Cut-Puzzle mix: Scribble Guided Medical Image Segmentation without Segmentation Masks.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Fully supervised segmentation algorithms are vital for human anatomy quantification but require extensive pixel-wise annotated datasets.
• Creating fully annotated datasets is labor-intensive and costly, hindering the application of deep neural networks.
• There is a growing need for efficient segmentation methods that utilize limited annotation data.

Purpose of the Study:

• To explore training strategies for pixel-wise segmentation networks using only scribble annotations.
• To investigate the effectiveness of cut-mix and puzzle mix augmentation techniques in limited data scenarios.
• To enhance segmentation performance through consistency losses combined with cross-entropy.

Main Methods:

• Developed training strategies to learn segmentation network parameters exclusively from scribble annotations.
• Employed cut-mix and puzzle mix data augmentation techniques to maximize information from limited scribbles.
• Integrated consistency losses with cross-entropy to enforce regularization and penalize segmentation inconsistencies.

Main Results:

• The proposed methods achieved noteworthy improvements in segmentation performance.
• Evaluations on cardiac (ACDC) and MSCMR datasets demonstrated superior results compared to existing state-of-the-art methods.
• The approach effectively leverages limited scribble annotations for accurate anatomical segmentation.

Conclusions:

• Training deep neural networks for medical image segmentation using only scribble annotations is feasible and effective.
• The combination of advanced augmentation strategies and consistency losses significantly enhances segmentation accuracy.
• This work offers a promising direction for reducing annotation burden in medical image segmentation.