.549 in the test set. The DTLR nomogram had the highest AUC of 0.866 (95% CI: 0.7984-0.9340), with an accuracy of 0.804, a sensitivity of 0.745, and a specificity of 0.863 in the test set. Compared with the other models, the DTLR nomogram had the greatest net benefit according to the DCA. There was a significant difference between the DTLR nomogram and the clinical model, no significant difference between the rest models in DeLong test.The DTLR nomogram has superior predictive value in DCA and could be a valuable method in clinical decision-making. Given the results of DeLong test, only the radiomics model is sufficient and there is no need to add DTL features. As a new attempt, the DTLR nomogram needs to be improved in the future study.

Application of deep learning convolutional neural networks to identify gastric squamous cell carcinoma in mice

Objective

In non-clinical safety evaluation of drugs, pathological result is one of the gold standards for determining toxic effects. However, pathological diagnosis might be challenging and affected by pathologist expertise. In carcinogenicity studies, drug-induced squamous cell carcinoma (SCC) of the mouse stomach represents a diagnostic challenge for toxicopathologists. This study aims to establish a detection model for mouse gastric squamous cell carcinoma (GSCC) using deep learning algorithms, to improve the accuracy and consistency of pathological diagnoses.

Methods

A total of 93 cases of drug-induced mouse GSCC and 56 cases of normal mouse stomach tissue from carcinogenicity studies were collected. After scanning into digital slides, semi-automated data annotation was performed. All images underwent preprocessing, including tissue extraction, artifact removal, and exclusion of normal epithelial regions. The images were then randomly divided into training, validation, and test sets in an 8:1:1 ratio. Five different convolutional neural networks (CNNs)-FCN, LR-ASPP, DeepLabv3+, U-Net, and DenseNet were applied to identify GSCC and non-GSCC regions. Tumor prediction images (algorithm results shown as overlays) derived from the slide images were compared, and the performance of the constructed models was evaluated using Precision, Recall, and F1-score.

Results

The Precision, Recall, and F1-scores of DenseNet, U-Net, and DeepLabv3 + algorithms were all above 90%. Specifically, the DenseNet model achieved an overall Precision of 0.9044, Recall of 0.9291, and F1-score of 0.9157 in the test set. Compared to the other algorithms, DenseNet exhibited the highest F1-score and Recall, demonstrating superior generalization ability.

Conclusion

The DenseNet algorithm model developed in this study shown promising application potential for assisting in the diagnosis of mouse GSCC. As artificial intelligence (AI) technology continues to advance in non-clinical safety evaluation of drugs, CNN-based toxicological pathology detection models will become essential tools to assist pathologists in precise diagnosis and consistency evaluation.

Augmented reality microscopy to bridge trust between AI and pathologists

Diagnostic certainty is the cornerstone of modern medicine and critical for maximal treatment benefit. When evaluating biomarker expression by immunohistochemistry (IHC), however, pathologists are hindered by complex scoring methodologies, unique positivity cut-offs and subjective staining interpretation. Artificial intelligence (AI) can potentially eliminate diagnostic uncertainty, especially when AI "trustworthiness" is proven by expert pathologists in the context of real-world clinical practice. Building on an IHC foundation model, we employed pathologists-in-the-loop finetuning to produce a programmed cell death ligand 1 (PD-L1) CPS AI Model. We devised a multi-head augmented reality microscope (ARM) system overlayed with the PD-L1 CPS AI Model to assess interobserver variability and gauge the pathologists' trust in AI model outputs. Using difficult to interpret regions on gastroesophageal biopsies, we show that AI-assistance improved case agreement between any 2 pathologists by 14% (agreement on 77% vs 91%) and among 11 pathologists by 26% (agreement on 43% vs 69%). At a clinical cutoff of PD-L1 CPS ≥ 5, the number of cases diagnosed as positive by all 11 pathologists increased by 31%. Our findings underscore the benefits of fully engaging pathologists as active participants in the development and deployment of IHC AI models and frame the roadmap for trustworthy AI as a bridge to increased adoption in routine pathology practice.

The role of nanomedicine and artificial intelligence in cancer health care: individual applications and emerging integrations-a narrative review

Cancer remains one of the deadliest diseases globally, significantly impacting patients' quality of life. Addressing the rising incidence of cancer deaths necessitates innovative approaches such as nanomedicine and artificial intelligence (AI). The convergence of nanomedicine and AI represents a transformative frontier in cancer healthcare, promising unprecedented advancements in diagnosis, treatment, and patient management. This narrative review explores the distinct applications of nanomedicine and AI in oncology, alongside their synergistic potential. Nanomedicine leverages nanoparticles for targeted drug delivery, enhancing therapeutic efficacy while minimizing adverse effects. Concurrently, AI algorithms facilitate early cancer detection, personalized treatment planning, and predictive analytics, thereby optimizing clinical outcomes. Emerging integrations of these technologies could transform cancer care by facilitating precise, personalized, and adaptive treatment strategies. This review synthesizes current research, highlights innovative individual applications, and discusses the emerging integrations of nanomedicine and AI in oncology. The goal is to provide a comprehensive understanding of how these cutting-edge technologies can collaboratively improve cancer diagnosis, treatment, and patient prognosis.

An automated hip fracture detection, classification system on pelvic radiographs and comparison with 35 clinicians

Accurate diagnosis of orthopedic injuries, especially pelvic and hip fractures, is vital in trauma management. While pelvic radiographs (PXRs) are widely used, misdiagnosis is common. This study proposes an automated system that uses convolutional neural networks (CNNs) to detect potential fracture areas and predict fracture conditions, aiming to outperform traditional object detection-based systems. We developed two deep learning models for hip fracture detection and prediction, trained on PXRs from three hospitals. The first model utilized automated hip area detection, cropping, and classification of the resulting patches. The images were preprocessed using the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm. The YOLOv5 architecture was employed for the object detection model, while three different pre-trained deep neural network (DNN) architectures were used for classification, applying transfer learning. Their performance was evaluated on a test dataset, and compared with 35 clinicians. YOLOv5 achieved a 92.66% accuracy on regular images and 88.89% on CLAHE-enhanced images. The classifier models, MobileNetV2, Xception, and InceptionResNetV2, achieved accuracies between 94.66% and 97.67%. In contrast, the clinicians demonstrated a mean accuracy of 84.53% and longer prediction durations. The DNN models showed significantly better accuracy and speed compared to human evaluators (p < 0.0005, p < 0.01). These DNN models highlight promising utility in trauma diagnosis due to their high accuracy and speed. Integrating such systems into clinical practices may enhance the diagnostic efficiency of PXRs.

Machine learning-based approaches for distinguishing viral and bacterial pneumonia in paediatrics: A scoping review

BACKGROUND AND OBJECTIVE

Pneumonia is the leading cause of hospitalisation and mortality among children under five, particularly in low-resource settings. Accurate differentiation between viral and bacterial pneumonia is essential for guiding appropriate treatment, yet it remains challenging due to overlapping clinical and radiographic features. Advances in machine learning (ML), particularly deep learning (DL), have shown promise in classifying pneumonia using chest X-ray (CXR) images. This scoping review summarises the evidence on ML techniques for classifying viral and bacterial pneumonia using CXR images in paediatric patients.

METHODS

This scoping review was conducted following the Joanna Briggs Institute methodology and the PRISMA-ScR guidelines. A comprehensive search was performed in PubMed, Embase, and Scopus to identify studies involving children (0-18 years) with pneumonia diagnosed through CXR, using ML models for binary or multiclass classification. Data extraction included ML models, dataset characteristics, and performance metrics.

RESULTS

A total of 35 studies, published between 2018 and 2025, were included in this review. Of these, 31 studies used the publicly available Kermany dataset, raising concerns about overfitting and limited generalisability to broader, real-world clinical populations. Most studies (n=33) used convolutional neural networks (CNNs) for pneumonia classification. While many models demonstrated promising performance, significant variability was observed due to differences in methodologies, dataset sizes, and validation strategies, complicating direct comparisons. For binary classification (viral vs bacterial pneumonia), a median accuracy of 92.3% (range: 80.8% to 97.9%) was reported. For multiclass classification (healthy, viral pneumonia, and bacterial pneumonia), the median accuracy was 91.8% (range: 76.8% to 99.7%).

CONCLUSIONS

Current evidence is constrained by a predominant reliance on a single dataset and variability in methodologies, which limit the generalisability and clinical applicability of findings. To address these limitations, future research should focus on developing diverse and representative datasets while adhering to standardised reporting guidelines. Such efforts are essential to improve the reliability, reproducibility, and translational potential of machine learning models in clinical settings.

Deep Learning for Automated Ventricle and Periventricular Space Segmentation on CT and T1CE MRI in Neuro-Oncology Patients

PURPOSE

This study aims to create a deep learning (DL) model capable of accurately delineating the ventricles, and by extension, the periventricular space (PVS), following the 2021 EPTN Neuro-Oncology Atlas guidelines on T1-weighted contrast-enhanced MRI scans (T1CE). The performance of this DL model was quantitatively and qualitatively compared with an off-the-shelf model.

MATERIALS AND METHODS

An nnU-Net was trained for ventricle segmentation using both CT and T1CE MRI images from 78 patients. Its performance was compared to that of a publicly available pretrained segmentation model, SynthSeg. The evaluation was conducted on both internal (N = 18) and external (n = 18) test sets, with each consisting of paired CT and T1CE MRI images and expert-delineated ground truths (GTs). Segmentation accuracy was assessed using the volumetric Dice Similarity Coefficient (DSC), 95th percentile Hausdorff distance (HD95), surface DSC, and added path length (APL). Additionally, a local evaluation of ventricle segmentations quantified differences between manual and automatic segmentations across both test sets. All segmentations were scored by radiotherapy technicians for clinical acceptability using a 4-point Likert scale.

