Enhanced Anomaly Detection for Capsule Endoscopy Using Ensemble Learning Strategies.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Capsule endoscopy captures gastrointestinal images but faces challenges with AI model size and limited data for anomaly detection.
• Embedding AI directly into capsules is difficult due to size constraints, hindering effective disease screening.
• Scarcity of relevant datasets complicates the development of robust anomaly detection models for capsule endoscopy.

Purpose of the Study:

• To develop an efficient ensemble strategy for anomaly detection in video capsule endoscopy.
• To address the limitations of model size and data scarcity in AI-powered capsule endoscopy.
• To enhance the accuracy and robustness of anomaly detection while minimizing computational resources.

Main Methods:

• An ensemble learning strategy was employed, training a small number of individual neural networks.
• Diverse loss functions from the anomaly detection field were used to train each network independently.
• The approach was validated on the Galar and Kvasir-Capsule datasets, the largest public datasets for video capsule endoscopy.

Main Results:

• The proposed ensemble strategy achieved an AUC score of 76.86% on the Kvasir-Capsule dataset.
• An AUC score of 76.98% was achieved on the Galar dataset.
• The method outperformed existing baselines with significantly fewer model parameters, demonstrating improved efficiency.

Conclusions:

• The developed ensemble strategy offers an effective solution for anomaly detection in video capsule endoscopy.
• This approach is crucial for integrating AI into capsule endoscopy systems due to its efficiency and reduced computational requirements.
• The findings pave the way for more advanced AI applications in minimally invasive gastrointestinal diagnostics.