Time-based self-supervised learning for Wireless Capsule Endoscopy.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Diagnostics

Background:

• Deep learning models require extensive labeled data, which is scarce in medical imaging.
• Class imbalance, with more healthy samples than pathological ones, poses a challenge for diagnostic systems.

Purpose of the Study:

• To develop a self-supervised learning method for wireless endoscopy videos that does not require initial labels or balanced data.
• To leverage the inherent structure of temporal data for improved diagnostic performance.

Main Methods:

• A custom self-supervised learning method tailored for wireless endoscopy videos.
• Utilizing the temporal axis to infer inherent data structure.
• Evaluating performance on polyp detection and other domain-specific applications.

Main Results:

• Achieved state-of-the-art results in polyp detection on the CAD-CAP dataset.
• Demonstrated improved detection rates even with severe class imbalance.
• Obtained 95.00 ± 2.09% Area Under the Curve and 92.77 ± 1.20% accuracy for polyp detection.

Conclusions:

• Self-supervised learning is effective for medical imaging tasks with limited or imbalanced data.
• The proposed method successfully extracts valuable information from the temporal dynamics of endoscopy videos.
• This approach offers a promising solution for enhancing computer-aided diagnostic systems.