Deep ensemble framework with Bayesian optimization for multi-lesion recognition in capsule endoscopy images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Wireless capsule endoscopy (WCE) generates numerous images, increasing the risk of fatigue-induced errors and misdiagnosis.
• Accurate identification of gastrointestinal (GI) lesions is crucial for timely diagnosis and treatment.

Purpose of the Study:

• To develop and evaluate a deep ensemble framework for automatic recognition of four common GI lesions in WCE images.
• To improve diagnostic accuracy and reduce clinician workload in WCE analysis.

Main Methods:

• A deep ensemble framework combining CA-EfficientNet-B0, ECA-RegNetY, and Swin transformer was utilized.
• Transfer learning and attention modules were incorporated into base learners for optimization.
• Bayesian optimization determined weights for combining base learner outputs for multi-lesion and normal image classification.

Main Results:

• The ensemble model achieved an accuracy of 84.31%, m-Precision of 88.60%, m-Recall of 79.36%, and m-F1-score of 81.08% on a dataset of 8358 images.
• The model demonstrated superior performance compared to mainstream deep learning models.
• The framework effectively recognized angioectasia, bleeding, erosions, and polyps.

Conclusions:

• The proposed deep ensemble framework significantly improves the classification performance for GI diseases detected via WCE.
• This AI-driven approach can assist clinicians in making initial diagnoses, potentially reducing diagnostic errors and enhancing patient care.