Endo-SemiS: Towards Robust Semi-Supervised Image Segmentation for Endoscopic Video.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate segmentation of endoscopic video frames is crucial for medical diagnosis and procedures.
• Limited availability of annotated data hinders the performance of deep learning models in medical image segmentation.
• Existing methods often struggle to leverage unlabeled data effectively.

Purpose of the Study:

• To introduce Endo-SemiS, a semi-supervised segmentation framework designed for reliable segmentation of endoscopic video frames.
• To improve the utilization of both labeled and unlabeled data in endoscopic video analysis.
• To achieve state-of-the-art performance in segmentation tasks with minimal annotations.

Main Methods:

• Endo-SemiS employs cross-supervision between two networks, uncertainty-guided pseudo-labeling, joint pseudo-label supervision, and mutual learning.
• A separate corrective network leverages spatiotemporal information from video data.
• The framework is designed to maximize the utility of unlabeled data.

Main Results:

• Endo-SemiS demonstrated substantially superior results compared to state-of-the-art methods on two clinical datasets.
• The framework achieved high performance in kidney stone laser lithotomy (ureteroscopy) and polyp screening (colonoscopy).
• Effective utilization of limited labeled data and unlabeled data was key to the performance gains.

Conclusions:

• Endo-SemiS offers a robust solution for semi-supervised segmentation in endoscopic videos, especially when labeled data is scarce.
• The proposed framework significantly advances the capabilities of automated analysis in medical endoscopy.
• The publicly available code facilitates further research and application in the field.