Curriculum learning based overcomplete U-Net for liver tumor segmentation from computed tomography images

Abstract

In this paper, we have proposed an overcomplete U-Net to perform liver tumor segmentation jointly using a curriculum learning strategy. Liver tumor segmentation is the most prominent and primary step in treating liver cancer and can also help doctors with proper diagnosis and therapy planning. However, it is challenging because of variations in shape, position, and depth of tumors and adjacent boundaries with internal organs around the liver. We have presented a promising solution by designing a U-Net-based segmentation network with two branches: an overcomplete branch to fine grade the small structures and an undercomplete branch to fine grade the high-level structures. This combination allows the network to learn all types of tumor artifacts more accurately. We also changed the conventional learning paradigm to curriculum learning where the input images are fed to the network from easy to hard ones to achieve faster convergence. Finally, our network segments the tumors directly from the whole medical images without the need for segmented liver region of interests (ROIs). The proposed network achieved a DICE score of 75% in tumor segmentation which is a decent value when compared with some existing deep learning methods for liver tumor segmentation.