A Novel Weakly Supervised Framework Based On Noisy-Label Learning For Medical Image Segmentation

Obtaining accurately annotated medical images for training segmentation models is expensive, time-consuming and labor-intensive. Although a variety of approaches based on weak annotations like points, scribbles and bounding boxes have been designed to address this problem, their performance is still limited. Inspired by recent studies on noisy-label learning, we propose a novel two-stage framework where a size-constrained loss is used to directly learn from the weak annotations in the first stage and a noise-robust loss is introduced to learn from pseudo labels in the second stage. The noise-robust loss function, named Edge-Dice, is based on the confidence in the network's prediction and the pseudo labels. Our approach differs from previous works by taking a natural step towards stronger supervision, in which predictions made by weak supervision methods are incorporated into another round of training using noise-robust methods. Experiments with the ACDC 2017 dataset showed that our method achieved 86.27% Dice for left ventricular segmentation with only 1% of the full annotations, and it outperformed existing methods with the same set of weak annotations.