MedCL: Learn Consistent Anatomy Distribution for Scribble-supervised Medical Image Segmentation
Keywords: medical image segmentation, weakly supervised learning, Scribble annotation, Data augmentation
TL;DR: We propose a clustering-based framework, called MedCL, to learn the inherent anatomy distribution of medical labels for scribble-supervised segmentation.
Abstract: Curating large-scale fully annotated datasets is expensive, laborious, and cumbersome, especially for medical images. Several methods have been proposed in the literature that make use of weak annotations in the form of scribbles. However, these approaches require large amounts of scribble annotations, and are only applied to the segmentation of regular organs, which are often unavailable for the disease species that fall in the long-tailed distribution. Motivated by the fact that the medical labels have anatomy distribution priors, we propose a scribble-supervised clustering-based framework, called MedCL, to learn the inherent anatomy distribution of medical labels. Our approach consists of two steps:i) Shuffle the features with intra- and inter-image mix operations, and ii) Perform feature clustering and regularize the anatomy distribution at both local and global levels. Combined with a small amount of weak supervision, the proposed MedCL is able to segment both regular organs and challenging irregular pathologies. We implement MedCL based on SAM and UNet backbones, and evaluate the performance on three open datasets of regular structure (MSCMRseg), multiple organs (BTCV) and irregular pathology (MyoPS). It is shown that even with less scribble supervision, MedCL substantially outperforms the conventional segmentation methods. Our code will be released upon acceptance.
Primary Subject Area: Segmentation
Secondary Subject Area: Learning with Noisy Labels and Limited Data
Paper Type: Methodological Development
Registration Requirement: Yes
Visa & Travel: Yes
Submission Number: 21
Loading