Towards accurate Diabetic Foot Ulcer image classification: Leveraging CNN pre-trained features and extreme learning machine

Abstract

Diabetes mellitus (DM) can cause irreversible tissue damage in the legs, leading to foot ulcers that are difficult to heal. Early detection is crucial in preventing further complications. This study proposes a detection system for foot ulcers using a hybrid approach that combines deep convolutional neural networks (CNN) with an extreme learning machine (ELM). We explore the features of popular pre-trained models, including ResNet101, DenseNet201, MobileNetv2, EfficientNetB0, InceptionResNetv2, and NasNet mobile. Given the challenge of a limited dataset, traditional data augmentation may introduce inter-class bias. Therefore, we adopt a fusion of CNN and ELM to mitigate this issue. The experiments show promising results, with ResNet101, DenseNet201, InceptionResNetv2, MobileNetV2, NasNet mobile, and EfficientNetB0 achieving accuracies of 80%, 76.67%, 80%, 83.34%, 80%, and 80%, respectively. Our analysis reveals that MobileNetV2 provides the best feature representation, achieving the highest accuracy rate of 83.34% with zero false positives. Based on the findings, we suggest that the proposed hybrid method can accurately recognize DM foot images, providing a potential tool for early diagnosis and treatment of foot ulcers.