RESULTS

The nnU-Net significantly outperformed the SynthSeg model on the internal test dataset in terms of median [range] DSC, 0.93 [0.86-0.95] vs. 0.85 [0.67-0.91], HD95, 0.9 [0.7-2.5] mm vs. 2.2 [1.7-4.8] mm, surface DSC, 0.97 [0.90-0.98] vs. 0.84 [0.70-0.89], and APL, 876 [407-1298] mm vs. 2809 [2311-3622] mm, all with p < 0.001. No significant differences in these metrics were found in the external test set. However clinical ratings favored nnU-Net segmentations on the internal and external test sets. In addition, the nnU-Net had higher clinical ratings than the GT delineation on the internal and external test set.

CONCLUSIONS

The nnU-Net model outperformed the SynthSeg model on the internal dataset in both segmentation metrics and clinician ratings. While segmentation metrics showed no significant differences between the models on the external set, clinician ratings favored nnU-Net, suggesting enhanced clinical acceptability. This suggests that nnU-Net could contribute to more time-efficient and streamlined radiotherapy planning workflows.

Revealing sleep and pain reciprocity with wearables and machine learning

Sleep disturbance and chronic pain share a bidirectional relationship with poor sleep exacerbating pain and pain disrupting sleep. Despite the substantial burden of sleep disturbance and pain, current treatments fail to address their interplay effectively, largely due to the lack of longitudinal data capturing their complex dynamics. Traditional sleep measurement methods that could be used to quantitate daily changes in sleep, such as polysomnography, are costly and unsuitable for large-scale studies in chronic pain populations. New wearable polysomnography devices combined with machine learning algorithms offer a scalable solution, enabling comprehensive, longitudinal analyses of sleep-pain dynamics. In this Perspective, we highlight how these technologies can overcome current limitations in sleep assessment to uncover mechanisms linking sleep and pain. These tools could transform our understanding of the sleep and pain relationship and guide the development of personalized, data-driven treatments.

Diagnosis of Sarcopenia Using Convolutional Neural Network Models Based on Muscle Ultrasound Images: Prospective Multicenter Study

BACKGROUND

Early detection is clinically crucial for the strategic handling of sarcopenia, yet the screening process, which includes assessments of muscle mass, strength, and function, remains complex and difficult to access.

OBJECTIVE

This study aims to develop a convolutional neural network model based on ultrasound images to simplify the diagnostic process and promote its accessibility.

METHODS

This study prospectively evaluated 357 participants (101 with sarcopenia and 256 without sarcopenia) for training, encompassing three types of data: muscle ultrasound images, clinical information, and laboratory information. Three monomodal models based on each data type were developed in the training cohort. The data type with the best diagnostic performance was selected to develop the bimodal and multimodal model by adding another one or two data types. Subsequently, the diagnostic performance of the above models was compared. The contribution ratios of different data types were further analyzed for the multimodal model. A sensitivity analysis was performed by excluding 86 cases with missing values and retaining 271 complete cases for robustness validation. By comprehensive comparison, we finally identified the optimal model (SARCO model) as the convenient solution. Moreover, the SARCO model underwent an external validation with 145 participants (68 with sarcopenia and 77 without sarcopenia) and a proof-of-concept validation with 82 participants (19 with sarcopenia and 63 without sarcopenia) from two other hospitals.

RESULTS

The monomodal model based on ultrasound images achieved the highest area under the receiver operator characteristic curve (AUC) of 0.827 and F1-score of 0.738 among the three monomodal models. Sensitivity analysis on complete data further confirmed the superiority of the ultrasound images model (AUC: 0.851; F1-score: 0.698). The performance of the multimodal model demonstrated statistical differences compared to the best monomodal model (AUC: 0.845 vs 0.827; P=.02) as well as the two bimodal models based on ultrasound images+clinical information (AUC: 0.845 vs 0.826; P=.03) and ultrasound images+laboratory information (AUC: 0.845 vs 0.832, P=0.035). On the other hand, ultrasound images contributed the most evidence for diagnosing sarcopenia (0.787) and nonsarcopenia (0.823) in the multimodal models. Sensitivity analysis showed consistent performance trends, with ultrasound images remaining the dominant contributor (Shapley additive explanation values: 0.810 for sarcopenia and 0.795 for nonsarcopenia). After comprehensive clinical analysis, the monomodal model based on ultrasound images was identified as the SARCO model. Subsequently, the SARCO model achieved satisfactory prediction performance in the external validation and proof-of-concept validation, with AUCs of 0.801 and 0.757 and F1-scores of 0.727 and 0.666, respectively.

CONCLUSIONS

All three types of data contributed to sarcopenia diagnosis, while ultrasound images played a dominant role in model decision-making. The SARCO model based on ultrasound images is potentially the most convenient solution for diagnosing sarcopenia.

TRIAL REGISTRATION

Chinese Clinical Trial Registry ChiCTR2300073651; https://www.chictr.org.cn/showproj.html?proj=199199.

Machine learning-based Diagnostic model for determining the etiology of pleural effusion using Age, ADA and LDH

BACKGROUND

Classification of the etiologies of pleural effusion is a critical challenge in clinical practice. Traditional diagnostic methods rely on a simple cut-off method based on the laboratory tests. However, machine learning (ML) offers a novel approach based on artificial intelligence to improving diagnostic accuracy and capture the non-linear relationships.

METHOD

A retrospective study was conducted using data from patients diagnosed with pleural effusion. The dataset was divided into training and test set with a ratio of 7:3 with 6 machine learning algorithms implemented to diagnosis pleural effusion. Model performances were assessed by accuracy, precision, recall, F1 scores and area under the receiver operating characteristic curve (AUC). Feature importance and average prediction of age, Adenosine (ADA) and Lactate dehydrogenase (LDH) was analyzed. Decision tree was visualized.

RESULTS

A total of 742 patients were included (training cohort: 522, test cohort: 220), 397 (53.3%) diagnosed with malignant pleural effusion (MPE) and 253 (34.1%) with tuberculous pleural effusion (TPE) in the cohort. All of the 6 models performed well in the diagnosis of MPE, TPE and transudates. Extreme Gradient Boosting and Random Forest performed better in the diagnosis of the MPE, with F1 scores above 0.890, while K-Nearest Neighbors and Tabular Transformer performed better in the diagnosis of the TPE, with F1 scores above 0.870. ADA was identified as the most important feature. The ROC of machine learning model outperformed those of conventional diagnostic thresholds.

CONCLUSIONS

This study demonstrates that ML models using age, ADA, and LDH can effectively classify the etiologies of pleural effusion, suggesting that ML-based approaches may enhance diagnostic decision-making.

Artificial intelligence for osteoporosis detection on panoramic radiography: A systematic review and meta analysis

INTRODUCTION

Osteoporosis is a disease characterized by low bone mineral density and an increased risk of fractures. In dentistry, mandibular bone morphology, assessed for example on panoramic images, has been employed to detect osteoporosis. Artificial intelligence (AI) can aid in diagnosing bone diseases from radiographs. We aimed to systematically review, synthesize and appraise the available evidence supporting AI in detecting osteoporosis on panoramic radiographs.

DATA

Studies that used AI to detect osteoporosis on dental panoramic images were included.

SOURCES

On April 8, 2023, a first comprehensive search of electronic databases was conducted, including PubMed, Scopus, Embase, IEEE, arXiv, and Google Scholar (grey literature). This search was subsequently updated on October 6, 2024.

STUDY SELECTION

The Quality Assessment and Diagnostic Accuracy Tool-2 was employed to determine the risk of bias in the studies. Quantitative analyses involved meta-analyses of diagnostic accuracy measures, including sensitivity and specificity, yielding Diagnostic Odds Ratios (DOR) and synthesized positive likelihood ratios (LR+). The certainty of evidence was assessed using the Grading of Recommendations Assessment, Development, and Evaluation system.

RESULTS

A total of 24 studies were included. Accuracy ranged from 50% to 99%, sensitivity from 50% to 100%, and specificity from 38% to 100%. A minority of studies (n=10) had a low risk of bias in all domains, while the majority (n=18) showed low risk of applicability concerns. Pooled sensitivity was 87.92% and specificity 81.93%. DOR was 32.99, and L+ 4.87. Meta-regression analysis indicated that sample size had only a marginal impact on heterogeneity (R² = 0.078, p = 0.052), suggesting other study-level factors may contribute to variability. Egger's test suggested potential small-study effects (p < 0.001), indicating a risk of publication bias.

CONCLUSION

AI, particularly deep learning, showed high diagnostic accuracy in detecting osteoporosis on panoramic radiographs. The results indicate a strong potential for AI to enhance osteoporosis screening in dental settings. However, significant heterogeneity across studies and potential small-study effects highlight the need for further validation, standardization, and larger, well-powered studies to improve model generalizability.

CLINICAL SIGNIFICANCE

The application of AI in analyzing panoramic radiographs could transform osteoporosis screening in routine dental practice by providing early and accurate diagnosis. This has the potential to integrate osteoporosis detection seamlessly into dental workflows, improving patient outcomes and enabling timely referrals for medical intervention. Addressing issues of model validation and comparability is critical to translating these findings into widespread clinical use.

Forensic sex classification by convolutional neural network approach by VGG16 model: accuracy, precision and sensitivity

INTRODUCTION

In the reconstructive phase of medico-legal human identification, the sex estimation is crucial in the reconstruction of the biological profile and can be applied both in identifying victims of mass disasters and in the autopsy room. Due to the inherent subjectivity associated with traditional methods, artificial intelligence, specifically, convolutional neural networks (CNN) may present a competitive option.

OBJECTIVES

This study evaluates the reliability of VGG16 model as an accurate forensic sex prediction algorithm and its performance using orthopantomography (OPGs).

MATERIALS AND METHODS

This study included 1050 OPGs from patients at the Santa Maria Local Health Unit Stomatology Department. Using Python, the OPGs were pre-processed, resized and similar copies were created using data augmentation methods. The model was evaluated for precision, sensitivity, F1-score and accuracy, and heatmaps were created.

