Pick and mix reliable pseudo labels for scribble-supervised medical image segmentation

Abstract

Scribble-supervised segmentation methods have attracted significant attention in the field of medical imaging because of their potential to alleviate the data annotation burden. However, these methods often underperform due to a lack of sufficient supervision. Various methods have attempted to enrich the supervisory signals in different ways, including mixing pseudo labels from different samples (referred as Mixup-based method). However, these methods primarily focus on the quantity of enriched supervisory signals, disregarding their quality. This oversight presents a major drawback in that low-quality signals are often contaminated with the noise, thus can lead to undermine performance. Therefore, it is crucial to not only introduce diverse supervisory signals but also ensure their quality and reliability. Motivated by this understanding, we propose a new framework named Pick & Mix, which builds upon the Mixup-based method. In the first step, we leverage the consistency of intra-class features to assess the reliability of pseudo-labels. To enhance the quality of pseudo labels, we assign lower weights to those unreliable pseudo-labels to mitigate the noise effect in the training process. Furthermore, we utilize a threshold to pick reliable pseudo-labels based on their reliability score. In the second step, we mix the reliable pseudo-labels from various samples and generate corresponding mixed images to provide richer supervisory signals for model training. In this manner, we enhance the quality of supervisory signals by generating and picking reliable ones, as well as enrich the quantity of these signals through a process of mixing. Finally, we evaluated our framework on three publicly available datasets: ACDC, MSCMRseg, and BraTS2020. The experimental results demonstrate that our approach achieves state-of-the-art performance.