DV-Net: Explainable model for hemorrhage prediction in AIS patients via physician logic

Document Type

Article

Publication Date

9-15-2026

Department

Department of Biomedical Engineering

Abstract

In the treatment of acute ischemic stroke, intravenous thrombolysis (IVT) carries the risk of hemorrhagic transformation (HT), necessitating accurate risk assessment for clinical decisions. However, existing deep learning methods often process the entire non-contrast CT (NCCT) slice sequence simultaneously, making it difficult to distinguish informative slices from redundant information and potentially reducing computational efficiency. This study proposes DV-Net, a multimodal framework combining NCCT images with clinical data. It introduces a “key slice focus” mechanism that mimics the diagnostic reasoning process of radiologists for precise HT prediction. The CNN-DVSS hybrid architecture implicitly emphasizes informative slices while suppressing redundant information within the slice sequence, thereby improving efficiency and accuracy. We evaluated DV-Net on 518 patients from multiple centers. The model achieves an average AUC value of 0.9523 in five-fold cross-validation, with AUC values of 0.9517 and 0.9041 for the overall training and test sets, outperforming mainstream architectures. The results of our qualitative, quantitative, and visualization-based interpretability analyses show good agreement with physician annotations, supporting model interpretability. Standardized preprocessing and BYOL-based self-supervised pre-training alleviate small sample variations across centers, improving generalization. DV-Net provides a decision support tool with both clinical feasibility and interpretability for HT risk assessment before IVT treatment in AIS patients, facilitating timely intervention and precise treatment.

Publication Title

Biomedical Signal Processing and Control

Link to Full Text

Share

COinS