Feature Ranking by Variational Dropout for Classification Using Thermograms from Diabetic Foot Ulcers

Area of Science:

• Medical imaging
• Machine learning
• Diabetology

Background:

• Diabetic foot ulcers (DFUs) are a severe complication of diabetes mellitus, affecting millions globally.
• Early detection of DFU risk is crucial for preventing costly and debilitating outcomes.
• Infrared thermography shows promise in identifying abnormal foot patterns indicative of DFU risk in diabetic patients.

Purpose of the Study:

• To extract and evaluate novel state-of-the-art features from infrared thermograms for efficient diabetic foot ulcer risk classification.
• To compare the performance of classical machine learning feature extraction methods with deep learning approaches.
• To enhance a public dataset with private and synthetic data for robust model training.

Main Methods:

• Utilized the INAOE thermogram dataset, augmented with private local data and synthetic data generated via SMOTE.
• Extracted features using LASSO, random forest, and variational deep learning methods (concrete dropout, variational dropout).
• Classified subjects using a Support Vector Machine (SVM) classifier to evaluate feature robustness and performance.

Main Results:

• Variational deep learning methods, particularly concrete dropout, yielded the best performance.
• Achieved an F1 score of 90% for DFU risk classification using concrete dropout features and SVM.
• The novel feature set demonstrated a 15% performance improvement over previously established state-of-the-art features.

Conclusions:

• Deep learning-based feature extraction from infrared thermography offers a superior approach for diabetic foot ulcer risk assessment.
• The developed feature set and methodology provide a robust and accurate tool for early DFU risk identification.
• This advancement holds significant potential for improving patient outcomes and reducing healthcare burdens associated with diabetes.