Deep learning models achieve strong performance in chest radiograph (CXR) interpretation, yet fairness and reliability concerns persist. Models often show uneven accuracy across patient subgroups, leading to hidden failures not reflected in aggregate metrics. Existing error detection approaches -- based on confidence calibration or out-of-distribution (OOD) detection -- struggle with subtle within-distribution errors, while image- and representation-level consistency-based methods remain underexplored in medical imaging. We propose an augmentation-sensitivity risk scoring (ASRS) framework to identify error-prone CXR cases. ASRS applies clinically plausible rotations ($\pm 15^\circ$/$\pm 30^\circ$) and measures embedding shifts with the RAD-DINO encoder. Sensitivity scores stratify samples into stability quartiles, where highly sensitive cases show substantially lower recall ($-0.2$ to $-0.3$) despite high AUROC and confidence. ASRS provides a label-free means for selective prediction and clinician review, improving fairness and safety in medical AI.

Large Language Models (LLMs), such as GPT-4o, are increasingly investigated for clinical decision support in emergency medicine. However, their real-world performance in disposition prediction remains insufficiently studied. This study evaluated the diagnostic accuracy of GPT-4o in predicting ED disposition-discharge, ward admission, or ICU admission-in complex emergency respiratory cases requiring pulmonology consultation and chest CT, representing a selective high-acuity subgroup of ED patients. We conducted a retrospective observational study in a tertiary ED between November 2024 and February 2025. We retrospectively included ED patients with complex respiratory presentations who underwent pulmonology consultation and chest CT, representing a selective high-acuity subgroup rather than the general ED respiratory population. GPT-4o was prompted to predict the most appropriate ED disposition using three progressively enriched input models: Model 1 (age, sex, oxygen saturation, home oxygen therapy, and venous blood gas parameters); Model 2 (Model 1 plus laboratory data); and Model 3 (Model 2 plus chest CT findings). Model performance was assessed using accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. Among the 221 patients included, 69.2% were admitted to the ward, 9.0% to the intensive care unit (ICU), and 21.7% were discharged. For hospital admission prediction, Model 3 demonstrated the highest sensitivity (91.9%) and overall accuracy (76.5%), but the lowest specificity (20.8%). In contrast, for discharge prediction, Model 3 achieved the highest specificity (91.9%) but the lowest sensitivity (20.8%). Numerical improvements were observed across models, but none reached statistical significance (all p > 0.22). Model 1 therefore performed comparably to Models 2-3 while being less complex. Among patients who were discharged despite GPT-4o predicting admission, the 14-day ED re-presentation rates were 23.8% (5/21) for Model 1, 30.0% (9/30) for Model 2, and 28.9% (11/38) for Model 3. GPT-4o demonstrated high sensitivity in identifying ED patients requiring hospital admission, particularly those needing intensive care, when provided with progressively enriched clinical input. However, its low sensitivity for discharge prediction resulted in frequent overtriage, limiting its utility for autonomous decision-making. This proof-of-concept study demonstrates GPT-4o's capacity to stratify disposition decisions in complex respiratory cases under varying levels of limited input data. However, these findings should be interpreted in light of key limitations, including the selective high-acuity cohort and the absence of vital signs, and require prospective validation before clinical implementation.

Exploring the dynamics of a complex system, such as the human brain, poses significant challenges due to inherent uncertainties and limited data. In this study, we enhance the capabilities of noisy linear recurrent neural networks (lRNNs) within the reservoir computing framework, demonstrating their effectiveness in creating autonomous in silico replicas - digital twins - of brain activity. Our findings reveal that the poles of the Laplace transform of high-dimensional inferred lRNNs are directly linked to the spectral properties of observed systems and to the kernels of auto-regressive models. Applying this theoretical framework to resting-state fMRI, we successfully predict and decompose BOLD signals into spatiotemporal modes of a low-dimensional latent state space confined around a single equilibrium point. lRNNs provide an interpretable proxy for clustering among subjects and different brain areas. This adaptable digital-twin framework not only enables virtual experiments but also offers computational efficiency for real-time learning, highlighting its potential for personalized medicine and intervention strategies.

