A weighted Gompertz fuzzy ranking-based ensemble model for diabetic foot ulcer detection in skin and thermal imagery

Abstract

Diabetic Foot Ulcers (DFUs) pose a significant health risk, often leading to severe complications in diabetic patients. Early and accurate detection of DFUs is crucial for effective intervention and management. This paper introduces an innovative ensemble methodology that leverages a weighted Gompertz function-based fuzzy ranking strategy for accurate and reliable DFU detection. The proposed technique adopts an ensemble framework where the weighted Gompertz function forms fuzzy rankings for the three fundamental classifiers, and their decision scores are integrated to yield final predictions on test cases. The ensemble model is constructed using three transfer learning-based models: InceptionV3, MobileNet, and InceptionResNetV2, which generate decision scores that are subsequently fused. The method showcases remarkable enhancements in classification accuracy, thoroughly evaluated on a dataset of 1055 foot images. Grad-CAM analysis showcases the models' focus on relevant regions, while ablation studies and comparison with the traditional ensemble method confirm our approach's reliability. These results highlight the crucial significance of hyperparameter tuning in enhancing model performance. Furthermore, our ensemble's efficacy extends to thermal imagery domains, validated by experiments on a diverse dataset comprising thermal images of diabetic foot cases. This adaptability across imaging modalities positions our methodology as a useful diagnostic tool, potentially aiding medical practitioners in efficiently detecting diabetic foot ulcers in both skin and thermal imagery domains. The proposed approach achieves an accuracy of 99.53 % on the initial dataset of foot images and attains a remarkable accuracy of 93.60 % on a separate dataset comprising thermal images. By improving diagnostic accuracy and adaptability, this approach can facilitate early DFU detection, reducing the risk of complications and enhancing patient outcomes.