Interpretable deep learning for enhanced multi-class classification of gastrointestinal endoscopic images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Gastrointestinal (GI) endoscopy is crucial for diagnosing GI disorders.
• Deep learning models have advanced anomaly detection in medical images.
• Existing methods often rely heavily on data augmentation.

Purpose of the Study:

• To develop an enhanced deep learning approach for improved classification accuracy in GI endoscopic images.
• To create a model that maintains moderate complexity without extensive data augmentation.
• To enhance the interpretability and accessibility of AI-driven diagnostic tools.

Main Methods:

• Utilized the Kvasir dataset with 8,000 labeled endoscopic images across eight classes.
• Employed EfficientNetB3 as the backbone architecture.
• Integrated Local Interpretable Model-agnostic Explanation (LIME) for interpretability.
• Developed a user-friendly interface using Gradio for practical application.

Main Results:

• Achieved a test accuracy of 94.25%.
• Obtained precision of 94.29% and recall of 94.24%.
• LIME saliency maps effectively highlighted critical image regions influencing predictions.

Conclusions:

• The proposed model demonstrates high accuracy and interpretability for GI endoscopic image analysis.
• The integration of LIME and a Gradio interface enhances clinical usability.
• This work highlights the potential of AI to advance medical imaging diagnostics.