This study conducts a comprehensive comparison of four neural network architectures: Convolutional Neural Network, Capsule Network, Convolutional Kolmogorov--Arnold Network, and the newly proposed Capsule--Convolutional Kolmogorov--Arnold Network. The proposed Capsule-ConvKAN architecture combines the dynamic routing and spatial hierarchy capabilities of Capsule Network with the flexible and interpretable function approximation of Convolutional Kolmogorov--Arnold Networks. This novel hybrid model was developed to improve feature representation and classification accuracy, particularly in challenging real-world biomedical image data. The architectures were evaluated on a histopathological image dataset, where Capsule-ConvKAN achieved the highest classification performance with an accuracy of 91.21\%. The results demonstrate the potential of the newly introduced Capsule-ConvKAN in capturing spatial patterns, managing complex features, and addressing the limitations of traditional convolutional models in medical image classification.

In recent years, artificial intelligence has significantly advanced medical image segmentation. However, challenges remain, including efficient 3D medical image processing across diverse modalities and handling data variability. In this work, we introduce Hierarchical Soft Mixture-of-Experts (HoME), a two-level token-routing layer for efficient long-context modeling, specifically designed for 3D medical image segmentation. Built on the Mamba state-space model (SSM) backbone, HoME enhances sequential modeling through sparse, adaptive expert routing. The first stage employs a Soft Mixture-of-Experts (SMoE) layer to partition input sequences into local groups, routing tokens to specialized per-group experts for localized feature extraction. The second stage aggregates these outputs via a global SMoE layer, enabling cross-group information fusion and global context refinement. This hierarchical design, combining local expert routing with global expert refinement improves generalizability and segmentation performance, surpassing state-of-the-art results across datasets from the three most commonly used 3D medical imaging modalities and data quality.

Plain X-ray is one of the most common image modalities for clinical diagnosis (e.g. bone fracture, pneumonia, cancer screening, etc.). X-ray image segmentation is an essential step for many computer-aided diagnostic systems, yet it remains challenging. Deep-learning-based methods have achieved superior performance in medical image segmentation tasks but often require a large amount of high-quality annotated data for model training. Providing such an annotated dataset is not only time-consuming but also requires a high level of expertise. This is particularly challenging in wrist bone segmentation in X-rays, due to the interposition of multiple small carpal bones in the image. To overcome the data annotation issue, this work utilizes a large number of simulated X-ray images generated from Computed Tomography (CT) volumes with their corresponding 10 bone labels to train a deep learning-based model for wrist bone segmentation in real X-ray images. The proposed method was evaluated using both simulated images and real images. The method achieved Dice scores ranging from 0.80 to 0.92 for the simulated dataset generated from different view angles. Qualitative analysis of the segmentation results of the real X-ray images also demonstrated the superior performance of the trained model. The trained model and X-ray simulation code are freely available for research purposes: the link will be provided upon acceptance.

Breast density assessment is a crucial component of mammographic interpretation, with high breast density (BI-RADS categories C and D) representing both a significant risk factor for developing breast cancer and a technical challenge for tumor detection. This study proposes an automated deep learning system for robust binary classification of breast density (low: A/B vs. high: C/D) using the VinDr-Mammo dataset. We implemented and compared four advanced convolutional neural networks: ResNet18, ResNet50, EfficientNet-B0, and DenseNet121, each enhanced with channel attention mechanisms. To address the inherent class imbalance, we developed a novel Combined Focal Label Smoothing Loss function that integrates focal loss, label smoothing, and class-balanced weighting. Our preprocessing pipeline incorporated advanced techniques, including contrast-limited adaptive histogram equalization (CLAHE) and comprehensive data augmentation. The individual models were combined through an optimized ensemble voting approach, achieving superior performance (AUC: 0.963, F1-score: 0.952) compared to any single model. This system demonstrates significant potential to standardize density assessments in clinical practice, potentially improving screening efficiency and early cancer detection rates while reducing inter-observer variability among radiologists.

