A deep weakly semi-supervised framework for endoscopic lesion segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Accurate medical image segmentation is crucial for diagnosis and treatment planning.
• Deep learning methods show promise but require extensive pixel-level annotations.
• Pixel-level annotation is time-consuming and labor-intensive for radiologists.

Purpose of the Study:

• To propose a weakly semi-supervised segmentation framework, Point Segmentation Transformer (Point SEGTR).
• To reduce the significant demand for pixel-level annotations in medical image segmentation.
• To leverage both pixel-level and point-level annotations effectively for network training.

Main Methods:

• Developed a framework utilizing a small amount of fully-supervised data (pixel-level masks) and abundant weakly-supervised data (point-level annotations).
• Introduced two regularization terms: multi-point consistency and symmetric consistency, to enhance pseudo-label quality.
• Trained a student model for inference using the improved pseudo labels.

Main Results:

• Demonstrated the effectiveness and generality of Point SEGTR on three diverse endoscopy datasets.
• Showcased the method's ability to significantly reduce the requirement for pixel-level annotations.
• Achieved promising segmentation performance with reduced annotation effort.

Conclusions:

• Point SEGTR offers a valuable solution for medical image segmentation by minimizing annotation burden.
• The proposed framework is effective and generalizable across different lesion types and anatomical sites.
• This approach holds significant potential for practical clinical applications due to reduced annotation requirements.