With the advancement of modern medicine and the development of technologies such as MRI, CT, and cellular analysis, it has become increasingly critical for clinicians to accurately interpret various diagnostic images. However, modern medical education often faces challenges due to limited access to high-quality teaching materials, stemming from privacy concerns and a shortage of educational resources (Balogh et al., 2015). In this context, image data generated by machine learning models, particularly generative models, presents a promising solution. These models can create diverse and comparable imaging datasets without compromising patient privacy, thereby supporting modern medical education. In this study, we explore the use of convolutional neural networks (CNNs) and CycleGAN (Zhu et al., 2017) for generating synthetic medical images. The source code is available at https://github.com/mliuby/COMP4211-Project.

To develop and validate an automated diabetic macular oedema (DME) classification system based on the images from different three-dimensional optical coherence tomography (3-D OCT) devices. A multicentre, platform-based development study using retrospective and cross-sectional data. Data were subjected to a two-level grading system by trained graders and a retina specialist, and categorised into three types: no DME, non-centre-involved DME and centre-involved DME (CI-DME). The 3-D convolutional neural networks algorithm was used for DME classification system development. The deep learning (DL) performance was compared with the diabetic retinopathy experts. Data were collected from Joint Shantou International Eye Center of Shantou University and the Chinese University of Hong Kong, Chaozhou People's Hospital and The Second Affiliated Hospital of Shantou University Medical College from January 2010 to December 2023. 7790 volumes of 7146 eyes from 4254 patients were annotated, of which 6281 images were used as the development set and 1509 images were used as the external validation set, split based on the centres. Accuracy, F1-score, sensitivity, specificity, area under receiver operating characteristic curve (AUROC) and Cohen's kappa were calculated to evaluate the performance of the DL algorithm. In classifying DME with non-DME, our model achieved an AUROCs of 0.990 (95% CI 0.983 to 0.996) and 0.916 (95% CI 0.902 to 0.930) for hold-out testing dataset and external validation dataset, respectively. To distinguish CI-DME from non-centre-involved-DME, our model achieved AUROCs of 0.859 (95% CI 0.812 to 0.906) and 0.881 (95% CI 0.859 to 0.902), respectively. In addition, our system showed comparable performance (Cohen's κ: 0.85 and 0.75) to the retina experts (Cohen's κ: 0.58-0.92 and 0.70-0.71). Our DL system achieved high accuracy in multiclassification tasks on DME classification with 3-D OCT images, which can be applied to population-based DME screening.

The aim of this investigation is to assess the clinical usefulness of a machine learning model using contrast-enhanced ultrasound (CEUS) radiomics in discriminating clear cell renal cell carcinoma (ccRCC) from non-ccRCC. A total of 292 patients with pathologically confirmed RCC subtypes underwent CEUS (development set. n = 231; validation set, n = 61) in a retrospective study. Radiomics features were derived from CEUS images acquired during the cortical and parenchymal phases. Radiomics models were developed using logistic regression (LR), support vector machine, decision tree, naive Bayes, gradient boosting machine, and random forest. The suitable model was identified based on the area under the receiver operating characteristic curve (AUC). Appropriate clinical CEUS features were identified through univariate and multivariate LR analyses to develop a clinical model. By integrating radiomics and clinical CEUS features, a combined model was established. A comprehensive evaluation of the models' performance was conducted. After the reduction and selection process were applied to 2250 radiomics features, the final set of 8 features was considered valuable. Among the models, the LR model had the highest performance on the validation set and showed good robustness. In both the development and validation sets, both the radiomics (AUC, 0.946 and 0.927) and the combined models (AUC, 0.949 and 0.925) outperformed the clinical model (AUC, 0.851 and 0.768), showing higher AUC values (all p < 0.05). The combined model exhibited favorable calibration and clinical benefit. The combined model integrating clinical CEUS and CEUS radiomics features demonstrated good diagnostic performance in discriminating ccRCC from non-ccRCC.

