Imaging studies for predicting hematoma expansion: from traditional imaging signs to artificial intelligence-based multimodal fusion

Hematoma expansion (HE) is a critical and modifiable event following acute intracerebral hemorrhage (ICH). Predicting HE accurately can inform individualized treatment and improve patient outcomes. This review systematically outlines the evolution of imaging-based HE prediction. We first define the core concepts of traditional HE, revised HE (rHE), and ultra-early hematoma growth (uHG). We then summarize predictive studies that employ traditional imaging markers, such as the computed tomography angiography (CTA) spot sign, non-contrast CT (NCCT) signs, and combined clinical-imaging scoring systems. Subsequent sections focus on AI-driven methodologies, encompassing radiomics, deep learning, and multi-task learning. The discussion extends to precision prediction through multimodal data fusion and subgroup analyses based on hemorrhage location and onset time. Finally, we address persistent challenges, including model interpretability, generalizability, and translational gaps, and suggest future directions involving federated learning, explainable AI, dynamic prediction, and closed-loop decision systems. This review offers a structured framework to guide both clinical practice and future research.