To improve the accuracy of hematoma delineation in non-contrast CT imaging for intracerebral hemorrhage (ICH), which is crucial for clinical decision-making, using a contour-centric framework.
Key Findings:
HemaContour achieved a Dice score of 87.2%, surpassing the best baseline (Swin–UNETR) at 85.0%, indicating a significant advancement in segmentation accuracy.
Reduced Hausdorff distance (HD95) from 8.5 mm to 7.3 mm (~14.1% improvement), reflecting better boundary adherence.
On external validation (PhysioNet CT–ICH), maintained Dice score of 84.3% vs. 81.8% and HD95 of 8.5 mm vs. 9.9 mm, demonstrating robustness across datasets.
Better volumetric agreement with average volume error (AVE) of 4.3 mL vs. 5.0 mL, highlighting improved clinical utility.
Interpretation:
HemaContour enhances boundary fidelity and volumetric accuracy in ICH segmentation, offering a robust alternative to voxel-centric methods, which is critical for clinical applications.
Limitations:
The study may not address all variations in hematoma shapes and sizes, particularly in complex cases.
Performance may vary with different imaging artifacts, such as motion blur or noise, which were not tested.
Conclusion:
HemaContour provides a clinically interpretable and efficient method for ICH segmentation, bridging manual tracing and automated deep learning, with potential applications in real-time clinical settings.
