Segmentation of Tiny Intracranial Hemorrhage Via Learning-to-Rank Local Feature Enhancement

Intracranial hemorrhage (ICH) is a common head disease that can result in significant disability or mortality. Segmentation of ICH is an important yet challenging step for medical image interpretation. While numerous methods have been proposed for automatic ICH segmentation, they still suffer from poor performance when dealing with small-sized hemorrhages. In this paper, we propose an effective framework to improve the segmentation accuracy of cases involving tiny hemorrhages. Our approach integrates a designed learning-to-rank module within the segmentation network, which can enhance the discrimination capability of local features. This enhancement assists the network in distinguishing patches w/ and w/o hemorrhages. Furthermore, we incorporate a hard negative mining strategy during the pairing process, ensuring the network focuses on the most ambiguous and challenging cases. We have conducted experiments on a labelled in-house CT dataset consisting of 267 head CT scans. The results demonstrate that our method can achieve state-of-the-art performance, particularly for tiny hemorrhages with a volume size smaller than 1 ml. Our method has outperformed nnU-Net by a large margin¹.