Precise Bleeding and Red lesions localization from Capsule Endoscopy using Compact U-Net.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Wireless capsule endoscopy (WCE) generates extensive video data, demanding significant time for manual diagnosis.
• Computational methods are crucial for enhancing the efficiency and accuracy of WCE diagnostic procedures.
• Detecting gastrointestinal bleeding and red lesions requires precise image analysis.

Purpose of the Study:

• To introduce a compact U-Net model for automated detection and segmentation of bleeding and red lesions in WCE videos.
• To evaluate the performance of the proposed compact U-Net against the original U-Net and other existing methods.
• To demonstrate the model's efficiency in terms of training time and memory consumption.

Main Methods:

• Development of a compact U-Net architecture with reduced encoder-decoder pairs.
• Training and evaluation of the compact U-Net on WCE datasets.
• Comparative analysis against the original U-Net and other state-of-the-art methods using metrics like Dice score.

Main Results:

• The compact U-Net achieved performance on par with the original U-Net.
• The proposed model demonstrated significantly faster training and reduced memory requirements.
• A Dice score of 91% was achieved on a blind-tested WCE dataset, outperforming other reported methods.

Conclusions:

• The compact U-Net offers an efficient and accurate solution for analyzing WCE videos.
• This AI model can improve diagnostic efficiency and accuracy in detecting gastrointestinal abnormalities.
• The reduced computational demands make the compact U-Net suitable for practical clinical applications.