Diabetic foot thermal image segmentation using Double Encoder-ResUnet (DE-ResUnet).

The use of thermography in the early diagnosis of Diabetic Foot (DF) has proven its effectiveness in identifying areas of the plantar foot that are susceptible to ulcer development. Segmentation of the foot sole is one of the most pertinent technical issues that must be performed with great precision. However, because of the inherent difficulties of foot thermal images, such as unclarity and the existence of ambiguities, segmentation approaches have not demonstrated sufficiently accurate and reliable results for clinical use. In this study, we aim to develop a fully automated, robust and accurate segmentation of the diabetic foot. To this end, we propose a deep neural network architecture adopting the encoder-decoder concept called Double Encoder-ResUnet (DE-ResUnet). This network combines the strengths of residual network and U-Net architecture. Moreover, it takes advantage of RGB (Red, Green, Blue) colour images and fuses thermal and colour information to improve segmentation accuracy. Our database consists of 398 pairs of thermal and RGB images. The population includes two groups. The first group of 54 healthy subjects. And a second group of 145 diabetic patients from the National Hospital Dos de Mayo in Peru. The dataset is splitted into 50% for training, 25% for validation and the last 25% is used for testing. This proposed model provided robust and accurate automatic segmentations of the DF and outperformed other state of the art methods with an average intersection over union (IoU) of 97%. In addition, it is able to accurately delineate the part of toes and heels which are high risk regions for ulceration.