Cognitive decline in older adults, particularly during the preclinical stages of Alzheimer's disease (AD), presents a critical opportunity for early detection and intervention. While T1-weighted MRI is widely used in AD research, its capacity to identify early vulnerability and monitor longitudinal progression remains incompletely characterized. We analyzed longitudinal T1-weighted MRI data from 224 cognitively unimpaired older adults followed for up to 12 years. Participants were stratified by clinical outcome into converters to mild cognitive impairment (HC-converters, n = 112) and stable controls (HC-stable, n = 112). Groups were matched at baseline for age (mean ~ 74-75 years), education (~ 16.4 years), and cognitive scores (MMSE ≈ 29; CDR-SB ≈ 0.04). Four MRI-derived biomarkers were examined: brain-predicted age difference (brain-PAD), mean cortical thickness, AD-cortical signature, and hippocampal volume. Brain-PAD showed the strongest baseline association with future conversion (β = 1.25, t = 3.52, p = 0.0009) and highest classification accuracy (AUC = 0.66; sensitivity = 62%, and specificity = 67%). Longitudinal mixed-effects models focusing on the group × time interaction revealed a significant positive slope in brain-PAD for converters (β = 0.0079, p = 0.003) and a non-significant trend in stable controls (β = 0.0047, p = 0.075), indicating incipient divergence in brain aging trajectories during the preclinical window. Hippocampal volume and AD-cortical signature declined similarly in both groups. The mean cortical thickness had limited discriminative or dynamic utility. These findings support brain-PAD, derived from routine T1-weighted MRI using machine learning, as a sensitive, performance-independent biomarker for early risk stratification and monitoring of cognitive aging trajectories.

The objective is to develop a differential diagnosis model for tuberculous spondylitis (TS) and pyogenic spondylitis (PS) by integrating MRI morphological features and computed tomography (CT) density parameters (Hounsfield Units, HU). This study aims to leverage multimodal data complementarity to achieve fusion of qualitative and quantitative information, thereby providing clinicians with a rapid and objective decision support tool for spinal inflammatory lesion characterization. Imaging data were extracted from MRI and CT scans of patients with TS and PS, then compared and summarized. Receiver operating characteristic (ROC) curves were used to determine optimal HU value thresholds. The least absolute shrinkage and selection operator (Lasso) regression was applied to identify the most predictive features for model construction. A logistic regression-based predictive model was developed and visualized as a nomogram. Model validation was performed using bootstrap resampling, ROC analysis, and decision curve analysis (DCA). A total of 171 patients with TS (n = 91) or PS (n = 80) were included. Statistically significant differences in MRI features were observed between the two groups (P < 0.05). Additionally, significant HU value differences were found in diseased vertebral endplates, small cavitary abscesses, large cavitary abscesses, and intravertebral abscesses between TS and PS patients (P < 0.05). The predictive model incorporated seven independent predictors. Calibration curves, ROC analysis, and DCA all demonstrated excellent model performance. Combined MRI and CT HU value analysis effectively differentiates TS from PS. The predictive model integrating imaging features and quantitative parameters demonstrates high accuracy and clinical utility, offering a novel approach to optimize diagnostic and treatment strategies for spinal infectious diseases.

This study aimed to develop interpretable machine learning models using radiomic and dosiomic features from radiotherapy target volumes to predict treatment response in glioma patients. A retrospective analysis was conducted on 176 glioma patients. Treatment response was categorized into disease control rate (DCR) and non-DCR groups (training cohort: 71 vs. 44; validation cohort: 34 vs. 27). Five regions of interest (ROIs) were identified: gross tumor volume (GTV), gross tumor volume with tumor bed (GTVtb), clinical target volume (CTV), GTV-GTV and CTV-GTVtb. For each ROI, radiomic features and dosiomic features were separately extracted from CT images and dose maps. Feature selection was performed. Six dosimetric parameters and six clinical variables were also included in model development. Five predictive models were constructed using four machine learning algorithms: Radiomic, Dosiomic, Dose-Volume Histogram (DVH), Combined (integrating clinical, radiomic, dosiomic, and DVH features), and Clinical models. Model performance was evaluated using accuracy, precision, recall, F1-score, and area under the curve (AUC). SHAP analysis was applied to explain model predictions. The CTV_combined support vector machine (SVM) model achieved the best performance, with an AUC of 0.728 in the validation cohort. SHAP summary plots showed that dosiomic features contributed significantly to prediction. Force plots further illustrated how individual features affected classification outcomes. The SHAP-interpretable CTV_combined SVM model demonstrated strong predictive ability for treatment response in glioma patients. This approach may support radiation oncologists in identifying the underlying pathological mechanisms of poor treatment response and adjusting dose distribution accordingly, thereby aiding the development of personalized radiotherapy strategies. Not applicable.

