Detecting and Locating Gastrointestinal Anomalies Using Deep Learning and Iterative Cluster Unification.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Gastrointestinal (GI) anomaly detection in endoscopic videos is crucial for diagnosis.
• Pixel-level annotations for training deep learning models are labor-intensive and costly.
• Existing methods often require detailed annotations, limiting scalability.

Purpose of the Study:

• To develop a novel, cost-effective methodology for automatic detection and localization of GI anomalies.
• To enable analysis of large videoendoscopy repositories using weakly annotated data.
• To improve the efficiency and accuracy of identifying abnormalities in endoscopic videos.

Main Methods:

• A three-phase approach using a weakly supervised convolutional neural network (WCNN).
• Phase 1: Frame classification (abnormal/normal) via WCNN.
• Phase 2: Salient point detection from WCNN layers.
• Phase 3: GI anomaly localization using an iterative cluster unification (ICU) algorithm with pointwise cross-feature-map (PCFM) descriptors.

Main Results:

• Achieved >80% Area Under the Curve (AUC) for both anomaly detection and localization.
• Highest AUC for anomaly detection reached 96% on conventional gastroscopy images.
• Highest AUC for anomaly localization reached 88% on wireless capsule endoscopy images.

Conclusions:

• The proposed weakly supervised methodology is effective for GI anomaly detection and localization.
• This approach offers a cost-efficient solution for analyzing large endoscopic video datasets.
• The method demonstrates generality through automatically derived image features and robust performance across different endoscopy types.