To determine whether a machine learning model of voxel level [¹⁸f]fluorodeoxyglucose positron emission tomography (PET) data could predict progressive supranuclear palsy (PSP) pathology, as well as outperform currently available biomarkers. One hundred and thirty-seven autopsied patients with PSP (n = 42) and other neurodegenerative diseases (n = 95) who underwent antemortem [¹⁸f]fluorodeoxyglucose PET and 3.0 Tesla magnetic resonance imaging (MRI) scans were analyzed. A linear support vector machine was applied to differentiate pathological groups with sensitivity analyses performed to assess the influence of voxel size and region removal. A radial basis function was also prepared to create a secondary model using the most important voxels. The models were optimized on the main dataset (n = 104), and their performance was compared with the magnetic resonance parkinsonism index measured on MRI in the independent test dataset (n = 33). The model had the highest accuracy (0.91) and F-score (0.86) when voxel size was 6mm. In this optimized model, important voxels for differentiating the groups were observed in the thalamus, midbrain, and cerebellar dentate. The secondary models found the combination of thalamus and dentate to have the highest accuracy (0.89) and F-score (0.81). The optimized secondary model showed the highest accuracy (0.91) and F-scores (0.86) in the test dataset and outperformed the magnetic resonance parkinsonism index (0.81 and 0.70, respectively). The results suggest that glucose hypometabolism in the thalamus and cerebellar dentate have the highest potential for predicting PSP pathology. Our optimized machine learning model outperformed the best currently available biomarker to predict PSP pathology. ANN NEUROL 2025.

Uncertainty estimation methods are essential for the application of artificial intelligence (AI) models in medical image segmentation, particularly in addressing reliability and feasibility challenges in clinical deployment. Despite their significance, the adoption of uncertainty estimation methods in clinical practice remains limited due to the lack of a comprehensive evaluation framework tailored to their clinical usage. To address this gap, a simulation of uncertainty-assisted clinical workflows is conducted, highlighting the roles of uncertainty in model selection, sample screening, and risk visualization. Furthermore, uncertainty evaluation is extended to pixel, sample, and model levels to enable a more thorough assessment. At the pixel level, the Uncertainty Confusion Metric (UCM) is proposed, utilizing density curves to improve robustness against variability in uncertainty distributions and to assess the ability of pixel uncertainty to identify potential errors. At the sample level, the Expected Segmentation Calibration Error (ESCE) is introduced to provide more accurate calibration aligned with Dice, enabling more effective identification of low-quality samples. At the model level, the Harmonic Dice (HDice) metric is developed to integrate uncertainty and accuracy, mitigating the influence of dataset biases and offering a more robust evaluation of model performance on unseen data. Using this systematic evaluation framework, five mainstream uncertainty estimation methods are compared on organ and tumor datasets, providing new insights into their clinical applicability. Extensive experimental analyses validated the practicality and effectiveness of the proposed metrics. This study offers clear guidance for selecting appropriate uncertainty estimation methods in clinical settings, facilitating their integration into clinical workflows and ultimately improving diagnostic efficiency and patient outcomes.

Gastrointestinal (GI) endoscopy is essential in identifying GI tract abnormalities in order to detect diseases in their early stages and improve patient outcomes. Although deep learning has shown success in supporting GI diagnostics and decision-making, these models require curated datasets with labels that are expensive to acquire. Foundation models offer a promising solution by learning general-purpose representations, which can be finetuned for specific tasks, overcoming data scarcity. Developing foundation models for medical imaging holds significant potential, but the sensitive and protected nature of medical data presents unique challenges. Foundation model training typically requires extensive datasets, and while hospitals generate large volumes of data, privacy restrictions prevent direct data sharing, making foundation model training infeasible in most scenarios. In this work, we propose a FL framework for training foundation models for gastroendoscopy imaging, enabling data to remain within local hospital environments while contributing to a shared model. We explore several established FL algorithms, assessing their suitability for training foundation models without relying on task-specific labels, conducting experiments in both homogeneous and heterogeneous settings. We evaluate the trained foundation model on three critical downstream tasks--classification, detection, and segmentation--and demonstrate that it achieves improved performance across all tasks, highlighting the effectiveness of our approach in a federated, privacy-preserving setting.

