A feature explainability-based deep learning technique for diabetic foot ulcer identification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Diabetology

Background:

• Diabetic foot ulcers (DFUs) are severe diabetes complications, risking infection and amputation.
• Current detection methods are labor-intensive and costly.
• AI, especially deep learning, offers potential for improved DFU diagnosis and treatment.

Purpose of the Study:

• To introduce the DFU_XAI framework for interpretable deep learning in DFU detection.
• To evaluate and compare the performance of six deep learning models for DFU labeling and localization.
• To enhance the clinical relevance and trustworthiness of AI in DFU management.

Main Methods:

• The DFU_XAI framework was developed to assess deep learning models.
• Six models (Xception, DenseNet121, ResNet50, InceptionV3, MobileNetV2, SNN) were evaluated.
• Interpretability techniques (SHAP, LIME, Grad-CAM) were employed.

Main Results:

• The Siamese Neural Network (SNN) model demonstrated superior performance with 98.76% accuracy.
• High precision (99.3%), recall (97.7%), F1-score (98.5%), and AUC (98.6%) were achieved by the SNN model.
• Grad-CAM heat maps provided visually interpretable ulcer localization, aiding clinical decisions.

Conclusions:

• The DFU_XAI framework enhances transparency and clinical utility of AI for DFU detection.
• The SNN model shows significant promise for accurate and interpretable DFU diagnosis.
• This AI-driven approach offers a reliable and efficient alternative to traditional DFU management methods.