Annotating medical data for training AI models is often costly and limited due to the shortage of specialists with relevant clinical expertise. This challenge is further compounded by privacy and ethical concerns associated with sensitive patient information. As a result, well-trained medical segmentation models on private datasets constitute valuable intellectual property requiring robust protection mechanisms. Existing model protection techniques primarily focus on classification and generative tasks, while segmentation models-crucial to medical image analysis-remain largely underexplored. In this paper, we propose a novel, stealthy, and harmless method, StealthMark, for verifying the ownership of medical segmentation models under closed-box conditions. Our approach subtly modulates model uncertainty without altering the final segmentation outputs, thereby preserving the model's performance. To enable ownership verification, we incorporate model-agnostic explanation methods, e.g. LIME, to extract feature attributions from the model outputs. Under specific triggering conditions, these explanations reveal a distinct and verifiable watermark. We further design the watermark as a QR code to facilitate robust and recognizable ownership claims. We conducted extensive experiments across four medical imaging datasets (CMR dataset from UK Biobank, the SEG fundus dataset, the EchoNet echocardiography dataset, and the PraNet colonoscopy dataset) and five mainstream segmentation models. The results demonstrate the effectiveness, stealthiness, and harmlessness of our method on the original model's segmentation performance. For example, when applied to the SAM model, StealthMark consistently achieved attack success rates (ASR) above 95% across various datasets while maintaining less than a 1% drop in Dice and AUC scores-significantly outperforming backdoor-based watermarking methods and highlighting its strong potential for practical deployment. Our implementation code is made available at https://github.com/Qinkaiyu/StealthMark.

Atrial fibrillation (AF), impacting nearly 50 million individuals globally, is a major contributor to ischaemic strokes, predominantly originating from the left atrial appendage (LAA). Current clinical scores like CHA₂DS₂-VASc, while useful, provide limited insight into the pro-thrombotic mechanisms of Virchow's triad-blood stasis, endothelial damage, and hypercoagulability. This study leverages biophysical computational modelling to deepen our understanding of thrombogenesis in AF patients. Utilising high temporal resolution Cine magnetic resonance imaging (MRI), a 3D patient-specific modelling pipeline for simulating patient-specific flow in the left atrium was developed. This computational fluid dynamics (CFD) approach was coupled with reaction-diffusion-convection equations for key clotting proteins, leading to an innovative risk stratification score that combines clinical and modelling data. This approach categorises thrombogenic risk into low (A), moderate (B), and high (C) levels. Applied to a cohort of nine patients, pre- and post-catheter ablation therapy, this approach generates novel risk scores of thrombus formation, which are based of mechanistic characterisation of all aspects of the Virchow's triad. Currently, thrombogenesis mechanisms are not factored in widespread clinical risks scores based on demographic characteristics and co-morbidities. Notably, some patients with a CHA₂DS₂-VASc score of 0 (lowest clinical risk) exhibited much higher risks once the individual pathophysiology was accounted for. This discrepancy highlights the limitations of the CHA₂DS₂-VASc score in providing detailed mechanistic insights into patient-specific thrombogenic risk. This work introduces a comprehensive method for assessing thrombus formation risks in AF patients, emphasising the value of integrating biophysical modelling with clinical scores to enhance personalised stroke prevention strategies.

Cardiovascular diseases remain one of the leading causes of death worldwide. Therefore, developing and validating automated tools to help identify high-risk patients are of paramount clinical utility. In this article, we tackle this task and introduce a genetic programming algorithm (called GIRAFFE) to build (deep) machine learning classification ensembles for arrhythmia classification from two-dimensional images of 12-lead electrocardiogram (ECG) tracings. GIRAFFE evolves the architecture, content, and fusion scheme of the ensemble, to obtain an accurate yet lightweight classification system. The experimental study performed over a large-scale dataset of ECG images revealed that our approach outperforms other ensemble methods and carefully fine-tuned deep models, elaborates compact heterogeneous ensembles, and does not require any user intervention hence it is easy to apply to other classification tasks.

Diagnosis assignment is the process of assigning disease codes to patients. Automatic diagnosis assignment has the potential to validate code assignments, correct erroneous codes, and register completion. Previous methods build on text-based techniques utilizing medical notes but are inapplicable in the absence of these notes. We propose using patients' medication data to assign diagnosis codes. We present a proof-of-concept study using medical data from an American dataset (MIMIC-III) and Danish nationwide registers to train a machine-learning-based model that predicts an extensive collection of diagnosis codes for multiple levels of aggregation over a disease hierarchy. We further suggest a specialized loss function designed to utilize the innate hierarchical nature of the disease hierarchy. We evaluate the proposed method on a subset of 567 disease codes. Moreover, we investigate the technique's generalizability and transferability by (1) training and testing models on the same subsets of disease codes over the two medical datasets and (2) training models on the American dataset while evaluating them on the Danish dataset, respectively. Results demonstrate the proposed method can correctly assign diagnosis codes on multiple levels of aggregation from the disease hierarchy over the American dataset with recall 70.0% and precision 69.48% for top-10 assigned codes; thereby being comparable to text-based techniques. Furthermore, the specialized loss function performs consistently better than the non-hierarchical state-of-the-art version. Moreover, results suggest the proposed method is language and dataset-agnostic, with initial indications of transferability over subsets of disease codes.