Early Detection of Diabetic Foot Using Thermal Imaging and Deep Learning Techniques

Abstract

Diabetes mellitus is one of the most predominant diseases worldwide, affecting people of all ages. Foot ulcers are one of the most horrific complications of diabetes mellitus, which could result in amputation if not discovered quickly or if the ulcer is not taken care of properly. Infrared thermography has advanced significantly in recent years and has shown great promise in medical field, diagnosis of diabetic foot ulcers is one of the uses of this technology. It is a non-invasive method that involves acquiring thermal images of the patient's foot and preprocessing them before entering a deep learning trained model where the images would be classified and diagnosed.

This research is focusing on constructing a convolutional neural network to classify thermal images of feet into either "normal" or "diabetic" and increasing the model’s accuracy by conducting several experiments on different parameters such as optimizers, activation functions, and transfer learning base models. The code is scripted using python language, TensorFlow, and Keras libraries, and the experiments are simulated and plotted in form of training and validation accuracy and loss graphs using matplotlib and Excel. in each experiment only one parameter is changed for comparison and the rest remain constant throughout the trials. For example, when experimenting to find a more efficient optimizer, other parameters remain unchanged for all trials so that the results are only affected by the change in optimizer, the same methodology is repeated for the rest of the experiments. The constructed convolutional neural network uses transfer learning to compensate for the small dataset and it also increases the efficiency and speed of the training. A comparison is made between some of the best transfer learning models, VGG16, ResNet-50, and EffecientNetB3. For optimizers, Adam and SGD were compared. And regarding activation functions, sigmoid was compared to ReLU.

It was assumed that by combining the best parameter from each experiment, the best result would be obtained. And indeed, it was an improvement from other results, it reached a validation accuracy of 90.5%. then, the model was tested on some thermal images that were left out from the training dataset and got an accuracy of 86.7%.