DMSPS: Dynamically mixed soft pseudo-label supervision for scribble-supervised medical image segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Accurate medical image segmentation requires extensive pixel-level annotations, which are time-consuming and costly for experts.
• Weakly-supervised learning with sparse labels offers a solution to reduce annotation burden while maintaining segmentation performance.

Purpose of the Study:

• To develop an efficient scribble-based framework for medical image segmentation.
• To reduce the reliance on dense pixel-level annotations.
• To improve segmentation accuracy in 3D medical imaging.

Main Methods:

• Introduced Dynamically Mixed Soft Pseudo-label Supervision (DMSPS), a scribble-based framework.
• Employed a dual-branch network with an auxiliary decoder to enhance feature capture.
• Utilized soft pseudo-labels generated by dynamically mixing decoder predictions for supervision.
• Implemented a two-stage approach to expand sparse scribbles using low-uncertainty predictions.

Main Results:

• Achieved significant improvements in average Dice Similarity Coefficient (DSC) across multiple datasets: ACDC (50.46% to 89.51%), WORD (75.46% to 87.56%), and BraTS2020 (52.61% to 76.53%).
• Outperformed five state-of-the-art scribble-supervised segmentation methods.
• Demonstrated generalizability across different segmentation backbones.

Conclusions:

• The DMSPS framework effectively reduces annotation costs for medical image segmentation.
• The proposed method achieves state-of-the-art performance in scribble-supervised medical image segmentation.
• DMSPS is a robust and generalizable approach applicable to various medical imaging tasks.