Explainable AI based automated segmentation and multi-stage classification of gastroesophageal reflux using machine...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Gastroenterology and Computational Pathology

Background:

• Gastrointestinal reflux diseases (GERD) affect millions globally, with diagnosis relying on time-consuming video endoscopy analysis.
• Manual analysis of extensive endoscopic imagery presents diagnostic challenges for healthcare providers.
• Computer-aided diagnostic techniques are crucial for efficient and accurate gastrointestinal disease detection.

Purpose of the Study:

• To develop and evaluate an AI-powered system for the automated diagnosis of GERD from video endoscopy images.
• To improve the speed and accuracy of GERD detection and segmentation using advanced machine learning models.

Main Methods:

• Utilized Yolov5 for object detection to identify regions of interest in endoscopic images.
• Employed DeepLabV3+ for precise segmentation of abnormal regions associated with GERD.
• Extracted features from segmented images for classification using machine learning models and a custom deep neural network.

Main Results:

• DeepLabV3+ achieved 95.2% segmentation accuracy and a 93.3% F1 score.
• A custom dense neural network reached 90.5% classification accuracy for GERD.
• Support Vector Machine (SVM) demonstrated the highest accuracy among traditional classifiers at 87%.

Conclusions:

• The integrated approach combining object detection, deep learning segmentation, and machine learning classification enables efficient GERD identification.
• This AI system offers a promising tool for timely surveillance and management of GERD by healthcare professionals.