Scribble-supervised method for cardiac tissue segmentation using position and temporal contrastive information.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Accurate pixel-level segmentation is crucial for medical diagnosis and treatment planning.
• Fully supervised methods require extensive high-quality annotations, which are often scarce and costly.
• Weakly supervised learning aims to reduce reliance on precise annotations.

Purpose of the Study:

• To develop an innovative segmentation method that combines contrastive learning with weak supervision.
• To reduce the dependency on precise annotations in medical image segmentation.
• To improve segmentation performance using limited annotation data.

Main Methods:

• Proposed Scribble Position and Temporal Contrast Learning (SPTCL) method.
• Leveraged spatial continuity in 3D medical volumes and temporal similarities across cardiac phases for contrastive learning.
• Utilized a pre-trained encoder, fine-tuned on a weakly supervised segmentation network with a dual-branch decoder.
• Fused predictions to generate pseudo-labels for iterative training with scribble annotations.

Main Results:

• Achieved a Dice coefficient of 90.5% on the ACDC dataset.
• Outperformed existing models with a 2.5% improvement over the baseline and 1.7% over the latest model.
• Reduced training time by approximately 33%.

Conclusions:

• SPTCL effectively addresses annotation scarcity in medical image segmentation.
• The method demonstrates strong potential for practical clinical deployment.
• Unites contrastive learning with weak supervision for robust feature representation and segmentation.