RESULTS AND DISCUSSION

The training revealed a discrepancy between the validation and training loss values. In the general test, the model showed a general balance between sexes, with F1-scores of 0.89. In the test by age group, contrary to expectations, the model was most accurate in the 16-20 age group (90%). Apart from the mandibular symphysis, analysis of the heatmaps showed that the model did not focus on anatomically relevant areas, possibly due to the lack of application of image extraction techniques.

CONCLUSIONS

The results indicate that CNNs are accurate in classifying human remains based on the generic factor sex for medico-legal identification, achieving an overall accuracy of 89%. However, further research is necessary to enhance the models' performance.

A domain adaptation model for carotid ultrasound: Image harmonization, noise reduction, and impact on cardiovascular risk markers

Deep learning has been used extensively for medical image analysis applications, assuming the training and test data adhere to the same probability distributions. However, a common challenge arises when dealing with medical images generated by different systems or even the same system with varying parameter settings. Such images often contain diverse textures and noise patterns, violating the assumption. Consequently, models trained on data from one machine or setting usually struggle to perform effectively on data from another. To address this issue in ultrasound images, we proposed a Generative Adversarial Network (GAN) based model in this paper. We formulated image harmonization and denoising tasks as an image-to-image translation task, wherein we adapt the texture pattern and reduced noise in Carotid ultrasound images while keeping the image content (the anatomy) unchanged. The performance was evaluated using feature distribution and pixel-space similarity metrics. In addition, blood-to-tissue contrast and influence on computed risk markers (Grey scale median, GSM) were evaluated. The results showed that domain adaptation was achieved in both tasks (histogram correlation 0.920 (0.043) and 0.844 (0.062)), as compared to no adaptation (0.890 (0.077) and 0.707 (0.098)), and that the anatomy of the images was retained (structure similarity index measure e.g. the arterial wall 0.71 (0.09) and 0.80 (0.08)). In addition, the image noise level (contrast) did not change in the image harmonization task (-34.1 (3.8) vs -35.2 (4.1) dB) but was improved in the noise reduction task (-23.5 (3.2) vs -46.7 (18.1) dB). To validate the performance of the proposed model, we compare its results with CycleGAN, the current state-of-the-art model. Our model outperformed CycleGAN in both tasks. Finally, the risk marker GSM was significantly changed in the noise reduction but not in the image harmonization task. We conclude that domain translation models are powerful tools for improving ultrasound image while retaining the underlying anatomy, but downstream calculations of risk markers may be affected.

CrossConvPyramid: Deep Multimodal Fusion for Epileptic Magnetoencephalography Spike Detection

Magnetoencephalography (MEG) is a vital non-invasive tool for epilepsy analysis, as it captures high-resolution signals that reflect changes in brain activity over time. The automated detection of epileptic spikes within these signals can significantly reduce the labor and time required for manual annotation of MEG recording data, thereby aiding clinicians in identifying epileptogenic foci and evaluating treatment prognosis. Research in this domain often utilizes the raw, multi-channel signals from MEG scans for spike detection, commonly neglecting the multi-channel spiking patterns from spatially adjacent channels. Moreover, epileptic spikes share considerable morphological similarities with artifact signals within the recordings, posing a challenge for models to differentiate between the two. In this paper, we introduce a multimodal fusion framework that addresses these two challenges collectively. Instead of relying solely on the signal recordings, our framework also mines knowledge from their corresponding topography-map images, which encapsulate the spatial context and amplitude distribution of the input signals. To facilitate more effective data fusion, we present a novel multimodal feature fusion technique called CrossConvPyramid, built upon a convolutional pyramid architecture augmented by an attention mechanism. It initially employs cross-attention and a convolutional pyramid to encode inter-modal correlations within the intermediate features extracted by individual unimodal networks. Subsequently, it utilizes a self-attention mechanism to refine and select the most salient features from both inter-modal and unimodal features, specifically tailored for the spike classification task. Our method achieved the average F1 scores of 92.88% and 95.23% across two distinct real-world MEG datasets from separate centers, respectively outperforming the current state-of-the-art by 2.31% and 0.88%. We plan to release the code on GitHub later.

AI-driven healthcare: Fairness in AI healthcare: A survey

Artificial intelligence (AI) is rapidly advancing in healthcare, enhancing the efficiency and effectiveness of services across various specialties, including cardiology, ophthalmology, dermatology, emergency medicine, etc. AI applications have significantly improved diagnostic accuracy, treatment personalization, and patient outcome predictions by leveraging technologies such as machine learning, neural networks, and natural language processing. However, these advancements also introduce substantial ethical and fairness challenges, particularly related to biases in data and algorithms. These biases can lead to disparities in healthcare delivery, affecting diagnostic accuracy and treatment outcomes across different demographic groups. This review paper examines the integration of AI in healthcare, highlighting critical challenges related to bias and exploring strategies for mitigation. We emphasize the necessity of diverse datasets, fairness-aware algorithms, and regulatory frameworks to ensure equitable healthcare delivery. The paper concludes with recommendations for future research, advocating for interdisciplinary approaches, transparency in AI decision-making, and the development of innovative and inclusive AI applications.

Significant Radiation Reduction Using Cloud-Based AI Imaging in Manually Matched Cohort of Complex Aneurysm Repair

BACKGROUND

Cloud-based, surgical augmented intelligence (Cydar Medical, Cambridge, United Kingdom) can be used for surgical planning and intraoperative imaging guidance during complex endovascular aortic procedures. We aim to evaluate radiation exposure, operative safety metrics, and postoperative renal outcomes following implementation of Cydar imaging guidance using a manually matched cohort of aortic procedures.

METHODS

We retrospectively reviewed our prospectively maintained database of endovascular aortic cases. Patients repaired using Cydar imaging were matched to patients who underwent a similar procedure without using Cydar. Matching was performed manually on a 1:1 basis using anatomy, device configuration, number of branches/fenestrations, and adjunctive procedures including in-situ laser fenestration. Radiation, contrast use, and other operative metrics were compared. Preoperative and postoperative maximum creatinine was compared to assess for acute kidney injury (AKI) based on risk, injury, failure, loss of kidney function, and end-stage kidney disease (RIFLE) criteria.

RESULTS

Hundred patients from 2012 to 2023 were identified: 50 cases (38 fenestrated endovascular aortic repairs, 2 thoracic endovascular aortic repairs, 3 octopus-type thoracoabdominal aortic aneurysm repair, 7 endovascular aneurysm repairs) where Cydar imaging was used, with suitable matches to 50 non-Cydar cases. Baseline characteristics including body mass index did not differ significantly between the 2 groups (27.8 ± 5.6 vs. 26.7 ± 6.1; P = 0.31). Radiation dose was significantly lower in the Cydar group (2529 ± 2256 vs. 3676 ± 2976 mGy; P < 0.03), despite there being no difference in fluoroscopy time (51 ± 29.4 vs. 58 ± 37.2 min; P = 0.37). Contrast volume (94 ± 37.4 vs. 93 ± 43.9 mL; P = 0.73), estimated blood loss (169 ± 223 vs. 193 ± 222 mL; P = 0.97), and procedure time (154 ± 78 vs. 165 ± 89.1 min) did not differ significantly. Additionally, Cydar versus non-Cydar patients did not show a significant difference between precreatinine and postcreatinine changes (0.13 ± 0.08 vs. 0.05 ± 0.07; P = 0.34). Only one patient in the non-Cydar group met RIFLE criteria for AKI postoperatively.

CONCLUSION

The use of cloud-based augmented intelligence imaging was associated with a significant reduction in radiation dose in a cohort of matched aortic procedures but did not appear to affect other parameters or renal function. Even with advanced imaging, surgeons should remain conscientious about radiation safety and administration of nephrotoxic contrast agents.

Integrating language into medical visual recognition and reasoning: A survey

Vision-Language Models (VLMs) are regarded as efficient paradigms that build a bridge between visual perception and textual interpretation. For medical visual tasks, they can benefit from expert observation and physician knowledge extracted from textual context, thereby improving the visual understanding of models. Motivated by the fact that extensive medical reports are commonly attached to medical imaging, medical VLMs have triggered more and more interest, serving not only as self-supervised learning in the pretraining stage but also as a means to introduce auxiliary information into medical visual perception. To strengthen the understanding of such a promising direction, this survey aims to provide an in-depth exploration and review of medical VLMs for various visual recognition and reasoning tasks. Firstly, we present an introduction to medical VLMs. Then, we provide preliminaries and delve into how to exploit language in medical visual tasks from diverse perspectives. Further, we investigate publicly available VLM datasets and discuss the challenges and future perspectives. We expect that the comprehensive discussion about state-of-the-art medical VLMs will make researchers realize their significant potential.

Feasibility of deep learning algorithm in diagnosing lumbar central canal stenosis using abdominal CT

OBJECTIVE

To develop a deep learning algorithm for diagnosing lumbar central canal stenosis (LCCS) using abdominal CT (ACT) and lumbar spine CT (LCT).

MATERIALS AND METHODS

This retrospective study involved 109 patients undergoing LCTs and ACTs between January 2014 and July 2021. The dural sac on CT images was manually segmented and classified as normal or stenosed (dural sac cross-sectional area ≥ 100 mm2 or < 100 mm2, respectively). A deep learning model based on U-Net architecture was developed to automatically segment the dural sac and classify the central canal stenosis. The classification performance of the model was compared on a testing set (990 images from 9 patients). The accuracy, sensitivity, and specificity of automatic segmentation were quantitatively evaluated by comparing its Dice similarity coefficient (DSC) and intraclass correlation coefficient (ICC) with those of manual segmentation.

