3D Liver and Tumor Segmentation in CT Scans with Robust Self-Paced Learning

Obed Tettey Nartey,Guowu Yang,Sarpong Kwadwo Asare,Jinzhao Wu,Dorothy Araba Yakoba Agyapong,Lady Nadia Frempong

DOI: https://doi.org/10.1109/icisce50968.2020.00302

2020-01-01

Abstract:Liver segmentation is a challenging problem where fully supervised deep learning models require large amounts of voxel-wise labels, which are usually laborious, expensive, and time consuming to obtain. However, massive non-annotated 3D CT volumes are usually easily accessible. Most deep learning based works to address semi-supervised CT medical image segmentation problems have mostly been graph-based or adversarial-based. In sharp contrast, this work investigates the potential of a semi-supervised learning method; novel self-training framework utilizing self-paced learning strategy to generate and select pseudo-labels from easy-to-transfer(well represented classes) to hard-to-transfer(less represented classes) samples. The proposed scheme introduces an effective pseudo-label generation and selection algorithm to generate highly confident pseudo-labels, jointly learn and optimize a model which can be solved by alternatively generating pseudo labels on unlabeled target data and re-training the model with the generated labels. Further, we also propose a novel class balancing framework to avoid the gradual eminence of easy-to-transfer discriminative features on pseudo-label generation utilizing class-wise normalization of confidence scores. Comprehensive experiments conducted on publicly available datasets show that the proposed method outperformed other state-of-the-art semi-supervised and some fully supervised methods on the segmentation results of tumors and achieved very competitive performance for liver segmentation.