Self-supervised 3D medical image segmentation by flow-guided mask propagation learning.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Manual annotation of 3D medical images is a time-consuming bottleneck in deep learning segmentation.
• Existing self-supervised mask propagation (SMP) methods often overlook volumetric context, relying on 2D information.
• Previous methods like Vol2Flow have limitations in capturing local and global information, leading to error accumulation.

Purpose of the Study:

• To introduce Flow2Mask, a novel SMP method to overcome limitations of prior approaches.
• To improve unsupervised learning of inter-slice flow fields using a Local-to-Global (L2G) loss with curriculum learning.
• To enhance segmentation accuracy and reduce annotation burden in 3D medical imaging.

Main Methods:

• Developed Flow2Mask, a novel self-supervised mask propagation method for 3D medical image segmentation.
• Introduced Local-to-Global (L2G) loss for unsupervised learning of inter-slice flow fields, incorporating curriculum learning.
• Implemented Inter-Slice Smoothness (ISS) loss for consistent and continuous slice-to-slice changes and an automatic slice selection strategy.

Main Results:

• Flow2Mask significantly outperforms previous SMP methods on abdominal datasets (Sliver, CHAOS, 3D-IRCAD).
• Achieved improvements in mean Dice Similarity Coefficient (DSC) of +2.1%, +8.2%, and +4.0% over Vol2Flow.
• Demonstrated substantial gains when used as a mask completion tool for weakly-supervised and self-supervised few-shot segmentation.

Conclusions:

• Flow2Mask effectively addresses limitations of prior SMP methods by leveraging volumetric context and improved loss functions.
• The method offers a more accurate and efficient solution for 3D medical image segmentation, reducing manual annotation effort.
• Flow2Mask provides a valuable contribution to medical image analysis, with code available for further research.