RESULTS

In total, 990 CT images from nine patients (mean age ± standard deviation, 77 ± 7 years; six men) were evaluated. The algorithm achieved high segmentation performance with a DSC of 0.85 ± 0.10 and ICC of 0.82 (95% confidence interval [CI]: 0.80,0.85). The ICC between ACTs and LCTs on the deep learning algorithm was 0.89 (95%CI: 0.87,0.91). The accuracy of the algorithm in diagnosing LCCS with dichotomous classification was 84%(95%CI: 0.82,0.86). In dataset analysis, the accuracy of ACTs and LCTs was 85%(95%CI: 0.82,0.88) and 83%(95%CI: 0.79,0.86), respectively. The model showed better accuracy for ACT than LCT.

CONCLUSION

The deep learning algorithm automatically diagnosed LCCS on LCTs and ACTs. ACT had a diagnostic performance for LCCS comparable to that of LCT.

Application of deep learning models for accurate classification of fluid collections in acute necrotizing pancreatitis on computed tomography: a multicenter study

PURPOSE

To apply CT-based deep learning (DL) models for accurate solid debris-based classification of pancreatic fluid collections (PFC) in acute pancreatitis (AP).

MATERIAL AND METHODS

This retrospective study comprised four tertiary care hospitals. Consecutive patients with AP and PFCs who had computed tomography (CT) prior to drainage were screened. Those who had magnetic resonance imaging (MRI) or endoscopic ultrasound (EUS) within 20 days of CT were considered for inclusion. Axial CT images were utilized for model training. Images were labelled as those with≤30% solid debris and >30% solid debris based on MRI or EUS. Single center data was used for model training and validation. Data from other three centers comprised the held out external test cohort. We experimented with ResNet 50, Vision transformer (ViT), and MedViT architectures.

RESULTS

Overall, we recruited 152 patients (129 training/validation and 23 testing). There were 1334, 334 and 512 images in the training, validation, and test cohorts, respectively. In the overall training and validation cohorts, ViT and MedVit models had high diagnostic performance (sensitivity 92.4-98.7%, specificity 89.7-98.4%, and AUC 0.908-0.980). The sensitivity (85.3-98.6%), specificity (69.4-99.4%), and AUC (0.779-0.984) of all the models was high in all the subgroups in the training and validation cohorts. In the overall external test cohort, MedViT had the best diagnostic performance (sensitivity 75.2%, specificity 75.3%, and AUC 0.753). MedVit had sensitivity, specificity, and AUC of 75.2%, 74.3%, and 0.748, in walled off necrosis and 79%, 74.2%, 75.3%, and 0.767 for collections >5 cm.

CONCLUSION

DL-models have moderate diagnostic performance for solid-debris based classification of WON and collections greater than 5 cm on CT.

Deep-learning-based diagnosis framework for ankle-brachial index defined peripheral arterial disease of lower extremity wound: Comparison with physicians

BACKGROUND AND OBJECTIVE

Few studies have evaluated peripheral artery disease (PAD) in patients with lower extremity wounds by a convolutional neural network (CNN)-based deep learning algorithm. We aimed to establish a framework for PAD detection, peripheral arterial occlusive disease (PAOD) detection, and PAD classification in patients with lower extremity wounds by the AlexNet, GoogleNet, and ResNet101V2 algorithms.

METHODS

Our proposed framework was based on a CNN-based AlexNet, GoogleNet, or ResNet 101V2 model devoted to performing optimized detection and classification of PAD in patients with lower extremity wounds. We also evaluated the performance of the plastic and reconstructive surgeons (PRS) and general practitioner (GP).

RESULTS

Compared to the performance of AlexNet or GoogleNet, a slight increase in ResNet101V2-based performance of PAD detection, PAOD detection, and PAD classification with original images was observed. A similar observation was found for PAD detection, PAOD detection, and PAD classification with background-removal or cropped images. GP group had a lower performance for PAD and PAOD detection than did the three models with original images, while a similar performance for PAD detection was observed in PRS group and the 3 models.

CONCLUSIONS

We proposed a promising framework using CNN-based deep learning based on objective ankle-brachial index (ABI) values and image preprocessing to characterize PAD detection, PAOD detection, and PAD classification for lower extremity wounds, which provides an easily implemented and objective and reliable computational tool for physicians to automatically identify and classify PAD.

An inherently interpretable AI model improves screening speed and accuracy for early diabetic retinopathy

Diabetic retinopathy (DR) is a frequent complication of diabetes, affecting millions worldwide. Screening for this disease based on fundus images has been one of the first successful use cases for modern artificial intelligence in medicine. However, current state-of-the-art systems typically use black-box models to make referral decisions, requiring post-hoc methods for AI-human interaction and clinical decision support. We developed and evaluated an inherently interpretable deep learning model, which explicitly models the local evidence of DR as part of its network architecture, for clinical decision support in early DR screening. We trained the network on 34,350 high-quality fundus images from a publicly available dataset and validated its performance on a large range of ten external datasets. The inherently interpretable model was compared to post-hoc explainability techniques applied to a standard DNN architecture. For comparison, we obtained detailed lesion annotations from ophthalmologists on 65 images to study if the class evidence maps highlight clinically relevant information. We tested the clinical usefulness of our model in a retrospective reader study, where we compared screening for DR without AI support to screening with AI support with and without AI explanations. The inherently interpretable deep learning model obtained an accuracy of .906 [.900-.913] (95%-confidence interval) and an AUC of .904 [.894-.913] on the internal test set and similar performance on external datasets, comparable to the standard DNN. High evidence regions directly extracted from the model contained clinically relevant lesions such as microaneurysms or hemorrhages with a high precision of .960 [.941-.976], surpassing post-hoc techniques applied to a standard DNN. Decision support by the model highlighting high-evidence regions in the image improved screening accuracy for difficult decisions and improved screening speed. This shows that inherently interpretable deep learning models can provide clinical decision support while obtaining state-of-the-art performance improving human-AI collaboration.

Leveraging deep learning for improving parameter extraction from perfusion MR images: A narrative review

BACKGROUND

Perfusion magnetic resonance imaging (MRI) is a non-invasive technique essential for assessing tissue microcirculation and perfusion dynamics. Various perfusion MRI techniques like Dynamic Contrast-Enhanced (DCE), Dynamic Susceptibility Contrast (DSC), Arterial Spin Labeling (ASL), and Intravoxel Incoherent Motion (IVIM) provide critical insights into physiological and pathological processes. However, traditional methods for quantifying perfusion parameters are time-consuming, often prone to variability, and limited by noise and complex tissue dynamics. Recent advancements in artificial intelligence (AI), particularly in deep learning (DL), offer potential solutions to these challenges. DL algorithms can process large datasets efficiently, providing faster, more accurate parameter extraction with reduced subjectivity.

AIM

This paper reviews the state-of-the-art DL-based techniques applied to perfusion MRI, considering DCE, DSC, ASL and IVIM acquisitions, focusing on their advantages, challenges, and potential clinical applications.

MAIN FINDINGS

DL-driven methods promise significant improvements over conventional approaches, addressing limitations like noise, manual intervention, and inter-observer variability. Deep learning techniques such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and generative adversarial networks (GANs) are particularly valuable in handling spatial and temporal data, enhancing image quality, and facilitating precise parameter extraction.

CONCLUSIONS

These innovations could revolutionize diagnostic accuracy and treatment planning, offering a new frontier in perfusion MRI by integrating DL with traditional imaging methods. As the demand for precise, efficient imaging grows, DL's role in perfusion MRI could significantly improve clinical outcomes, making personalized treatment a more realistic goal.

Machine Learning of Endoscopy Images to Identify, Classify, and Segment Sinonasal Masses

BACKGROUND

We developed and assessed the performance of a machine learning model (MLM) to identify, classify, and segment sinonasal masses based on endoscopic appearance.

METHODS

A convolutional neural network-based model was constructed from nasal endoscopy images from patients evaluated at an otolaryngology center between 2013 and 2024. Images were classified into four groups: normal endoscopy, nasal polyps, benign, and malignant tumors. Polyps and tumors were confirmed with histopathological diagnosis. Images were annotated by an otolaryngologist and independently verified by two other otolaryngologists. We used high- and low-quality images to mirror real-world conditions. The models used for classification (EfficientNet-B2) and segmentation (nnUNet) were trained, validated, and tested at an 8:1:1 ratio. The performance accuracy was averaged across a 10-fold cross-validation assessment. Segmentation accuracy was assessed via Dice similarity coefficients.

RESULTS

A total of 1242 images from 311 patients were used. The MLM was trained, validated, and tested on 663 normal, 276 polyps, 157 benign, and 146 malignant tumors images. Overall, the model performed at 84.1 ± 4.3% accuracy in the validation set and 80.4 ± 1.7% in the test set. The model correctly identified the presence of a sinonasal mass at 90.5 ± 1.2% accuracy rate. The MLM accuracy performance rates were 86.2 ± 1.0% for polyps and 84.1 ± 1.8% for tumors. Benign and malignant tumor subclassification achieved 87.8 ± 2.1% and 94.0 ± 2.4% accuracy, respectively. Segmentation accuracies for polyps were 72.3% and 72.8% for tumors.

CONCLUSIONS

An MLM for nasal endoscopy images can perform with moderate to high accuracy in identifying, classifying, and segmenting sinonasal masses. Performance in future iterations may improve with larger and more diverse training datasets.

MERIT: Multi-view evidential learning for reliable and interpretable liver fibrosis staging

Accurate staging of liver fibrosis from magnetic resonance imaging (MRI) is crucial in clinical practice. While conventional methods often focus on a specific sub-region, multi-view learning captures more information by analyzing multiple patches simultaneously. However, previous multi-view approaches could not typically calculate uncertainty by nature, and they generally integrate features from different views in a black-box fashion, hence compromising reliability as well as interpretability of the resulting models. In this work, we propose a new multi-view method based on evidential learning, referred to as MERIT, which tackles the two challenges in a unified framework. MERIT enables uncertainty quantification of the predictions to enhance reliability, and employs a logic-based combination rule to improve interpretability. Specifically, MERIT models the prediction from each sub-view as an opinion with quantified uncertainty under the guidance of the subjective logic theory. Furthermore, a distribution-aware base rate is introduced to enhance performance, particularly in scenarios involving class distribution shifts. Finally, MERIT adopts a feature-specific combination rule to explicitly fuse multi-view predictions, thereby enhancing interpretability. Results have showcased the effectiveness of the proposed MERIT, highlighting the reliability and offering both ad-hoc and post-hoc interpretability. They also illustrate that MERIT can elucidate the significance of each view in the decision-making process for liver fibrosis staging. Our code will be released via https://github.com/HenryLau7/MERIT.

