ICH-HPINet: a hybrid propagation interaction network for intelligent and interactive 3D intracerebral hemorrhage segmentation

Intracerebral hemorrhage (ICH) is the deadliest form of stroke and is associated with high disability rates. Accurate segmentation and quantitative analysis are critical for effective patient management, yet current 3D ICH segmentation methods often require extensive manual annotations and 2D methods fail to capture inter-slice relationships. Currently, prompt-based ICH segmentation methods have only one-time interaction and lack feedback functions. We propose ICH-HPINet, a novel hybrid propagation interaction network for intelligent and interactive segmentation of ICH regions in 3D images to address these challenges. The model reduces annotation needs while maintaining or improving segmentation performance. ICH-HPINet consists of four key components: the Volume Interaction Module, the Slice Interaction Module, the Feature Convert Module, and the Multi-Propagation Feature Fusion Module, enabling hybrid propagation and intelligent interaction. We validated ICH-HPINet on both a private dataset and the Physionet dataset, demonstrating superior performance compared to existing state-of-the-art methods with fewer prompts.