Multi-Site Benchmarking of Deep Learning Models for Intraparenchymal Hemorrhage Segmentation on NCCT

Abstract

Intraparenchymal hemorrhage (IPH) is a critical and often fatal subtype of hemorrhagic stroke, requiring rapid and accurate diagnosis on non-contrast computed tomography (NCCT) scans for effective treatment. While deep learning (DL) models, particularly convolutional neural networks (CNNs), offer potential for automating IPH segmentation, their real-world clinical utility is often limited by the lack of explicit data integration across diverse hospital sites with varying imaging protocols. This study conducted a multi-site benchmarking of black{five} prominent CNN architectures: baseline U-Net, Attention U-Net, Feature Pyramid Network (FPN), black{Swin U-Net}, and Trans U-Net, for IPH segmentation on a heterogeneous dataset from 17 clinical sites. Models were rigorously evaluated using F-measure (a.k.a., Dice), Intersection over Union (IoU), and 95% Hausdorff Distance ($d_{H95}$). The advanced CNN variants (Attention U-Net, FPN, Trans U-Net) significantly outperformed the baseline U-Net in F-measure and IoU (e.g., FPN F-measure: $0.868$ vs. U-Net: $0.819$, $p<0.001$), with no significant difference among them. For boundary error, FPN reduced $d_{H95}$ compared to the baseline, whereas Trans U-Net showed improvement, though it was not significant. These models exhibited robust cross-site generalization across hemorrhage volumes, with minimal site-specific effects on performance. This study demonstrates that advanced CNN variants can be adopted for IPH segmentation to standardize and potentially accelerate IPH diagnosis.