Towards High-Quality MRI Reconstruction With Anisotropic Diffusion-Assisted Generative Adversarial Networks and Its Multi-Modal Images Extension

Recently, fast Magnetic Resonance Imaging reconstruction technology has emerged as a promising way to improve the clinical diagnostic experience by significantly reducing scan times. While existing studies have used Generative Adversarial Networks to achieve impressive results in reconstructing MR images, they still suffer from challenges such as blurred zones/boundaries and abnormal spots caused by inevitable noise in the reconstruction process. To this end, we propose a novel deep framework termed Anisotropic Diffusion-Assisted Generative Adversarial Networks, which aims to maximally preserve valid high-frequency information and structural details while minimizing noises in reconstructed images by optimizing a joint loss function in a unified framework. In doing so, it enables more authentic and accurate MR image generation. To specifically handle unforeseeable noises, an Anisotropic Diffused Reconstruction Module is developed and added aside the backbone network as a denoise assistant, which improves the final image quality by minimizing reconstruction losses between targets and iteratively denoised generative outputs with no extra computational complexity during the testing phase. To make the most of valuable MRI data, we extend its application to support multi-modal learning to boost reconstructed image quality by aggregating more valid information from images of diverse modalities. Extensive experiments on public datasets show that the proposed framework can achieve superior performance in polishing up the quality of reconstructed MR images. For example, the proposed method obtains average PSNR and mSSIM values of 35.785 dB and 0.9765 on the MRNet dataset, which are at least about 2.9 dB and 0.07 higher than those from the baselines.

PARADISE: Personalized and regional adaptation for HIE disease identification and segmentation

Hypoxic ischemic encephalopathy (HIE) is a brain dysfunction occurring in approximately 1-5/1000 term-born neonates. Accurate segmentation of HIE lesions in brain MRI is crucial for prognosis and diagnosis but presents a unique challenge due to the diffuse and small nature of these abnormalities, which resulted in a substantial gap between the performance of machine learning-based segmentation methods and clinical expert annotations for HIE. To address this challenge, we introduce ParadiseNet, an algorithm specifically designed for HIE lesion segmentation. ParadiseNet incorporates global-local learning, progressive uncertainty learning, and self-evolution learning modules, all inspired by clinical interpretation of neonatal brain MRIs. These modules target issues such as unbalanced data distribution, boundary uncertainty, and imprecise lesion detection, respectively. Extensive experiments demonstrate that ParadiseNet significantly enhances small lesion detection (<1%) accuracy in HIE, achieving an over 4% improvement in Dice, 6% improvement in NSD compared to U-Net and other general medical image segmentation algorithms.

Automatic cerebral microbleeds detection from MR images via multi-channel and multi-scale CNNs

BACKGROUND

Computer-aided detection (CAD) systems have been widely used to assist medical professionals in interpreting medical images, aiding in the detection of potential diseases. Despite their usefulness, CAD systems cannot yet fully replace doctors in diagnosing many conditions due to limitations in current algorithms. Cerebral microbleeds (CMBs) are a critical area of concern for neurological health, and accurate detection of CMBs is essential for understanding their impact on brain function. This study aims to improve CMB detection by enhancing existing machine learning algorithms.

METHODS

This paper presents four CNN-based algorithms designed to enhance CMB detection. The detection methods are categorized into traditional machine learning approaches and deep learning-based methods. The traditional methods, while computationally efficient, offer lower sensitivity, while CNN-based approaches promise greater accuracy. The algorithms proposed in this study include a multi-channel CNN with optimized architecture and a multiscale CNN structure, both of which were designed to reduce false positives and improve overall performance.

RESULTS

The first CNN algorithm, with an optimized multi-channel architecture, demonstrated a sensitivity of 99.6 %, specificity of 99.3 %, and accuracy of 99.5 %. The fourth algorithm, based on a stable multiscale CNN structure, achieved sensitivity of 98.2 %, specificity of 97.4 %, and accuracy of 97.8 %. Both algorithms exhibited a significant reduction in false positives compared to traditional methods. The experiments conducted confirm the effectiveness of these algorithms in improving the precision and reliability of CMB detection.

CONCLUSION

The proposed CNN-based algorithms demonstrate a significant advancement in the automated detection of CMBs, with notable improvements in sensitivity, specificity, and accuracy. These results underscore the potential of deep learning models, particularly CNNs, in enhancing CAD systems for neurological disease detection and reducing diagnostic errors. Further research and optimization may allow these algorithms to be integrated into clinical practices, providing more reliable support for healthcare professionals.

AutoFRS: an externally validated, annotation-free approach to computational preoperative complication risk stratification in pancreatic surgery - an experimental study

BACKGROUND

The risk of postoperative pancreatic fistula (POPF), one of the most dreaded complications after pancreatic surgery, can be predicted from preoperative imaging and tabular clinical routine data. However, existing studies suffer from limited clinical applicability due to a need for manual data annotation and a lack of external validation. We propose AutoFRS (automated fistula risk score software), an externally validated end-to-end prediction tool for POPF risk stratification based on multimodal preoperative data.

MATERIALS AND METHODS

We trained AutoFRS on preoperative contrast-enhanced computed tomography imaging and clinical data from 108 patients undergoing pancreatic head resection and validated it on an external cohort of 61 patients. Prediction performance was assessed using the area under the receiver operating characteristic curve (AUC) and balanced accuracy. In addition, model performance was compared to the updated alternative fistula risk score (ua-FRS), the current clinical gold standard method for intraoperative POPF risk stratification.

RESULTS

AutoFRS achieved an AUC of 0.81 and a balanced accuracy of 0.72 in internal validation and an AUC of 0.79 and a balanced accuracy of 0.70 in external validation. In a patient subset with documented intraoperative POPF risk factors, AutoFRS (AUC: 0.84 ± 0.05) performed on par with the uaFRS (AUC: 0.85 ± 0.06). The AutoFRS web application facilitates annotation-free prediction of POPF from preoperative imaging and clinical data based on the AutoFRS prediction model.

CONCLUSION

POPF can be predicted from multimodal clinical routine data without human data annotation, automating the risk prediction process. We provide additional evidence of the clinical feasibility of preoperative POPF risk stratification and introduce a software pipeline for future prospective evaluation.

GRAPHICAL ABSTRACT

Harmonizing foundation models in healthcare: A comprehensive survey of their roles, relationships, and impact in artificial intelligence's advancing terrain

The lightning development of artificial intelligence (AI) has revolutionized healthcare, helping significant improvements in various applications. This paper provides a comprehensive review of foundation models in healthcare, highlighting their transformative potential in areas such as diagnostics, personalized treatment, and operational efficiency. We argue the key capabilities of these models, including their ability to process diverse data types such as medical images, clinical notes, and structured health records. Regardless their assurance, difficulties remain, including data privacy concerns, bias in AI algorithms, and the need for extensive computational resources. Our analysis identifies emerging trends and future directions, emphasizing the importance of ethical AI deployment, improved interoperability over healthcare systems, and the development of more robust, domain-specific models. Future research should focus on enhancing model interpretability, ensuring equitable access, and fostering collaboration between AI developers and healthcare professionals to maximize the advantages of these technologies.

Large blood vessel segmentation in quantitative DCE-MRI of brain tumors: A Swin UNETR approach

Brain tumor growth is associated with angiogenesis, wherein the density of newly developed blood vessels indicates tumor progression and correlates with the tumor grade. Quantitative dynamic contrast enhanced-magnetic resonance imaging (DCE-MRI) has shown potential in brain tumor grading and treatment response assessment. Segmentation of large-blood-vessels is crucial for automatic and accurate tumor grading using quantitative DCE-MRI. Traditional manual and semi-manual rule-based large-blood-vessel segmentation methods are time-intensive and prone to errors. This study proposes a novel deep learning-based technique for automatic large-blood-vessel segmentation using Swin UNETR architectures and comparing it with U-Net and Attention U-Net architectures. The study employed MRI data from 187 brain tumor patients, with training, validation, and testing datasets sourced from two centers, two vendors, and two field-strength magnetic resonance scanners. To test the generalizability of the developed model, testing was also carried out on different brain tumor types, including lymphoma and metastasis. Performance evaluation demonstrated that Swin UNETR outperformed other models in segmenting large-blood-vessel regions (achieving Dice scores of 0.979, and 0.973 on training and validation sets, respectively, with test set performance ranging from 0.835 to 0.982). Moreover, most quantitative parameters showed significant differences (p < 0.05) between with and without large-blood-vessel. After large-blood-vessel removal, using both ground truth and predicted masks, the values of parameters in non-vascular tumoral regions were statistically similar (p > 0.05). The proposed approach has potential applications in improving the accuracy of automatic grading of tumors as well as in treatment planning.

