Meta Pixel Loss Correction for Medical Image Segmentation with Noisy Labels

Supervised training with deep learning has exhibited impressive performance in numerous medical image domains. However, previous successes rely on the availability of well-labeled data. In practice, it is a great challenge to obtain a large high-quality labeled dataset, especially for the medical image segmentation task, which generally needs pixel-wise labels, and the inaccurate label (noisy label) may significantly degrade the segmentation performance. In this paper, we propose a novel Meta Pixel Loss Correction (MPLC) based on a simple meta guided network for the medical segmentation that is robust to noisy labels. The core idea is to estimate a pixel transition confidence map by meta guided network to take full advantage of noisy labels for pixel-wise loss correction. To achieve this, we introduce a small size of meta dataset with the meta-learning method to train the whole model and help the meta guided network automatically learn the pixel transition confidence map in an alternative training manner. Experiments have been conducted on three medical image datasets, and the results demonstrate that our method is able to achieve superior segmentation with noisy labels compared to the existing state-of-the-art approaches.