PrototypeGuided Meta Pixel Correction for 3D MultiLesion Segmentation With Partial Instance Annotations.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Pixel-level annotation of 3D medical volumes is time-consuming and expensive, especially for multiple lesions.
• Clinical constraints often lead to incomplete lesion annotation, posing challenges for automated segmentation.
• Existing weakly supervised methods fail under protocols with limited fully labeled instances and sparse background cues.

Purpose of the Study:

• To develop a weakly supervised segmentation framework (PIASeg) that addresses the clinical reality of extremely limited annotations.
• To enable accurate 3D medical volume segmentation even when only a few lesion instances are fully labeled.
• To improve lesion detection and segmentation efficiency in rapid screening scenarios.

Main Methods:

• PIASeg employs a meta-learning framework for iterative label refinement by correcting confident foreground predictions.
• Class-specific prototypes are utilized to filter inconsistent pseudo-labels through feature similarity.
• Prototype representations are optimized using contrastive and diversity objectives for robust feature learning.

Main Results:

• PIASeg demonstrates superior performance compared to state-of-the-art methods on three public 3D lesion datasets (LiTS, ISLES22, MS).
• The method achieves high segmentation accuracy even with extremely limited supervision, such as only one annotated lesion per volume.
• PIASeg effectively handles multilesion scenarios and heterogeneous lesion types under weak supervision.

Conclusions:

• PIASeg offers a practical solution for 3D medical volume segmentation under severe annotation constraints.
• The proposed framework significantly advances weakly supervised learning for medical imaging applications.
• This approach has the potential to reduce annotation costs and accelerate clinical screening processes.