Deep learning-based automated segmentation of cardiac real-time MRI in non-human primates

Advanced imaging techniques, like magnetic resonance imaging (MRI), have revolutionised cardiovascular disease diagnosis and monitoring in humans and animal models. Real-time (RT) MRI, which can capture a single slice during each consecutive heartbeat while the animal or patient breathes continuously, generates large data sets that necessitate automatic myocardium segmentation to fully exploit these technological advancements. While automatic segmentation is common in human adults, it remains underdeveloped in preclinical animal models. In this study, we developed and trained a fully automated 2D convolutional neural network (CNN) for segmenting the left and right ventricles and the myocardium in non-human primates (NHPs) using RT cardiac MR images of rhesus macaques, in the following referred to as PrimUNet. Based on the U-Net framework, PrimUNet achieved optimal performance with a learning rate of 0.0001, an initial kernel size of 64, a final kernel size of 512, and a batch size of 32. It attained an average Dice score of 0.9, comparable to human studies. Testing PrimUNet on additional RT MRI data from rhesus macaques demonstrated strong agreement with manual segmentation for left ventricular end-diastolic volume (LVEDV), left ventricular end-systolic volume (LVESV), and left ventricular myocardial volume (LVMV). It also performs well on cine MRI data of rhesus macaques and acceptably on those of baboons. PrimUNet is well-suited for automatically segmenting extensive RT MRI data, facilitating strain analyses of individual heartbeats. By eliminating human observer variability, PrimUNet enhances the reliability and reproducibility of data analysis in animal research, thereby advancing translational cardiovascular studies.

Merging Context Clustering With Visual State Space Models for Medical Image Segmentation

Medical image segmentation demands the aggregation of global and local feature representations, posing a challenge for current methodologies in handling both long-range and short-range feature interactions. Recently, vision mamba (ViM) models have emerged as promising solutions for addressing model complexities by excelling in long-range feature iterations with linear complexity. However, existing ViM approaches overlook the importance of preserving short-range local dependencies by directly flattening spatial tokens and are constrained by fixed scanning patterns that limit the capture of dynamic spatial context information. To address these challenges, we introduce a simple yet effective method named context clustering ViM (CCViM), which incorporates a context clustering module within the existing ViM models to segment image tokens into distinct windows for adaptable local clustering. Our method effectively combines long-range and short-range feature interactions, thereby enhancing spatial contextual representations for medical image segmentation tasks. Extensive experimental evaluations on diverse public datasets, i.e., Kumar, CPM17, ISIC17, ISIC18, and Synapse, demonstrate the superior performance of our method compared to current state-of-the-art methods. Our code can be found at https://github.com/zymissy/CCViM.

Ordinal Unsupervised Domain Adaptation With Recursively Conditional Gaussian Imposed Variational Disentanglement

There has been a growing interest in unsupervised domain adaptation (UDA) to alleviate the data scalability issue, while the existing works usually focus on classifying independently discrete labels. However, in many tasks (e.g., medical diagnosis), the labels are discrete and successively distributed. The UDA for ordinal classification requires inducing non-trivial ordinal distribution prior to the latent space. Target for this, the partially ordered set (poset) is defined for constraining the latent vector. Instead of the typically i.i.d. Gaussian latent prior, in this work, a recursively conditional Gaussian (RCG) set is proposed for ordered constraint modeling, which admits a tractable joint distribution prior. Furthermore, we are able to control the density of content vectors that violate the poset constraint by a simple "three-sigma rule." We explicitly disentangle the cross-domain images into a shared ordinal prior induced ordinal content space and two separate source/target ordinal-unrelated spaces, and the self-training is worked on the shared space exclusively for ordinal-aware domain alignment. Extensive experiments on UDA medical diagnoses and facial age estimation demonstrate its effectiveness.

Explainable CNN for brain tumor detection and classification through XAI based key features identification

Despite significant advancements in brain tumor classification, many existing models suffer from complex structures that make them difficult to interpret. This complexity can hinder the transparency of the decision-making process, causing models to rely on irrelevant features or normal soft tissues. Besides, these models often include additional layers and parameters, which further complicate the classification process. Our work addresses these limitations by introducing a novel methodology that combines Explainable AI (XAI) techniques with a Convolutional Neural Network (CNN) architecture. The major contribution of this paper is ensuring that the model focuses on the most relevant features for tumor detection and classification, while simultaneously reducing complexity, by minimizing the number of layers. This approach enhances the model's transparency and robustness, giving clear insights into its decision-making process through XAI techniques such as Gradient-weighted Class Activation Mapping (Grad-Cam), Shapley Additive explanations (Shap), and Local Interpretable Model-agnostic Explanations (LIME). Additionally, the approach demonstrates better performance, achieving 99% accuracy on seen data and 95% on unseen data, highlighting its generalizability and reliability. This balance of simplicity, interpretability, and high accuracy represents a significant advancement in the classification of brain tumor.

Evolving and Novel Applications of Artificial Intelligence in Cancer Imaging

Artificial intelligence (AI) is revolutionizing cancer imaging, enhancing screening, diagnosis, and treatment options for clinicians. AI-driven applications, particularly deep learning and machine learning, excel in risk assessment, tumor detection, classification, and predictive treatment prognosis. Machine learning algorithms, especially deep learning frameworks, improve lesion characterization and automated segmentation, leading to enhanced radiomic feature extraction and delineation. Radiomics, which quantifies imaging features, offers personalized treatment response predictions across various imaging modalities. AI models also facilitate technological improvements in non-diagnostic tasks, such as image optimization and automated medical reporting. Despite advancements, challenges persist in integrating AI into healthcare, tracking accurate data, and ensuring patient privacy. Validation through clinician input and multi-institutional studies is essential for patient safety and model generalizability. This requires support from radiologists worldwide and consideration of complex regulatory processes. Future directions include elaborating on existing optimizations, integrating advanced AI techniques, improving patient-centric medicine, and expanding healthcare accessibility. AI can enhance cancer imaging, optimizing precision medicine and improving patient outcomes. Ongoing multidisciplinary collaboration between radiologists, oncologists, software developers, and regulatory bodies is crucial for AI's growing role in clinical oncology. This review aims to provide an overview of the applications of AI in oncologic imaging while also discussing their limitations.

Medical Image Segmentation: A Comprehensive Review of Deep Learning-Based Methods

Medical image segmentation is a critical application of computer vision in the analysis of medical images. Its primary objective is to isolate regions of interest in medical images from the background, thereby assisting clinicians in accurately identifying lesions, their sizes, locations, and their relationships with surrounding tissues. However, compared to natural images, medical images present unique challenges, such as low resolution, poor contrast, inconsistency, and scattered target regions. Furthermore, the accuracy and stability of segmentation results are subject to more stringent requirements. In recent years, with the widespread application of Convolutional Neural Networks (CNNs) in computer vision, deep learning-based methods for medical image segmentation have become a focal point of research. This paper categorizes, reviews, and summarizes the current representative methods and research status in the field of medical image segmentation. A comparative analysis of relevant experiments is presented, along with an introduction to commonly used public datasets, performance evaluation metrics, and loss functions in medical image segmentation. Finally, potential future research directions and development trends in this field are predicted and analyzed.

From pixels to prognosis: leveraging radiomics and machine learning to predict IDH1 genotype in gliomas

Gliomas are the most common primary tumors of the central nervous system, and advances in genetics and molecular medicine have significantly transformed their classification and treatment. This study aims to predict the IDH1 genotype in gliomas using radiomics and machine learning (ML) methods. Retrospective data from 108 glioma patients were analyzed, including MRI data supported by demographic details such as age, sex, and comorbidities. Tumor segmentation was manually performed using 3D Slicer software, and 112 radiomic features were extracted with the PyRadiomics library. Feature selection using the mRMR algorithm identified 17 significant radiomic features. Various ML algorithms, including KNN, Ensemble, DT, LR, Discriminant and SVM, were applied to predict the IDH1 genotype. The KNN and Ensemble models achieved the highest sensitivity (92-100%) and specificity (100%), emerging as the most successful models. Comparative analyses demonstrated that KNN achieved an accuracy of 92.59%, sensitivity of 92.38%, specificity of 100%, precision of 100%, and an F1-score of 95.02%. Similarly, the Ensemble model achieved an accuracy of 90.74%, sensitivity of 90.65%, specificity of 100%, precision of 100%, and an F1-score of 95.13%. To evaluate their effectiveness, KNN and Ensemble models were compared with commonly used machine learning approaches in glioma classification. LR, a conventional statistical approach, exhibited lower predictive performance with an accuracy of 79.63%, while SVM, a frequently utilized ML model for radiomics-based tumor classification, achieved an accuracy of 85.19%. Our findings are consistent with previous research indicating that radiomics-based ML models achieve high accuracy in IDH1 mutation prediction, with reported performances typically exceeding 80%. These findings suggest that KNN and Ensemble models are more effective in capturing the non-linear radiomic patterns associated with IDH1 status, compared to traditional ML approaches. Our findings indicate that radiomic analyses provide comprehensive genotypic classification by assessing the entire tumor and present a safer, faster, and more patient-friendly alternative to traditional biopsies. This study highlights the potential of radiomics and ML techniques, particularly KNN, Ensemble, and SVM, as powerful tools for predicting the molecular characteristics of gliomas and developing personalized treatment strategies.

The performance of sonographic antenatal birth weight assessment assisted with artificial intelligence compared to that of manual examiners at term

PURPOSE

The aim of this study is to investigate the differences in the accuracy of sonographic antenatal fetal weight estimation at term with artificial intelligence (AI) compared to that of clinical sonographers at different levels of experience.

METHODS

This is a prospective cohort study where pregnant women at term scheduled for an imminent elective cesarean section were recruited. Three independent antenatal fetal weight estimations for each fetus were blindly measured by an experienced resident physician with level I qualification from the German Society for Ultrasound in Medicine (group 1), a senior physician with level II qualification (group 2), and an AI-supported algorithm (group 3) using Hadlock formula 3. The differences between the three groups and the actual birth weight were examined with a paired t-test. A variation within 10% of birth weight was deemed accurate, and the diagnostic accuracies of both groups 1 and 3 compared to group 2 were assessed using receiver operating characteristic (ROC) curves. The association between accuracy and potential influencing factors including gestational age, fetal position, maternal age, maternal body mass index (BMI), twins, neonatal gender, placental position, gestational diabetes, and amniotic fluid index was tested with univariate logistic regression. A sensitivity analysis by inflating the estimated weights by daily 25 grams (g) gain for days between examination and birth was conducted.

