Osteoporosis, characterized by reduced bone mineral density (BMD) and compromised bone microstructure, increases fracture risk in aging populations. While dual-energy X-ray absorptiometry (DXA) is the clinical standard for BMD assessment, its limited accessibility hinders diagnosis in resource-limited regions. Opportunistic computed tomography (CT) analysis has emerged as a promising alternative for osteoporosis diagnosis using existing imaging data. Current approaches, however, face three limitations: (1) underutilization of unlabeled vertebral data, (2) systematic bias from device-specific DXA discrepancies, and (3) insufficient integration of clinical knowledge such as spatial BMD distribution patterns. To address these, we propose a unified deep learning framework with three innovations. First, a self-supervised learning method using radiomic representations to leverage unlabeled CT data and preserve bone texture. Second, a Mixture of Experts (MoE) architecture with learned gating mechanisms to enhance cross-device adaptability. Third, a multi-task learning framework integrating osteoporosis diagnosis, BMD regression, and vertebra location prediction. Validated across three clinical sites and an external hospital, our approach demonstrates superior generalizability and accuracy over existing methods for opportunistic osteoporosis screening and diagnosis.

Accurate interpretation of knee MRI scans relies on expert clinical judgment, often with high variability and limited scalability. Existing radiomic approaches use a fixed set of radiomic features (the signature), selected at the population level and applied uniformly to all patients. While interpretable, these signatures are often too constrained to represent individual pathological variations. As a result, conventional radiomic-based approaches are found to be limited in performance, compared with recent end-to-end deep learning (DL) alternatives without using interpretable radiomic features. We argue that the individual-agnostic nature in current radiomic selection is not central to its intepretability, but is responsible for the poor generalization in our application. Here, we propose a novel radiomic fingerprint framework, in which a radiomic feature set (the fingerprint) is dynamically constructed for each patient, selected by a DL model. Unlike the existing radiomic signatures, our fingerprints are derived on a per-patient basis by predicting the feature relevance in a large radiomic feature pool, and selecting only those that are predictive of clinical conditions for individual patients. The radiomic-selecting model is trained simultaneously with a low-dimensional (considered relatively explainable) logistic regression for downstream classification. We validate our methods across multiple diagnostic tasks including general knee abnormalities, anterior cruciate ligament (ACL) tears, and meniscus tears, demonstrating comparable or superior diagnostic accuracy relative to state-of-the-art end-to-end DL models. More importantly, we show that the interpretability inherent in our approach facilitates meaningful clinical insights and potential biomarker discovery, with detailed discussion, quantitative and qualitative analysis of real-world clinical cases to evidence these advantages.

Positron Emission Tomography / Computed Tomography (PET/CT) plays a critical role in medical imaging, combining functional and anatomical information to aid in accurate diagnosis. However, image quality degradation due to noise, compression and other factors could potentially lead to diagnostic uncertainty and increase the risk of misdiagnosis. When evaluating the quality of a PET/CT image, both low-level features like distortions and high-level features like organ anatomical structures affect the diagnostic value of the image. However, existing medical image quality assessment (IQA) methods are unable to account for both feature types simultaneously. In this work, we propose MS-IQA, a novel multi-scale feature fusion network for PET/CT IQA, which utilizes multi-scale features from various intermediate layers of ResNet and Swin Transformer, enhancing its ability of perceiving both local and global information. In addition, a multi-scale feature fusion module is also introduced to effectively combine high-level and low-level information through a dynamically weighted channel attention mechanism. Finally, to fill the blank of PET/CT IQA dataset, we construct PET-CT-IQA-DS, a dataset containing 2,700 varying-quality PET/CT images with quality scores assigned by radiologists. Experiments on our dataset and the publicly available LDCTIQAC2023 dataset demonstrate that our proposed model has achieved superior performance against existing state-of-the-art methods in various IQA metrics. This work provides an accurate and efficient IQA method for PET/CT. Our code and dataset are available at https://github.com/MS-IQA/MS-IQA/.

To address the high computational complexity of the Transformer in the segmentation of ultrasound thyroid nodules and the loss of image details or omission of key spatial information caused by traditional image sampling techniques when dealing with high-resolution, complex texture or uneven density two-dimensional ultrasound images, this paper proposes a thyroid nodule segmentation method that integrates the receiving weighted key-value (RWKV) architecture and spherical geometry feature (SGF) sampling technology. This method effectively captures the details of adjacent regions through two-dimensional offset prediction and pixel-level sampling position adjustment, achieving precise segmentation. Additionally, this study introduces a patch attention module (PAM) to optimize the decoder feature map using a regional cross-attention mechanism, enabling it to focus more precisely on the high-resolution features of the encoder. Experiments on the thyroid nodule segmentation dataset (TN3K) and the digital database for thyroid images (DDTI) show that the proposed method achieves dice similarity coefficients (DSC) of 87.24% and 80.79% respectively, outperforming existing models while maintaining a lower computational complexity. This approach may provide an efficient solution for the precise segmentation of thyroid nodules.

