Anatomical Prior-Inspired Label Refinement for Weakly Supervised Liver Tumor Segmentation with Volume-Level Labels

Automatic liver tumor segmentation is important for assisting doctors in accurate diagnosis of liver cancer. Existing models for liver tumor segmentation usually require accurate pixel-level labels. However, acquiring such dense labels is laborious and costly. In this paper, we propose a weakly supervised method for liver tumor segmentation using volume-level labels, which can significantly reduce the manual annotation effort. Volume-level labels are propagated to image-level labels where all the slices in one CT volume share the same label, and pixel-level pseudo labels can be estimated from image-level labels. However, it will cause the label noise problem because not all slices contain tumors. To address this issue, we propose two label refinement strategies based on anatomical priors to reduce the training noise and improve the performance. Evaluation experiments on two public datasets demonstrate that our proposed method can achieve competitive results compared to other methods with stronger supervision.