Early detection of diabetic foot ulcers from thermal images using the bag of features technique

The most common diabetes complication, diabetic foot ulcer (DFU), can result in amputation if not managed effectively. Early diagnosis is thus critical, yet DFU symptoms cannot be reliably identified in the initial stages of the disease. This work develops an accurate, fast, reliable end-to-end traditional machine learning (ML) model that can be implemented in real-time computer-aided diagnosis (CAD) applications for DFU diagnosis. We selected the plantar thermogram database because it may help detect an increase in plantar temperature and thereby permit early diagnosis. This method is less affected by environmental conditions, such as smoke and weather. To extract features invariant to translation, rotation, and scaling transformation, we use the Scale Invariant Feature Transform (SIFT) and Speeded Up Robust Features (SURF) methods combined with the Bag Of Features (BOF) technique. We used different pretrained deep learning (DL) networks to compare ML models trained with DL and handcrafted features. This research is novel since no previous studies have used traditional ML algorithms for the diagnosis of DFU from thermal images, nor have they used SIFT, SURF, or BOF techniques. In addition, we use direct temperature files of each foot, later mapping them onto images to obtain accurate temperature distributions. Conventional ML classifiers are preferred in the final stage for binary classification between normal and DFU. The proposed model achieves 97.81% classification accuracy using a support vector machine (SVM) classifier trained with the SURF-BOF technique, along with 97.81% sensitivity, 97.9% precision, and a 0.9995 Area under the ROC curve (AUC) for DFU classification using thermal images.