Multi-pathology detection and lesion localization in WCE videos by using the instance segmentation approach.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Current automatic diagnostic systems primarily detect single, known pathologies.
• Existing systems for capsule endoscopy struggle with multi-pathology detection and lesion localization.
• The medical community requires systems capable of detecting coexisting pathologies and precisely localizing lesions.

Purpose of the Study:

• To develop a multi-pathology detection system for wireless capsule endoscopy (WCE) images.
• To enhance lesion localization capabilities in automatic diagnostic systems.
• To address the limitations of current systems in detecting and localizing multiple pathologies simultaneously.

Main Methods:

• Utilized an instance segmentation approach, specifically Mask Improved RCNN (MI-RCNN), for multi-pathology detection.
• Implemented a novel training strategy based on the second momentum for training RCNN-based systems.
• Tested the system on the public KID database, including pathologies like bleeding, angioectasias, polyps, and inflammatory lesions.

Main Results:

• The proposed MI-RCNN system demonstrated significant improvements in multi-pathology detection and localization.
• The novel training strategy enhanced the performance of RCNN and PANet models.
• Achieved performance increases of nearly 7% over the PANet model with the new training approach.

Conclusions:

• Instance segmentation, particularly with MI-RCNN, is highly effective for multi-pathology detection in WCE images.
• The novel training strategy offers a significant performance boost for RCNN-based systems.
• This research presents a novel, comprehensive system for WCE pathology diagnosis, meeting critical medical needs.