CoProLITE: Constrained Proxy Learning for lIver and hepaTic lesion sEgmentation

Yuchen Fu, Song Liu, Cong Wang, Zhiwei Jiang, Juan Du, Qing Gu

Abstract: Liver and hepatic lesion segmentation is an important task in medical image analysis, which plays a crucial role in diagnosis, treatment planning and monitoring of liver diseases. We observed an ordinal layout of the feature space that aligns with CT image characteristics will improve performance on liver and hepatic lesion segmentation task. In order to enforce the samples to conform to a specific layout of the feature space, we propose a novel liver and hepatic lesion segmentation method called CoProLITE, which learns a constrained proxy for each classes. Specifically, We replace the traditional FCN-based segmentation head by a proxy learning-based head to learn feature representations of the images, and introduces constraints during the training process to guide the learning of the proxies. We extensively evaluate CoProLITE on three public datasets and compare it to state-of-the-art methods. The experimental results demonstrate the effectiveness of the proposed method.