Cross Pseudo Supervision for Semi-supervised Medical Image Segmentation

Peng Liu

Abstract:

Multi-organ segmentation is an important step in many medical image applications. Due to time-consuming and challenging for physicians to annotate multi-organ segmentation dataset, the existing dataset
for multi-organ segmentation usually have small number of samples. A series of works have been proposed to make use of these limited annotated data for improving the performance of multi-organ segmentation. In this paper, In this paper, we present a novel context-aware cross pseudo supervision
algorithm for semi-supervised medical image segmentation. Our method first use two networks with different initialization strategies, then we fed two overlapped patches to the network, last we use the outputs of the overlapped regions of one network to get the pseudo label to supervised another network. Experimental Results show that our proposed method perform well in multi-organ segmentation.