Brain tumour segmentation is an important task in medical imaging, that requires accurate tumour localization for improved diagnostics and treatment planning. However, conventional segmentation models often struggle with boundary delineation and generalization across heterogeneous datasets. Furthermore, data privacy concerns limit centralized model training on large-scale, multi-institutional datasets. To address these drawbacks, we propose a Hybrid Dual Encoder-Decoder Segmentation Model in Federated Learning, that integrates EfficientNet with Swin Transformer as encoders and BASNet (Boundary-Aware Segmentation Network) decoder with MaskFormer as decoders. The proposed model aims to enhance segmentation accuracy and efficiency in terms of total training time. This model leverages hierarchical feature extraction, self-attention mechanisms, and boundary-aware segmentation for superior tumour delineation. The proposed model achieves a Dice Coefficient of 0.94, an Intersection over Union (IoU) of 0.87 and reduces total training time through faster convergence in fewer rounds. The proposed model exhibits strong boundary delineation performance, with a Hausdorff Distance (HD95) of 1.61, an Average Symmetric Surface Distance (ASSD) of 1.12, and a Boundary F1 Score (BF1) of 0.91, indicating precise segmentation contours. Evaluations on the Kaggle Mateuszbuda LGG-MRI segmentation dataset partitioned across multiple federated clients demonstrate consistent, high segmentation performance. These findings highlight that integrating transformers, lightweight CNNs, and advanced decoders within a federated setup supports enhanced segmentation accuracy while preserving medical data privacy.

Rapid and accurate quantitative assessment of muscle tissue characteristics is critical for the diagnosis and monitoring of neuromuscular diseases (NMDs). Quantitative magnetic resonance imaging enables non-invasive assessment of muscle pathology by using water T2 values to detect muscle damage and proton density fat fraction (PDFF) to quantify fat infiltration. However, conventional methods for simultaneous water-fat separation and T2 quantification often require long acquisition times. This study aims to develop an ultrafast method for simultaneous water-fat separation and T2 quantification.&#xD;Approach: A novel water-fat separation framework that combines chemical shift encoding with the multiple overlapping-echo detachment sequence (CSE-MOLED) was proposed. Synthetic training data and deep learning-based reconstruction were employed to address challenges in water-fat separation, including the complex multi-peak spectral characteristic of fat and the non-idealities in MRI acquisition. The proposed method was validated through numerical simulations, phantom studies, and in vivo experiments involving five healthy volunteers, one subject with muscle atrophy, and one with muscle damage.&#xD;Main results: In numerical experiments, the R2 values were all 0.999 for water T2, fat T2, and PDFF. In phantom experiments, the R2 values were 0.995, 0.733, and 0.996 for water T2, fat T2, and PDFF, respectively. High repeatability (coefficient of variation < 2.0%) was achieved in both phantom and in vivo experiments. In patient scans, CSE-MOLED successfully distinguished between fat infiltration and muscle damage.&#xD;Significance: CSE-MOLED simultaneously obtains T2 and proton density maps for both water and fat, along with T2-corrected PDFF map, in 162 ms per slice, offering the potential to enhance the diagnostic accuracy of NMDs without increasing the clinical scanning burden.

Accurate segmentation and anatomical classification of vertebrae in spinal CT scans are crucial for clinical diagnosis, surgical planning, and disease monitoring. However, the task is complicated by anatomical variability, degenerative changes, and the presence of rare vertebral anomalies. In this study, we propose a hybrid framework that combines a high-resolution WNet segmentation backbone with a Vision Transformer (ViT)-based classification module to perform vertebral identification and anomaly detection. Our model incorporates an attention-based anatomical variation module and leverages patient-specific metadata (age, sex, vertebral distribution) to improve the accuracy and personalization of vertebrae typing. Extensive experiments on the VerSe 2019 and 2020 datasets demonstrate that our approach outperforms state-of-the-art baselines such as nnUNet and SwinUNet, especially in detecting transitional vertebrae (e.g., T13, L6) and modeling morphological diversity. The system maintains high robustness under slice skipping, noise perturbation, and scanner variations, while offering interpretability through attention heatmaps and case-specific alerts. Our findings suggest that integrating anatomical priors and demographic context into transformer-based pipelines is a promising direction for personalized, intelligent spinal image analysis.

Metal implants and other high-density objects cause significant artifacts in computed tomography (CT) images, hindering clinical diagnosis. Traditional metal artifact reduction methods often leave residual artifacts due to sinogram edges discontinuities. Supervised deep learning approaches struggle due to reliance on paired data, while unsupervised methods often lack multi-domain information. In this paper, we propose TDMAR-Net, a diffusion model-based three-domain neural network that leverages priors from projection, image, and Fourier domains for removing metal artifact and enhancing CT image quality. To enhance the model's learning capability and gradient optimization while preventing reliance on a single data structure, we employ a two-stage training strategy that combines large-scale pretraining with masked data fine-tuning, improving both accuracy and adaptability in metal artifact removal. The specific process is to adjust the weight of the high frequency and low frequency components of the input image through the high-pass filter module in the Fourier domain, and process the image into blocks to extract the diffusion prior information. The prior information is then introduced iteratively into the sinogram and image domains to fill in the metal-induced artifacts. Our method overcomes the challenges of information sharing and complementarity across different domains, ensuring that each domain contributes effectively, thereby enhancing the precision and robustness of metal artifact elimination. Experiments show that our approach superior to existing unsupervised methods, which we have validated on both synthetic and clinical datasets.