Labeled-to-unlabeled distribution alignment for partially-supervised multi-organ medical image segmentation

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Partially-supervised multi-organ segmentation uses multiple datasets with single-organ labels.
• Limited labeled data and lack of background distinction cause distribution mismatch.
• Existing pseudo-labeling methods struggle with this distribution mismatch.

Purpose of the Study:

• To propose a novel framework, Labeled-to-Unlabeled Distribution Alignment (LTUDA), to address distribution mismatch in partially-supervised multi-organ segmentation.
• To align feature distributions and enhance the discriminative capabilities of segmentation models.
• To improve the performance of medical image segmentation using limited labeled data.

Main Methods:

• Introduced a cross-set data augmentation strategy for region-level mixing of labeled and unlabeled organs.
• Proposed a prototype-based distribution alignment method to reduce intra-class variation and improve foreground-background separation.
• Utilized consistency between prototype classifiers and a linear classifier for distribution alignment.

Main Results:

• The LTUDA framework demonstrated superior performance on the AbdomenCT-1K dataset.
• Significant improvements were observed on a combined dataset of LiTS, MSD-Spleen, KiTS, and NIH82.
• The proposed method outperformed existing state-of-the-art partially-supervised methods and even surpassed fully-supervised approaches.

Conclusions:

• The LTUDA framework effectively mitigates distribution mismatch in partially-supervised multi-organ segmentation.
• The proposed methods enhance model discriminative capability and segmentation accuracy.
• LTUDA offers a promising solution for medical image segmentation tasks with limited annotations.