Multi-dataset Collaborative Learning for Liver Tumor Segmentation.

Abstract

Automatic segmentation of biomedical images has emerged due to its potential in improving real-world clinical processes and has achieved great success in recent years thanks to the development of deep learning. However, it is the limited availability of certain modalities of datasets and the scarcity of labels that still present challenges. In this work, we propose a workflow of MRI liver and tumor segmentation methods utilizing external publicly available datasets. By employing pseudo-labeling, unpaired image-to-image translation, and self-ensemble learning, we improve the task performance from the nnU-Net baseline model with an average Dice score of 95.7% and 72.2%, and an average symmetric surface of 1.23 mm and 15.6 mm for the whole liver and the tumor, respectively, resulting in more robust and efficient segmentation. Our results demonstrate that the utilization of external datasets can significantly enhance liver tumor segmentation performance.