Mammographic breast cancer detection depends on high-quality positioning, which is traditionally assessed and monitored subjectively. This study used artificial intelligence (AI) to evaluate mammography positioning on digital screening mammograms to identify and quantify unmet mammography positioning quality (MPQ). Data were collected within an IRB-approved collaboration. In total, 126 367 digital mammography studies (553 339 images) were processed. Unmet MPQ criteria, including exaggeration, portion cutoff, posterior tissue missing, nipple not in profile, too high on image receptor, inadequate pectoralis length, sagging, and posterior nipple line (PNL) length difference, were evaluated using MPQ AI algorithms. The similarity of unmet MPQ occurrence and rank order was compared for each health system. Altogether, 163 759 and 219 785 unmet MPQ criteria were identified, respectively, at the health systems. The rank order and the probability distribution of the unmet MPQ criteria were not statistically significantly different between health systems (P = .844 and P = .92, respectively). The 3 most-common unmet MPQ criteria were: short PNL length on the craniocaudal (CC) view, inadequate pectoralis muscle, and excessive exaggeration on the CC view. The percentages of unmet positioning criteria out of the total potential unmet positioning criteria at health system 1 and health system 2 were 8.4% (163 759/1 949 922) and 7.3% (219 785/3 030 129), respectively. Artificial intelligence identified a similar distribution of unmet MPQ criteria in 2 health systems' daily work. Knowledge of current commonly unmet MPQ criteria can facilitate the improvement of mammography quality through tailored education strategies.

Physics-driven deep learning (PD-DL) models have proven to be a powerful approach for improved reconstruction of rapid MRI scans. In order to train these models in scenarios where fully-sampled reference data is unavailable, self-supervised learning has gained prominence. However, its application at high acceleration rates frequently introduces artifacts, compromising image fidelity. To mitigate this shortcoming, we propose a novel way to train PD-DL networks via carefully-designed perturbations. In particular, we enhance the k-space masking idea of conventional self-supervised learning with a novel consistency term that assesses the model's ability to accurately predict the added perturbations in a sparse domain, leading to more reliable and artifact-free reconstructions. The results obtained from the fastMRI knee and brain datasets show that the proposed training strategy effectively reduces aliasing artifacts and mitigates noise amplification at high acceleration rates, outperforming state-of-the-art self-supervised methods both visually and quantitatively.

Recent advances in deep learning-based medical image registration have shown that training deep neural networks~(DNNs) does not necessarily require medical images. Previous work showed that DNNs trained on randomly generated images with carefully designed noise and contrast properties can still generalize well to unseen medical data. Building on this insight, we propose using registration between random images as a proxy task for pretraining a foundation model for image registration. Empirical results show that our pretraining strategy improves registration accuracy, reduces the amount of domain-specific data needed to achieve competitive performance, and accelerates convergence during downstream training, thereby enhancing computational efficiency.

The U-shaped encoder-decoder architecture with skip connections has become a prevailing paradigm in medical image segmentation due to its simplicity and effectiveness. While many recent works aim to improve this framework by designing more powerful encoders and decoders, employing advanced convolutional neural networks (CNNs) for local feature extraction, Transformers or state space models (SSMs) such as Mamba for global context modeling, or hybrid combinations of both, these methods often struggle to fully utilize pretrained vision backbones (e.g., ResNet, ViT, VMamba) due to structural mismatches. To bridge this gap, we introduce ACM-UNet, a general-purpose segmentation framework that retains a simple UNet-like design while effectively incorporating pretrained CNNs and Mamba models through a lightweight adapter mechanism. This adapter resolves architectural incompatibilities and enables the model to harness the complementary strengths of CNNs and SSMs-namely, fine-grained local detail extraction and long-range dependency modeling. Additionally, we propose a hierarchical multi-scale wavelet transform module in the decoder to enhance feature fusion and reconstruction fidelity. Extensive experiments on the Synapse and ACDC benchmarks demonstrate that ACM-UNet achieves state-of-the-art performance while remaining computationally efficient. Notably, it reaches 85.12% Dice Score and 13.89mm HD95 on the Synapse dataset with 17.93G FLOPs, showcasing its effectiveness and scalability. Code is available at: https://github.com/zyklcode/ACM-UNet.