Boosting Medical Image Segmentation with Partial Class Supervision

Abstract

Medical image data are often limited due to expensive acquisition and annotation processes. Directly using such limited annotated samples can easily lead to the deep learning models overfitting on the training dataset. An alternative way is to leverage the unlabeled dataset which is free to obtain in most cases. Semi-supervised methods using a small set of labeled data and large amounts of unlabeled data have received much attention. In this paper, we propose a novel semi-supervised method for medical image segmentation that uses partial class supervision. Specifically, for a given multi-class label, we extend it to generate several labeled images with partial classes annotated while others remain unannotated. The unlabeled part in the partially annotated label is supervised by a pseudo-labels approach. In addition, we project the labeled pixel values into pseudo-labels to achieve rectified pixel-level pseudo-labels. In this way, our method can effectively increase the number of training samples. The experimental results on two public medical datasets of heart and prostate anatomy demonstrate that our method outperforms the state-of-the-art semi-supervised methods. Additional experiments also show that the proposed method gives better results compared to fully supervised segmentation methods.