RESULTS

300 fetuses at a mean gestational week of 38.7 ± 1.1 were included in this study and examined on median 2 (2-4) days prior to delivery. Average birth weight was 3264.6 ± 530.7 g and the mean difference of the sonographic estimated fetal weight compared to birthweight was -203.6 ± 325.4 g, -132.2 ± 294.1 g, and -338.4 ± 606.2 g for groups 1, 2, and 3 respectively. The estimated weight was accurate in 62% (56.2%, 67.5%), 70% (64.5%, 75,1%), and 48.3% (42.6%, 54.1%) for groups 1, 2, and 3 respectively. The diagnostic accuracy measures for groups 1 and 3 compared to group 2 resulted in 55.7% (48.7%, 62.5%) and 68.6% (61.8%, 74.8%) sensitivity, 68.9% (58.3%, 78.2%) and 53.3% (42.5%, 63.9%) specificity and 0.62 (0.56, 0.68) and 0.61 (0.55, 0.67) area under the ROC curves respectively. There was no association between accuracy and the investigated variables. Adjusting for sensitivity analysis increased the accuracy to 68% (62.4%, 73.2%), 75% (69.7%, 79.8%), and 51.3% (45.5%, 57.1%), and changed the mean difference compared to birth weight to -136.1 ± 321.8 g, -64.7 ± 291.2 g, and -270.7 ± 605.2 g for groups 1, 2, and 3 respectively.

CONCLUSION

The antenatal weight estimation by experienced specialists with high-level qualifications remains the gold standard and provides the highest precision. Nevertheless, the accuracy of this standard is less than 80% even after adjusting for daily weight gain. The tested AI-supported method exhibits high variability and requires optimization and validation before being reliably used in clinical practice.

Deep learning-assisted detection of meniscus and anterior cruciate ligament combined tears in adult knee magnetic resonance imaging: a crossover study with arthroscopy correlation

AIM

We aimed to compare the diagnostic performance of physicians in the detection of arthroscopically confirmed meniscus and anterior cruciate ligament (ACL) tears on knee magnetic resonance imaging (MRI), with and without assistance from a deep learning (DL) model.

METHODS

We obtained preoperative MR images from 88 knees of patients who underwent arthroscopic meniscal repair, with or without ACL reconstruction. Ninety-eight MR images of knees without signs of meniscus or ACL tears were obtained from a publicly available database after matching on age and ACL status (normal or torn), resulting in a global dataset of 186 MRI examinations. The Keros® (Incepto, Paris) DL algorithm, previously trained for the detection and characterization of meniscus and ACL tears, was used for MRI assessment. Magnetic resonance images were individually, and blindly annotated by three physicians and the DL algorithm. After three weeks, the three human raters repeated image assessment with model assistance, performed in a different order.

RESULTS

The Keros® algorithm achieved an area under the curve (AUC) of 0.96 (95% CI 0.93, 0.99), 0.91 (95% CI 0.85, 0.96), and 0.99 (95% CI 0.98, 0.997) in the detection of medial meniscus, lateral meniscus and ACL tears, respectively. With model assistance, physicians achieved higher sensitivity (91% vs. 83%, p = 0.04) and similar specificity (91% vs. 87%, p = 0.09) in the detection of medial meniscus tears. Regarding lateral meniscus tears, sensitivity and specificity were similar with/without model assistance. Regarding ACL tears, physicians achieved higher specificity when assisted by the algorithm (70% vs. 51%, p = 0.01) but similar sensitivity with/without model assistance (93% vs. 96%, p = 0.13).

CONCLUSIONS

The current model consistently helped physicians in the detection of medial meniscus and ACL tears, notably when they were combined.

LEVEL OF EVIDENCE

Diagnostic study, Level III.

Impact of fine-tuning parameters of convolutional neural network for skin cancer detection

Melanoma skin cancer is a deadly disease with a high mortality rate. A prompt diagnosis can aid in the treatment of the disease and potentially save the patient's life. Artificial intelligence methods can help diagnose cancer at a rapid speed. The literature has employed numerous Machine Learning (ML) and Deep Learning (DL) algorithms to detect skin cancer. ML algorithms perform well for small datasets but cannot comprehend larger ones. Conversely, DL algorithms exhibit strong performance on large datasets but misclassify when applied to smaller ones. We conduct extensive experiments using a convolutional neural network (CNN), varying its parameter values to determine which set of values yields the best performance measure. We discovered that adding layers, making each Conv2D layer have multiple filters, and getting rid of dropout layers greatly improves the accuracy of the classifiers, going from 62.5% to 85%. We have also discussed the parameters that have the potential to significantly impact the model's performance. This shows how powerful it is to fine-tune the parameters of a CNN-based model. These findings can assist researchers in fine-tuning their CNN-based models for use with skin cancer image datasets.

Personalized Medical Approach in Gastrointestinal Surgical Oncology: Current Trends and Future Perspectives

Advances in artificial intelligence (AI), multi-omic profiling, and sophisticated imaging technologies have significantly advanced personalized medicine in gastrointestinal surgical oncology. These technological innovations enable precise patient stratification, tailored surgical strategies, and individualized therapeutic approaches, thereby significantly enhancing clinical outcomes. Despite remarkable progress, challenges persist, including the standardization and integration of diverse data types, ethical concerns regarding patient privacy, and rigorous clinical validation of predictive models. Addressing these challenges requires establishing international standards for data interoperability, such as Fast Healthcare Interoperability Resources, and adopting advanced security methods, such as homomorphic encryption, to facilitate secure multi-institutional data sharing. Moreover, ensuring model transparency and explainability through techniques such as explainable AI is critical for fostering trust among clinicians and patients. The successful integration of these advanced technologies necessitates strong multidisciplinary collaboration among surgeons, radiologists, geneticists, pathologists, and oncologists. Ultimately, the continued development and effective implementation of these personalized medical strategies complemented by human expertise promise a transformative shift toward patient-centered care, improving long-term outcomes for patients with gastrointestinal cancer.

The use of a convolutional neural network to automate radiologic scoring of computed tomography of paranasal sinuses

BACKGROUND

Chronic rhinosinusitis (CRS) is diagnosed with symptoms and objective endoscopy or computed tomography (CT). The Lund-Mackay score (LMS) is often used to determine the radiologic severity of CRS and make clinical decisions. This proof-of-concept study aimed to develop an automated algorithm combining a convolutional neural network (CNN) for sinus segmentation with post-processing to compute LMS directly from CT scans.

RESULTS

Radiology Information System was queried for outpatient paranasal sinus CTs at a tertiary institution. We identified 1,399 CT scans which were manually labelled with LMS of individual sinuses. Seventy-seven CT scans with 13,668 coronal images were segmented manually for individual sinuses. Our model for segmentation achieved a mean Dice score of 0.85 for all sinus regions, except for the osteomeatal complex. For individual Dice scores were 0.95, 0.71, 0.78, 0.93, 0.86 for the maxillary, anterior ethmoid, posterior ethmoid, sphenoid, and frontal sinuses, respectively. LMS was computed automatically by applying adaptive image thresholding and pixel counting to the CNN's segmented regions. A convolutional neural network (CNN) model was trained to segment each sinus region. Overall, the LMS model showed a high degree of accuracy with a score of 0.92, 0.99, 0.99, 0.97, 0.99, 0.86 for the maxillary, anterior ethmoid, posterior ethmoid, sphenoid, and frontal sinuses, respectively.

CONCLUSIONS

Reporting of paranasal sinus CT can be automated and potentially standardized with a CNN model to provide accurate Lund-Mackay score.

A deep learning-based multimodal medical imaging model for breast cancer screening

In existing breast cancer prediction research, most models rely solely on a single type of imaging data, which limits their performance. To overcome this limitation, the present study explores breast cancer prediction models based on multimodal medical images (mammography and ultrasound images) and compares them with single-modal models. We collected medical imaging data from 790 patients, including 2,235 mammography images and 1,348 ultrasound images, and conducted a comparison using six deep learning classification models to identify the best model for constructing the multimodal classification model. Performance was evaluated using metrics such as area under the receiver operating characteristic curve (AUC), sensitivity, specificity, precision, and accuracy to compare the multimodal and single-modal classification models. Experimental results demonstrate that the multimodal classification model outperforms single-modal models in terms of specificity (96.41% (95% CI:93.10%-99.72%)), accuracy (93.78% (95% CI:87.67%-99.89%)), precision (83.66% (95% CI:76.27%-91.05%)), and AUC (0.968 (95% CI:0.947-0.989)), while single-modal models excel in sensitivity. Additionally, heatmap visualization was used to further validate the classification performance of the multimodal model. In conclusion, our multimodal classification model shows strong potential in breast cancer screening tasks, effectively assisting physicians in improving screening accuracy.

Bayesian-Optimized Convolutional Neural Networks for Classifying Primary Tumor Origin of Brain Metastases from MRI

Background/Objectives: This study evaluates whether convolutional neural networks (CNNs) can be trained to determine the primary tumor origin from MRI images alone in patients with metastatic brain lesions. Methods: This retrospective, monocentric study involved the segmentation of 1175 brain lesions from MRI scans of 436 patients with histologically confirmed primary tumor origins. The four most common tumor types-lung adenocarcinoma, small cell lung cancer, breast cancer, and melanoma-were selected, and a class-balanced dataset was created through under-sampling. This resulted in 276 training datasets and 88 hold-out test datasets. Bayesian optimization was employed to determine the optimal CNN architecture, the most relevant imaging sequences, and whether the masking of images was necessary. We compared the performance of the CNN with that of two expert radiologists specializing in neuro-oncological imaging. Results: The best-performing CNN from the Bayesian optimization process used masked images across all available MRI sequences. It achieved Area-Under-the-Curve (AUC) values of 0.75 for melanoma, 0.65 for small cell lung cancer, 0.64 for breast cancer, and 0.57 for lung adenocarcinoma. Masked images likely improved performance by focusing the CNN on relevant regions and reducing noise from surrounding tissues. In comparison, Radiologist 1 achieved AUCs of 0.55, 0.52, 0.45, and 0.51, and Radiologist 2 achieved AUCs of 0.68, 0.55, 0.64, and 0.43 for the same tumor types, respectively. The CNN consistently showed higher accuracy, particularly for melanoma and breast cancer. Conclusions: Bayesian optimization enabled the creation of a CNN that outperformed expert radiologists in classifying the primary tumor origin of brain metastases from MRI.

