Semi-supervised Breast Lesion Segmentation using Local Cross Triplet Loss for Ultrafast Dynamic Contrast-Enhanced MRI

YoungTack Oh, Eun Sook Ko, Hyunjin Park; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 2713-2728

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and its fast variant, ultrafast DCE-MRI, are useful for the management of breast cancer. Segmentation of breast lesions is necessary for automatic clinical decision support. Despite the advantage of acquisition time, existing segmentation studies on ultrafast DCE-MRI are scarce, and they are mostly fully supervised studies with high annotation costs. Herein, we propose a semi-supervised segmentation approach that can be trained with small amounts of annotations for ultrafast DCE-MRI. A time difference map is proposed to incorporate the distinct time-varying enhancement pattern of the lesion. Furthermore, we present a novel loss function that efficiently distinguishes breast lesions from non-lesions based on triple loss. This loss reduces the potential false positives induced by the time difference map. Our approach is compared to that of five competing methods using the dice similarity coefficient and two boundary-based metrics. Compared to other models, our approach achieves better segmentation results using small amounts of annotations, especially for boundary-based metrics relevant to spatially continuous breast lesions. An ablation study demonstrates the incremental effects of our study. Our code is available on GitHub (https://github.com/yt- oh96/SSL-CTL).