Hybrid deep feature fusion framework with maximum entropy based selection for accurate diabetic foot ulcer diagnosis from enhanced and segmented medical images

Abstract Diabetic foot ulcers (DFU) pose significant risks, including delayed wound healing, infection, and, in severe cases, amputation. Early diagnosis is critical to mitigate these risks and enhance patient outcomes. Manual assessment may be affected by inter-observer variability and workflow delays, which motivates the investigation of AI-assisted methods for DFU image classification. In this study, several hybrid artificial neural network (ANN) strategies were developed incorporating features from various convolutional neural network (CNN) models for diabetic ulcer image diagnosis. The use of feature selection algorithms, specifically principal component analysis (PCA) and maximum entropy score based selection (MESbS), enhanced the discriminatory power of the models. Image quality was improved by using fused filters - average and Laplacian - to enhance the images and highlight the features of the diabetic ulcer and highlight the boundaries of the lesion. The geometric active contour (GAC) algorithm facilitated isolation of damaged areas from healthy areas. AlexNet, MobileNetV2 and VGG16 models processed these regions and extracted deep feature maps. These maps are later combined to create global feature vectors, namely, AlexNet-MobileNetV2, MobileNetV2-VGG16, AlexNet-VGG16, and AlexNet-MobileNetV2-VGG16. The fusion of features from multiple CNN models gives rise to high-dimensional vectors, prompting the application of PCA and MESbS for feature selection. The proposed hybrid strategies demonstrated strong diagnostic performance on the evaluated dataset. Among the evaluated configurations, the AlexNet-MobileNetV2-VGG16-ANN model with MESbS achieved the highest performance, reaching an AUC of 99%, sensitivity of 99.7%, accuracy of 99.5%, precision of 99.5%, and specificity of 99.75%. These findings suggest that combining complementary CNN feature representations with feature selection may improve DFU classification under the evaluated experimental conditions.