Diagnosis Test Accuracy of Artificial Intelligence for Endometrial Cancer: Systematic Review and Meta-Analysis

BACKGROUND

Endometrial cancer is one of the most common gynecological tumors, and early screening and diagnosis are crucial for its treatment. Research on the application of artificial intelligence (AI) in the diagnosis of endometrial cancer is increasing, but there is currently no comprehensive meta-analysis to evaluate the diagnostic accuracy of AI in screening for endometrial cancer.

OBJECTIVE

This paper presents a systematic review of AI-based endometrial cancer screening, which is needed to clarify its diagnostic accuracy and provide evidence for the application of AI technology in screening for endometrial cancer.

METHODS

A search was conducted across PubMed, Embase, Cochrane Library, Web of Science, and Scopus databases to include studies published in English, which evaluated the performance of AI in endometrial cancer screening. A total of 2 independent reviewers screened the titles and abstracts, and the quality of the selected studies was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool. The certainty of the diagnostic test evidence was evaluated using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) system.

RESULTS

A total of 13 studies were included, and the hierarchical summary receiver operating characteristic model used for the meta-analysis showed that the overall sensitivity of AI-based endometrial cancer screening was 86% (95% CI 79%-90%) and specificity was 92% (95% CI 87%-95%). Subgroup analysis revealed similar results across AI type, study region, publication year, and study type, but the overall quality of evidence was low.

CONCLUSIONS

AI-based endometrial cancer screening can effectively detect patients with endometrial cancer, but large-scale population studies are needed in the future to further clarify the diagnostic accuracy of AI in screening for endometrial cancer.

TRIAL REGISTRATION

PROSPERO CRD42024519835; https://www.crd.york.ac.uk/PROSPERO/view/CRD42024519835.

Anchor point based image registration for absolute scale topographic structure detection in microscopy

Microscopy images obtained through remote sensing often suffer from misalignment and deformation, complicating accurate data analysis. As experimental instruments improve and scientific discoveries deepen, the volume of data requiring processing continues to grow. Image registration can contribute to microscopy automation, which enables more efficient data analysis and experimental workflows. For this implementation, image processing techniques that can handle both image registration and localized object analysis are required. This research introduces a computer interface designed to calibrate and analyze specific structures with prior knowledge of the observed target. Our method achieves image registration by aligning anchor points, which correspond to the coordinates of a structural model within the image. It employs homography transform to correct images, restoring them to their original, undistorted form, thus enabling consistent quantitative comparisons across different images on an absolute scale. Additionally, the method provides valuable information from the registered anchor points, allowing for the precise localization of local objects in the structure. We demonstrate this technique across various microscopy scenarios at different scales and evaluate its precision against a keypoint detection AI approach from our previous research, which promises its enhancement in microscopy data analysis and automation.

A fully automated, expert-perceptive image quality assessment system for whole-body [18F]FDG PET/CT

BACKGROUND

The quality of clinical PET/CT images is critical for both accurate diagnosis and image-based research. However, current image quality assessment (IQA) methods predominantly rely on handcrafted features and region-specific analyses, thereby limiting automation in whole-body and multicenter evaluations. This study aims to develop an expert-perceptive deep learning-based IQA system for [18F]FDG PET/CT to tackle the lack of automated, interpretable assessments of clinical whole-body PET/CT image quality.

METHODS

This retrospective multicenter study included clinical whole-body [18F]FDG PET/CT scans from 718 patients. Automated identification and localization algorithms were applied to select predefined pairs of PET and CT slices from whole-body images. Fifteen experienced experts, trained to conduct blinded slice-level subjective assessments, provided average visual scores as reference standards. Using the MANIQA framework, the developed IQA model integrates the Vision Transformer, Transposed Attention, and Scale Swin Transformer Blocks to categorize PET and CT images into five quality classes. The model's correlation, consistency, and accuracy with expert evaluations on both PET and CT test sets were statistically analysed to assess the system's IQA performance. Additionally, the model's ability to distinguish high-quality images was evaluated using receiver operating characteristic (ROC) curves.

RESULTS

The IQA model demonstrated high accuracy in predicting image quality categories and showed strong concordance with expert evaluations of PET/CT image quality. In predicting slice-level image quality across all body regions, the model achieved an average accuracy of 0.832 for PET and 0.902 for CT. The model's scores showed substantial agreement with expert assessments, achieving average Spearman coefficients (ρ) of 0.891 for PET and 0.624 for CT, while the average Intraclass Correlation Coefficient (ICC) reached 0.953 for PET and 0.92 for CT. The PET IQA model demonstrated strong discriminative performance, achieving an area under the curve (AUC) of ≥ 0.88 for both the thoracic and abdominal regions.

CONCLUSIONS

This fully automated IQA system provides a robust and comprehensive framework for the objective evaluation of clinical image quality. Furthermore, it demonstrates significant potential as an impartial, expert-level tool for standardised multicenter clinical IQA.

A comprehensive case study of deep learning on the detection of alpha thalassemia and beta thalassemia using public and private datasets

This study explores the performance of deep learning models, specifically Convolutional Neural Networks (CNN) and XGBoost, in predicting alpha and beta thalassemia using both public and private datasets. Thalassemia is a genetic disorder that impairs hemoglobin production, leading to anemia and other health complications. Early diagnosis is essential for effective management and prevention of severe health issues. The study applied CNN and XGBoost to two case studies: one for alpha-thalassemia and the other for beta-thalassemia. Public datasets were sourced from medical databases, while private datasets were collected from clinical records, offering a more comprehensive feature set and larger sample sizes. After data preprocessing and splitting, model performance was evaluated. XGBoost achieved 99.34% accuracy on the private dataset for alpha thalassemia, while CNN reached 98.10% accuracy on the private dataset for beta-thalassemia. The superior performance on private datasets was attributed to better data quality and volume. This study highlights the effectiveness of deep learning in medical diagnostics, demonstrating that high-quality data can significantly enhance the predictive capabilities of AI models. By integrating CNN and XGBoost, this approach offers a robust method for detecting thalassemia, potentially improving early diagnosis and reducing disease-related mortality.

Deep learning model DeepNeo predicts neointimal tissue characterization using optical coherence tomography

BACKGROUND

Accurate interpretation of optical coherence tomography (OCT) pullbacks is critical for assessing vascular healing after percutaneous coronary intervention (PCI). Manual analysis is time-consuming and subjective, highlighting the need for a fully automated solution.

METHODS

In this study, 1148 frames from 92 OCT pullbacks were manually annotated to classify neointima as homogeneous, heterogeneous, neoatherosclerosis, or not analyzable on a quadrant level. Stent and lumen contours were annotated in 305 frames for segmentation of the lumen, stent struts, and neointima. We used these annotations to train a deep learning algorithm called DeepNeo. Performance was further evaluated in an animal model (male New Zealand White Rabbits) of neoatherosclerosis using co-registered histopathology images as the gold standard.

RESULTS

DeepNeo demonstrates a strong classification performance for neointimal tissue, achieving an overall accuracy of 75%, which is comparable to manual classification accuracy by two clinical experts (75% and 71%). In the animal model of neoatherosclerosis, DeepNeo achieves an accuracy of 87% when compared with histopathological findings. For segmentation tasks in human pullbacks, the algorithm shows strong performance with mean Dice overlap scores of 0.99 for the lumen, 0.66 for stent struts, and 0.86 for neointima.

CONCLUSIONS

To the best of our knowledge, DeepNeo is the first deep learning algorithm enabling fully automated segmentation and classification of neointimal tissue with performance comparable to human experts. It could standardize vascular healing assessments after PCI, support therapeutic decisions, and improve risk detection for cardiac events.

Integrating Machine Learning and Deep Learning for Predicting Non-Surgical Root Canal Treatment Outcomes Using Two-Dimensional Periapical Radiographs

Background/Objectives: In a previous study, we utilized categorical variables and machine learning (ML) algorithms to predict the success of non-surgical root canal treatments (NSRCTs) in apical periodontitis (AP), classifying the outcome as either success (healed) or failure (not healed). Given the importance of radiographic imaging in diagnosis, the present study evaluates the efficacy of deep learning (DL) in predicting NSRCT outcomes using two-dimensional (2D) periapical radiographs, comparing its performance with ML models. Methods: The DL model was trained and validated using leave-one-out cross-validation (LOOCV). Its output was incorporated into the set of categorical variables, and the ML study was reproduced using backward stepwise selection (BSS). The chi-square test was applied to assess the association between this new variable and NSRCT outcomes. Finally, after identifying the best-performing method from the ML study reproduction, statistical comparisons were conducted between this method, clinical professionals, and the image-based model using Fisher's exact test. Results: The association study yielded a p-value of 0.000000127, highlighting the predictive capability of 2D radiographs. After incorporating the DL-based predictive variable, the ML algorithm that demonstrated the best performance was logistic regression (LR), differing from the previous study, where random forest (RF) was the top performer. When comparing the deep learning-logistic regression (DL-LR) model with the clinician's prognosis (DP), DL-LR showed superior performance with a statistically significant difference (p-value < 0.05) in sensitivity, NPV, and accuracy. The same trend was observed in the DL vs. DP comparison. However, no statistically significant differences were found in the comparisons of RF vs. DL-LR, RF vs. DL, or DL vs. DL-LR. Conclusions: The findings of this study suggest that image-based artificial intelligence models exhibit superior predictive capability compared with those relying exclusively on categorical data. Moreover, they outperform clinician prognosis.