DFU_VIRNet: A novel Visible-InfraRed CNN to improve diabetic foot ulcer classification and early detection of ulcer risk zones

Abstract

Diabetic Foot Ulcer (DFU) is considered the most serious complication suffered by people living with Diabetes Mellitus (DM). The development of an advanced stage of DFU may lead to complete or partial lower limb amputation. In this paper a novel Deep Convolutional Neural Network (CNN) architecture, DFU_VIRNet, is proposed for automatic classification of abnormal skin (DFU) versus normal skin (healthy skin). Visible and infrared (thermography) images are used to train and validate the proposed CNN. Furthermore, a new method based on estimation maps is presented to detect risk zones with high probability of the patient developing DFU. To verify the generality of the proposed approach, the performance of DFU_VIRNet was evaluated using four recently published visible datasets and one visible-infrared dataset presented in this paper; for a total of five datasets, which contain samples of abnormal skin (DFU), normal skin (healthy skin), Ischaemia and Infection. The results showed that DFU_VIRNet outperformed the current state-of-the-art results with 0.9923 (AUC) and 0.9600 (F1-score) for DFU classification (PEDIS-classification), 0.9982 (AUC) and 0.9928 (F1-score) for Ischaemia classification, and 0.9121 (AUC) and 0.8363 (F1-score) for Infection classification. The high performance of DFU_VIRNet is attributed to a novel learning mechanism proposed in this work, which is referred as GAP-2D-DLSA-IMG substructure and is used to significantly avoid overfitting and increase the perceptual field of DFU_VIRNet.