One-Shot Weakly-Supervised Segmentation in 3D Medical Images.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Deep neural networks (DNNs) for medical image segmentation demand large, accurate datasets.
• Reducing annotation effort is critical for clinical adoption.
• One-shot and weakly-supervised learning offer promising solutions.

Purpose of the Study:

• To develop an innovative framework for 3D medical image segmentation using one-shot and weakly-supervised learning.
• To minimize the requirement for extensive manual annotations in medical imaging tasks.
• To improve the robustness and accuracy of segmentation models under challenging conditions.

Main Methods:

• A propagation-reconstruction network to transfer annotations from a single volume.
• A multi-level similarity denoising module for refining initial annotations.
• Self-support prototypes for targeted refinement of segmentation borders.
• A noisy label training strategy for final model training.

Main Results:

• Significant improvements over state-of-the-art methods in 3D medical image segmentation.
• Robust performance demonstrated on CT and MRI datasets, even with class imbalance and low contrast.
• Successful application of one-shot and weakly-supervised learning principles to reduce annotation burden.

Conclusions:

• The proposed framework effectively addresses the challenge of limited annotations in 3D medical image segmentation.
• The method shows strong potential for practical clinical applications by enhancing efficiency and accuracy.
• Future work can explore further refinements and applications across diverse medical imaging modalities.