Misclassification Loss for Segmentation of the Aortic Vessel Tree

Abbas Khan, Muhammad Asad, Alexander M. Zolotarev, Caroline H. Roney, Anthony Mathur, Martin Benning, Gregory G. Slabaugh

Abstract: Common pixel-based loss functions for image segmentation struggle with the fine-scale structures often found in the aortic vessel tree. In this paper, we propose a Misclassification Loss (MC loss) function, which can effectively suppress false positives and rescue the false negatives. A differentiable eXclusive OR (XOR) operation is implemented to identify these false predictions, which are then minimized through a cross-entropy loss. The proposed MC loss helps the network achieve better performance by focusing on these difficult regions. On the Segmentation of the Aorta SEG.A. 2023 challenge, our method achieves a Dice score of 0.93 and a Hausdorff Distance (HD) of 3.50 mm on a 5-fold split of 56 training subjects. We participated in the SEG.A. 2023 challenge, and the proposed method ranks among the top-six approaches in the validation phase-1. The pre-trained models, source code, and implementation will be made public.