To predict the normalized glandular dose (DgN) coefficients and the related uncertainty in mammography and digital breast tomosynthesis (DBT) using a machine learning algorithm and patient-like digital breast models. 126 patient-like digital breast phantoms were used for DgN Monte Carlo ground truth calculations. An Automatic Relevance Determination Regression algorithm was used to predict DgN from anatomical breast features. These features included compressed breast thickness, glandular fraction by volume, glandular volume, center of mass and standard deviation of the glandular tissue distribution in the cranio-caudal direction. An algorithm for data imputation was explored to account for avoiding the use of the latter two features. 5-fold cross validation showed that the predictive model provides an estimation of DgN with 1% average difference from the ground truth; this difference was less than 3% in 50% of the cases. The average uncertainty of the estimated DgN values was 9%. Excluding the information related to the glandular distribution increased this uncertainty to 17% without inducing a significant discrepancy in estimated DgN values, with half of the predicted cases differing from the ground truth by less than 9%. The data imputation algorithm reduced the estimated uncertainty, without restoring the original performance. Predictive performance improved by increasing tube voltage. The proposed methodology predicts the DgN in mammography and DBT for patient-derived breasts with an uncertainty below 9%. Predicting test evaluations reported 1% average difference from the ground truth, with 50% of the cohort cases differing by less than 5%.

Medical self-supervised learning eliminates the reliance on labels, making feature extraction simple and efficient. The intricate design of pretext tasks in single-modal self-supervised analysis presents challenges, however, compounded by an excessive dependency on data augmentation, leading to a bottleneck in medical self-supervised learning research. Consequently, this paper reanalyzes the feature learnability introduced by data augmentation strategies in medical image self-supervised learning. We introduce an adaptive self-supervised learning data augmentation method from the perspective of batch fusion. Moreover, we propose a conv embedding block for learning the incremental representation between these batches. We tested 5 fused data tasks proposed by previous researchers and it achieved a linear classification protocol accuracy of 94.25% with only 150 self-supervised feature training in Vision Transformer(ViT), which is the best among the same methods. With a detailed ablation study on previous augmentation strategies, the results indicate that the proposed medical data augmentation strategy in this paper effectively represents ultrasound data features in the self-supervised learning process. The code and weights could be found at here.

With the advances in artificial intelligence (AI) and precision medicine, radiomics has emerged as a promising tool in the field of oncology. Radiogenomics integrates radiomics with genomic data, potentially offering a non-invasive method for identifying biomarkers relevant to cancer therapy. Liquid biopsy (LB) has further revolutionized cancer diagnostics by detecting circulating tumor DNA (ctDNA), enabling real-time molecular profiling. This study explores the integration of radiomics and LB to predict genomic alterations in solid tumors, including lung, colon, pancreatic, and prostate cancers. A retrospective study was conducted on 418 patients from the STING trial (NCT04932525), all of whom underwent both LB and CT imaging. Predictive models were developed using an XGBoost logistic classifier, with statistical analysis performed to compare tumor volumes, lesion counts, and affected organs across molecular subtypes. Performance was evaluated using area under the curve (AUC) values and cross-validation techniques. Radiomic models demonstrated moderate-to-good performance in predicting genomic alterations. KRAS mutations were best identified in pancreatic cancer (AUC=0.97), while moderate discrimination was noted in lung (AUC=0.66) and colon cancer (AUC=0.64). EGFR mutations in lung cancer were detected with an AUC of 0.74, while BRAF mutations showed good discriminatory ability in both lung (AUC=0.79) and colon cancer (AUC=0.76). In the radiomics predictive model, AR mutations in prostate cancer showed limited discrimination (AUC = 0.63). This study highlights the feasibility of integrating radiomics and LB for non-invasive genomic profiling in solid tumors, demonstrating significant potential in patient stratification and personalized oncology care. While promising, further prospective validation is required to enhance the generalizability of these models.