Radiologist's manual annotations limit robust deep learning in volumetric medical imaging. While supervised methods excel with large annotated datasets, few-shot learning performs well for large structures but struggles with small ones, such as lesions. This paper describes a novel method that leverages the advantages of both few-shot learning models and fully supervised models while reducing the cost of manual annotation. Our method inputs a small dataset of labeled scans and a large dataset of unlabeled scans and outputs a validated labeled dataset used to train a supervised model (nnU-Net). The estimated correction effort is reduced by having the radiologist correct a subset of the scan labels computed by a few-shot learning model (UniverSeg). The method uses an optimized support set of scan slice patches and prioritizes the resulting labeled scans that require the least correction. This process is repeated for the remaining unannotated scans until satisfactory performance is obtained. We validated our method on liver, lung, and brain lesions on CT and MRI scans (375 scans, 5933 lesions). It significantly reduces the estimated lesion detection correction effort by 34% for missed lesions, 387% for wrongly identified lesions, with 130% fewer lesion contour corrections, and 424% fewer pixels to correct in the lesion contours with respect to manual annotation from scratch. Our method effectively reduces the radiologist's annotation effort of small structures to produce sufficient high-quality annotated datasets to train deep learning models. The method is generic and can be applied to a variety of lesions in various organs imaged by different modalities.

Obesity is associated with functional alterations in the brain. Although spatial organisation changes in the brains of individuals with obesity have been widely studied, the temporal dynamics in their brains remain poorly understood. Therefore, in this study, we investigated variations in the intrinsic neural timescale (INT) across different degrees of obesity using resting-state functional and diffusion magnetic resonance imaging data from the enhanced Nathan Kline Institute Rockland Sample database. We examined the relationship between the INT and obesity phenotypes using supervised machine learning, controlling for age and sex. To further explore the structure-function characteristics of these regions, we assessed the modular network properties by analysing the participation coefficients and within-module degree derived from the structure-function coupling matrices. Finally, the INT values of the identified regions were used to predict eating behaviour traits. A significant negative correlation was observed, particularly in the default mode, limbic and reward networks. We found a negative association with the participation coefficients, suggesting that shorter INT values in higher-order association areas are related to reduced network integration. Moreover, the INT values of these identified regions moderately predicted eating behaviours, underscoring the potential of the INT as a candidate marker for obesity and eating behaviours. These findings provide insight into the temporal organisation of neural activity in obesity, highlighting the role of specific brain networks in shaping behavioural outcomes.

To assess whether the diagnostic performance of a commercial artificial intelligence (AI) algorithm for mammography differs between breast-level and lesion-level interpretations and to compare performance to a large population of specialised human readers. We retrospectively analysed 1200 mammograms from the NHS breast cancer screening programme using a commercial AI algorithm and assessments from 1258 trained human readers from the Personal Performance in Mammographic Screening (PERFORMS) external quality assurance programme. For breasts containing pathologically confirmed malignancies, a breast and lesion-level analysis was performed. The latter considered the locations of marked regions of interest for AI and humans. The highest score per lesion was recorded. For non-malignant breasts, a breast-level analysis recorded the highest score per breast. Area under the curve (AUC), sensitivity and specificity were calculated at the developer's recommended threshold for recall. The study was designed to detect a medium-sized effect (odds ratio 3.5 or 0.29) for sensitivity. The test set contained 882 non-malignant (73%) and 318 malignant breasts (27%), with 328 cancer lesions. The AI AUC was 0.942 at breast level and 0.929 at lesion level (difference -0.013, p < 0.01). The mean human AUC was 0.878 at breast level and 0.851 at lesion level (difference -0.027, p < 0.01). AI outperformed human readers at the breast and lesion level (ps < 0.01, respectively) according to the AUC. AI's diagnostic performance significantly decreased at the lesion level, indicating reduced accuracy in localising malignancies. However, its overall performance exceeded that of human readers. Question AI often recalls mammography cases not recalled by humans; to understand why, we as humans must consider the regions of interest it has marked as cancerous. Findings Evaluations of AI typically occur at the breast level, but performance decreases when AI is evaluated on a lesion level. This also occurs for humans. Clinical relevance To improve human-AI collaboration, AI should be assessed at the lesion level; poor accuracy here may lead to automation bias and unnecessary patient procedures.

