A multimodal prediction model for brain hematoma expansion based on deep neural network with transformer-driven collaborative attention fusion.

Abstract

Hematoma expansion (HE) in patients with spontaneous intracerebral hemorrhage (ICH) is strongly associated with death and disability, so accurate early prediction is important for clinical decision making. Existing methods often rely on single imaging signs or traditional scoring systems and are unable to consider both the three dimensional hematoma structure and the clinical status at the same time, which limits predictive performance and robustness. To address this, we propose a multimodal deep learning model named SResFormer based on position-aware Computed Tomography (CT) feature encoding. CT images and parenchymal hematoma masks are processed in a dual channel scheme. Spatial order across slices is then modeled using position-aware CT feature encoding, enabling explicit capture of three dimensional patterns of hematoma evolution. These representations are subsequently fused with structured clinical features through semantic alignment to achieve integrated cross modality prediction. SResFormer demonstrates significantly superior performance compared with traditional scoring systems, unimodal Convolutional Neural Network (CNN) based methods, and shallow multimodal fusion approaches on multicenter datasets. Ablation studies confirm both the independent contributions and the synergistic benefits of each module. SHapley Additive exPlanations (SHAP) analysis further indicates that the model reasoning is consistent with established medical knowledge.