Combining Oversampling and Pretrained Feature Extractor For Classification Diabetic Foot Uclear Thermogram Images

Abstract

Diabetic Foot Ulcers (DFUs) represent a significant health concern, often leading to severe complications if not diagnosed and treated promptly. Early and accurate classification of DFUs is crucial for effective patient management. However, In the realm of machine learning, the imbalanced data problem is a prevalent issue that arises when the classes in a dataset are not represented equally. This study proposes a novel approach to enhance the classification performance of DFU thermogram images by integrating oversampling techniques with pretrained feature extractors. This study use pretrained model method with InceptionV3 architecture to automatically obtain features in the DFU thermogram datasets. Overall, InceptionV3 as a feature extractor resulted in satisfactory performance, achieving an accuracy of 83.1% on non-diabetic data and 81.1% on diabetic data. Subsequently, the second experiment incorporated the oversampling technique SMOTE, leading to an improvement in performance, with accuracy rising to 98.1% on non-diabetic data and 96.1% on diabetic data. Finally, the SMOTE IPF method achieved accuracy of 98.7%, with a precision of 99.1% for the diabetic class and 98.7% for the non-diabetic class, a recall of 98.2% for the diabetic class and 98.1% for the non-diabetic class, and F-Measure of 98.1% for both the diabetic and non-diabetic classes.

Keywords

• diabetic foot uclear, smote, cost sensitive learning, machine learning, imbalanced data.