HemaContour: explicit parametric contour learning for robust ICH segmentation on non-contrast CT - Takeaways - MDSpire

HemaContour: explicit parametric contour learning for robust ICH segmentation on non-contrast CT

  • By

  • Cheng Zheng

  • Guomin Xie

  • Hongcai Wang

  • Bingxuan Ren

  • Xinru Lin

  • Jincheng Jiang

  • Xinchen Jiang

  • Zhixiang Zhang

  • Haifeng Wang

  • Wu Zheng

  • December 10, 2025

  • 0 min

Share

  • 1

    HemaContour is a contour-centric framework that enhances ICH segmentation in non-contrast CT imaging by fitting a closed parametric spline to the hematoma boundary.

  • 2

    The method utilizes a coarse CNN for contour seeding, followed by optimization through an implicit contour-regression network with a shape-aware objective.

  • 3

    HemaContour achieves superior performance with a Dice score of 87.2% and reduces Hausdorff distance by approximately 14.1% compared to the best baseline.

  • 4

    The framework maintains high accuracy in external validation, with Dice scores of 84.3% and improved volumetric agreement over traditional methods.

  • 5

    By focusing on explicit contour modeling, HemaContour improves boundary fidelity and volumetric accuracy, offering a robust alternative to voxel-centric segmentation.

Original Source(s)

HemaContour: explicit parametric contour learning for robust ICH segmentation on non-contrast CT

  • by Cheng Zheng, Guomin Xie, Hongcai Wang, Bingxuan Ren, Xinru Lin, Jincheng Jiang, Xinchen Jiang, Zhixiang Zhang, Haifeng Wang, Wu Zheng

  • December 10, 2025

Related Content

STD-Net: a spatio-temporal decoupling network for multiphasic liver lesion segmentation and characterization

PAM: a propagation-based model for segmenting any 3D objects across multi-modal medical images

Deep learning model for assessing survival benefits in hepatocellular carcinoma patients undergoing intra-arterial therapies based on proliferative subtype