Annotation-efficient medical image segmentation via cross-latent graphs and vector-quantized memory

Area of Science:

• Medical Imaging
• Computer-Assisted Diagnosis
• Machine Learning

Background:

• Medical image segmentation is crucial for diagnosis but hindered by extensive annotation requirements.
• Current methods struggle with scalability due to the need for large, pixel-level annotated datasets.

Purpose of the Study:

• To develop an annotation-efficient framework for medical image segmentation.
• To leverage sparse supervision (scribbles, points) to reduce annotation burden.
• To achieve high segmentation accuracy comparable to fully supervised methods.

Main Methods:

• Proposed a framework utilizing sparse supervision for medical image segmentation.
• Employed an auxiliary reconstruction branch for enhanced supervision and feature enrichment.
• Integrated a vector quantization (VQ) memory bank for dynamic pseudo-label generation.
• Utilized a cross-latent graph neural network (GNN) to capture non-local dependencies and improve predictions.

Main Results:

• Achieved competitive or superior performance against state-of-the-art weakly supervised methods on benchmark datasets (ACDC, BraTS'19, Pancreas-CT).
• Demonstrated segmentation quality approaching fully supervised accuracy.
• Showcased significant reduction in annotation costs.

Conclusions:

• The proposed framework effectively reduces annotation costs for medical image segmentation.
• It offers a scalable solution for clinical applications by minimizing reliance on pixel-level annotations.
• The approach maintains high segmentation quality, supporting reliable computer-assisted diagnosis.