Collaborative Learning for Annotation-Efficient Volumetric MR Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning excels in MR image segmentation with sufficient labeled data.
• Manual annotation of 3D MR images is labor-intensive and requires expert knowledge.

Purpose of the Study:

• To develop a deep learning method utilizing sparse annotations, specifically a single 2D slice label per 3D training MR image.
• To address the challenge of limited annotated data in 3D medical image segmentation.

Main Methods:

• A retrospective study involving 150 subjects from two public datasets (prostate and left atrium segmentation).
• A collaborative learning method integrating semi-supervised and self-supervised learning, trained on labeled central and unlabeled non-central slices.
• Quantitative evaluation using B-IoU, Dice coefficient, surface distance, and volume difference, with paired t-tests for statistical significance.

Main Results:

• The proposed method significantly improved segmentation accuracy compared to fully supervised and other semi-supervised methods.
• Mean B-IoU increased by over 10.0% for prostate segmentation (70.3% ± 7.6%) and over 6.0% for left atrium segmentation (66.1% ± 6.8%).
• The method demonstrated superior performance over interpolation consistency training (ICT) and other comparative techniques.

Conclusions:

• A collaborative learning approach effectively segments prostate and left atrium using sparse annotations.
• The developed method offers a highly accurate solution for 3D MR image segmentation with minimal labeling effort.