A semi-supervised multi-connection contrastive learning framework for x-ray lung segmentation based on mutual...

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning for medical image segmentation requires extensive labeled data, which is labor-intensive and costly.
• Limited computing resources and storage in portable medical devices necessitate offline, compact model deployment.
• High data security requirements often restrict cloud-based processing for medical data.

Purpose of the Study:

• To develop high-performance, tiny offline models for edge deployment in medical image segmentation.
• To create a segmentation model suitable for clinical practice with limited computational resources.
• To address the challenges of data labeling costs and device limitations in medical AI.

Main Methods:

• A semi-supervised framework utilizing contrastive learning for organ contour segmentation.
• Incorporation of multiple consistency alignment and mutual distillation mechanisms.
• Adaptable backbone design to balance performance and speed requirements for diverse applications.

Main Results:

• Achieved high Dice scores for lung segmentation (e.g., 0.9636 on JSRT dataset) using only two labeled images.
• Developed an inference model with only 1.15 million parameters, demonstrating significant model compression.
• Validated the framework on three distinct chest X-ray datasets (JSRT, Montgomery County, Shenzhen Hospital).

Conclusions:

• The proposed framework demonstrates leading performance and suitability for edge deployment in clinical settings.
• The model's small size and high accuracy make it practical for resource-constrained medical devices.
• The study highlights the potential of semi-supervised learning and contrastive methods for efficient medical image analysis.