Unsupervised domain adaptation (UDA) methods have achieved significant progress in medical image segmentation. Nevertheless, the significant differences between the source and target domains remain a daunting barrier, creating an urgent need for more robust cross-domain solutions. Current UDA techniques generally employ a fixed, unvarying feature alignment procedure to reduce inter-domain differences throughout the training process. This rigidity disregards the shifting nature of feature distributions throughout the training process, leading to suboptimal performance in boundary delineation and detail retention on the target domain. A novel confidence-guided unsupervised domain adaptation network (CUDA-Net) is introduced to overcome persistent domain gaps, adapt to shifting feature distributions during training, and enhance boundary delineation in the target domain. This proposed network adaptively aligns features by tracking cross-domain distribution shifts throughout training, starting with adversarial alignment at early stages (coarse) and transitioning to pseudo-label-driven alignment at later stages (fine-grained), thereby leading to more accurate segmentation in the target domain. A confidence-weighted mechanism then refines these pseudo labels by prioritizing high-confidence regions while allowing low-confidence areas to be gradually explored, thereby enhancing both label reliability and overall model stability. Experiments on three representative medical image datasets, namely MMWHS17, BraTS2021, and VS-Seg, confirm the superiority of CUDA-Net. Notably, CUDA-Net outperforms eight leading methods in terms of overall segmentation accuracy (Dice) and boundary extraction precision (ASD), highlighting that it offers an efficient and reliable solution for cross-domain medical image segmentation.

The generalization of deep learning-based low-dose computed tomography (CT) reconstruction models to doses unseen in the training data is important and remains challenging. Previous efforts heavily rely on paired data to improve the generalization performance and robustness through collecting either diverse CT data for re-training or a few test data for fine-tuning. Recently, diffusion models have shown promising and generalizable performance in low-dose CT (LDCT) reconstruction, however, they may produce unrealistic structures due to the CT image noise deviating from Gaussian distribution and imprecise prior information from the guidance of noisy LDCT images. In this paper, we propose a noise-inspired diffusion model for generalizable LDCT reconstruction, termed NEED, which tailors diffusion models for noise characteristics of each domain. First, we propose a novel shifted Poisson diffusion model to denoise projection data, which aligns the diffusion process with the noise model in pre-log LDCT projections. Second, we devise a doubly guided diffusion model to refine reconstructed images, which leverages LDCT images and initial reconstructions to more accurately locate prior information and enhance reconstruction fidelity. By cascading these two diffusion models for dual-domain reconstruction, our NEED requires only normal-dose data for training and can be effectively extended to various unseen dose levels during testing via a time step matching strategy. Extensive qualitative, quantitative, and segmentation-based evaluations on two datasets demonstrate that our NEED consistently outperforms state-of-the-art methods in reconstruction and generalization performance. Source code is made available at https://github.com/qgao21/NEED.

Generative AI models have demonstrated strong potential in radiology report generation, but their clinical adoption depends on physician trust. In this study, we conducted a radiology-focused Turing test to evaluate how well attendings and residents distinguish AI-generated reports from those written by radiologists, and how their confidence and decision time reflect trust. we developed an integrated web-based platform comprising two core modules: Report Generation and Report Evaluation. Using the web-based platform, eight participants evaluated 48 anonymized X-ray cases, each paired with two reports from three comparison groups: radiologist vs. AI model 1, radiologist vs. AI model 2, and AI model 1 vs. AI model 2. Participants selected the AI-generated report, rated their confidence, and indicated report preference. Attendings outperformed residents in identifying AI-generated reports (49.9% vs. 41.1%) and exhibited longer decision times, suggesting more deliberate judgment. Both groups took more time when both reports were AI-generated. Our findings highlight the role of clinical experience in AI acceptance and the need for design strategies that foster trust in clinical applications. The project page of the evaluation platform is available at: https://zachatalay89.github.io/Labsite.