A deep CNN model for anomaly detection and localization in wireless capsule endoscopy images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Wireless capsule endoscopy (WCE) is crucial for examining the gastrointestinal tract, but manual inspection of videos is challenging.
• Computer-aided diagnostic tools can improve the efficiency and accuracy of WCE video analysis.
• Existing methods often focus on specific anomalies, lacking a unified approach for simultaneous detection.

Purpose of the Study:

• To propose a novel deep learning model, WCENet, for comprehensive anomaly detection and localization in WCE images.
• To develop a generic method capable of identifying various common gastrointestinal anomalies simultaneously.

Main Methods:

• A two-phase deep convolutional neural network (CNN) approach named WCENet was developed.
• Phase 1: An attention-based CNN classifies images into normal, polyp, vascular, or inflammatory categories.
• Phase 2: Fusion of Grad-CAM++ and SegNet localizes anomalies in abnormal images.

Main Results:

• WCENet classifier achieved 98% accuracy and 99% area under the ROC curve.
• WCENet segmentation model obtained 81% frequency weighted intersection over union and 56% average dice score.
• WCENet outperformed nine other state-of-the-art models on the KID dataset.

Conclusions:

• WCENet demonstrates high performance in detecting and localizing multiple gastrointestinal anomalies from WCE images.
• The proposed model shows significant potential for integration into clinical diagnostic workflows.
• This generic approach offers a more efficient alternative to specialized anomaly detection methods.