PL-Seg: Partially labeled abdominal organ segmentation via classwise orthogonal contrastive learning and progressive...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate abdominal organ segmentation in CT scans is vital for diagnosis and treatment planning.
• Fully supervised deep learning requires extensive, costly voxel-level annotations, hindering multi-organ segmentation.
• Partially labeled data offers a solution to reduce annotation burden in medical image segmentation.

Purpose of the Study:

• To develop a novel framework (PL-Seg) for multi-organ segmentation in abdominal CT scans using partially labeled data.
• To improve segmentation performance by effectively utilizing both labeled and unlabeled organ data.
• To reduce the time and cost associated with obtaining dense annotations for medical imaging datasets.

Main Methods:

• PL-Seg framework utilizing partially labeled CT scans.
• Hardness-Aware Decoupled Foreground Loss (HADFL) to focus on annotated organs and adjust weights based on difficulty.
• Classwise Orthogonal Contrastive Loss (COCL) for reducing inter-class ambiguity and regularizing unlabeled regions.
• Progressive Self-Distillation (PSD) to enhance feature learning from high-resolution to low-resolution layers.

Main Results:

• PL-Seg demonstrated significant performance improvements by effectively leveraging unlabeled categories.
• Outperformed six state-of-the-art methods in segmentation accuracy, with a simpler pipeline and greater computational efficiency.
• Achieved superior results compared to existing semi-supervised methods under identical annotation costs.

Conclusions:

• PL-Seg offers an effective solution for multi-organ segmentation in abdominal CT scans using partially labeled data.
• The proposed methods (HADFL, COCL, PSD) enhance segmentation performance and efficiency.
• The release of a partially labeled medical image segmentation codebase and benchmark will foster further research in this domain.