A Deep Supervision-Based Method for Segmenting Liver Cancer Lesions in CT Scans

Liver cancer is a highly lethal disease, and both early clinical diagnosis and timely treatment require accurate identification of liver cancer lesions. This process typically involves radiologists manually extracting lesion areas from CT scans, which is time-consuming and labor-intensive. In this paper, we propose an automatic liver cancer segmentation method based on deep supervision to assist radiologists. The method is built upon the U-Net architecture and computes supervised loss between the encoder's output at each layer passing through the segmentation head and the ground truth mask. To facilitate the learning of features at different levels, we also introduce an edge Dice loss to achieve mixed deep supervision. This edge loss not only enhances edge information but also increases the emphasis on small lesions during the optimization process. Finally, we validated our method on the LITS dataset. The results demonstrate that our method improves liver cancer segmentation performance by 5.6% in Dice score and 7.0% in Volume Overlap Error (VOE), respectively, indicating its effectiveness in enhancing liver tumor segmentation performance.