Segmentation with mixed supervision: Confidence maximization helps knowledge distillation.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Machine Learning

Background:

• Deep neural networks (DNNs) achieve high performance in medical image segmentation.
• DNNs require large datasets with pixel-wise annotations, which are costly and time-consuming to obtain.
• Limited annotated data hinders DNN applicability in many real-world medical scenarios.

Purpose of the Study:

• To develop a novel mixed-supervision framework for medical image segmentation.
• To address the challenge of limited annotated data in deep learning models.
• To improve segmentation accuracy by effectively utilizing weakly supervised and unlabeled data.

Main Methods:

• Proposed a dual-branch architecture with a teacher (strong supervision) and student (limited supervision) branch.
• Integrated a Shannon entropy loss for confident predictions in the student branch.
• Incorporated a Kullback-Leibler (KL) divergence term to transfer knowledge from the teacher to the student branch.

Main Results:

• The proposed method significantly outperforms existing mixed-supervision and semi-supervised approaches for semantic segmentation.
• The student branch, trained with reduced supervision and guided by the teacher, achieved superior performance compared to the teacher branch.
• Demonstrated the effectiveness of Shannon entropy minimization over pseudo-mask generation for leveraging unlabeled data.

Conclusions:

• The synergistic combination of entropy and KL divergence losses enhances segmentation performance under mixed supervision.
• The developed dual-branch architecture effectively leverages limited annotations for robust medical image segmentation.
• This approach offers a promising solution for training deep learning models in data-scarce medical imaging domains.