Deep-learning models for prostate cancer detection typically require large datasets, limiting clinical applicability across institutions due to domain shift issues. This study aimed to develop a few-shot learning deep-learning model for prostate cancer detection on multiparametric MRI that requires minimal training data and to compare its diagnostic performance with experienced radiologists. In this retrospective study, we used 99 cases (80 positive, 19 negative) of biopsy-confirmed prostate cancer (2017-2022), with 20 cases for training, 5 for validation, and 74 for testing. A 2D transformer model was trained on T2-weighted, diffusion-weighted, and apparent diffusion coefficient map images. Model predictions were compared with two radiologists using Matthews correlation coefficient (MCC) and F1 score, with 95% confidence intervals (CIs) calculated via bootstrap method. The model achieved an MCC of 0.297 (95% CI: 0.095-0.474) and F1 score of 0.707 (95% CI: 0.598-0.847). Radiologist 1 had an MCC of 0.276 (95% CI: 0.054-0.484) and F1 score of 0.741; Radiologist 2 had an MCC of 0.504 (95% CI: 0.289-0.703) and F1 score of 0.871, showing that the model performance was comparable to Radiologist 1. External validation on the Prostate158 dataset revealed that ImageNet pretraining substantially improved model performance, increasing study-level ROC-AUC from 0.464 to 0.636 and study-level PR-AUC from 0.637 to 0.773 across all architectures. Our findings demonstrate that few-shot deep-learning models can achieve clinically relevant performance when using pretrained transformer architectures, offering a promising approach to address domain shift challenges across institutions.

The study aims to utilize interpretable machine learning models to predict the risk of myasthenia gravis onset in thymoma patients and investigate the intrinsic correlations and causal relationships between them. A comprehensive retrospective analysis was conducted on 172 thymoma patients diagnosed at two medical centers between 2018 and 2024. The cohort was bifurcated into a training set (n = 134) and test set (n = 38) to develop and validate risk predictive models. Radiomic and deep features were extracted from tumor regions across three CT phases: non-enhanced, arterial, and venous. Through rigorous feature selection employing Spearman's rank correlation coefficient and LASSO (Least Absolute Shrinkage and Selection Operator) regularization, 12 optimal imaging features were identified. These were integrated with 11 clinical parameters and one pathological subtype variable to form a multi-dimensional feature matrix. Six machine learning algorithms were subsequently implemented for model construction and comparative analysis. We utilized SHAP (SHapley Additive exPlanation) to interpret the model and employed doubly robust learner to perform a potential causal analysis between thymoma and myasthenia gravis (MG). All six models demonstrated satisfactory predictive capabilities, with the support vector machine (SVM) model exhibiting superior performance on the test cohort. It achieved an area under the curve (AUC) of 0.904 (95% confidence interval [CI] 0.798-1.000), outperforming other models such as logistic regression, multilayer perceptron (MLP), and others. The model's predictive result substantiates the strong correlation between thymoma and MG. Additionally, our analysis revealed the existence of a significant causal relationship between them, and high-risk tumors significantly elevated the risk of MG by an average treatment effect (ATE) of 9.2%. This implies that thymoma patients with types B2 and B3 face a considerably high risk of developing MG compared to those with types A, AB, and B1. The model provides a novel and effective tool for evaluating the risk of MG development in patients with thymoma. Furthermore, correlation and causal analysis have unveiled pathways that connect tumor to the risk of MG, with a notably higher incidence of MG observed in high risk pathological subtypes. These insights contribute to a deeper understanding of MG and drive a paradigm shift in medical practice from passive treatment to proactive intervention.

Aortic valve calcium scoring is an essential tool for diagnosing, treating, monitoring, and assessing the risk of aortic stenosis. The current gold standard for determining the aortic valve calcium score is computed tomography (CT). However, CT is costly and exposes patients to ionizing radiation, making it less ideal for frequent monitoring. Echocardiography, a safer and more affordable alternative that avoids radiation, is more widely accessible, but its variability between and within experts leads to subjective interpretations. Given these limitations, there is a clear need for an automated, objective method to measure the aortic valve calcium score from echocardiography, which could reduce costs and improve patient safety. In this paper, we first employ the YOLOv5 method to detect the region of interest in the aorta within echocardiography images. Building on this, we propose a novel approach that combines UNet and diffusion model architectures to segment calcified areas within the identified region, forming the foundation for automated aortic valve calcium scoring. This architecture leverages UNet's localization capabilities and the diffusion model's strengths in capturing fine-grained structures, enhancing both segmentation accuracy and consistency. The proposed method achieves 85.08% precision, 80.01% recall, and 71.13% Dice score on a novel dataset comprising 160 echocardiography images from 86 distinct patients. This system enables cardiologists to focus more on critical aspects of diagnosis by providing a faster, more objective, and cost-effective method for aortic valve calcium scoring and eliminating the risk of radiation exposure.