Semi-Supervised Medical Image Segmentation with Dual-View Differential Feature Reinjection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Large-scale 3D medical imaging data is increasingly available, enabling advanced segmentation studies.
• Voxel-wise annotation for 3D medical images is expensive and time-consuming.
• Current semi-supervised methods often fail to utilize semantic cues, leading to weak supervision and blurred boundaries.

Purpose of the Study:

• To develop a novel semi-supervised framework for 3D medical image segmentation that addresses the limitations of existing methods.
• To improve the accuracy and robustness of segmentation in low-annotation scenarios.
• To enhance the utilization of semantic information within the segmentation process.

Main Methods:

• Proposed a Dual-View Differential Feedback (DVDF) framework incorporating a feature-reinjection-guided feedback loop.
• Introduced a Semantic Refinement Decoder (SRD) to improve decoder consistency and capture fine details.
• Integrated a Semantic Aggregation Attention (SAA) module for enhanced contextual information aggregation and cross-scale semantic prior establishment.

Main Results:

• The DVDF framework demonstrated consistent improvements in key metrics on Left Atrium (LA) and Pancreas-CT datasets under low-annotation conditions.
• The proposed methods effectively transformed prediction discrepancies into learnable supervisory signals for uncertain regions.
• Enhanced decoder consistency and improved capture of fine details were observed.

Conclusions:

• The DVDF framework offers an effective and robust solution for 3D medical image segmentation, particularly in data-scarce environments.
• The integration of DVDF, SRD, and SAA modules significantly enhances segmentation performance on complex structures and ambiguous boundaries.
• The study validates the potential of leveraging differential feedback and semantic refinement for semi-supervised medical image analysis.