Transfer Learning to improve the Diagnosis of Type 2 Diabetes Mellitus (T2D)

Abstract

Transfer Learning is a Deep Learning technique that is currently being used in early and non-invasive diagnosis of T2D. The objective of this work is to design and implement a Transfer Learning model trained with images of skin patches belonging to healthy people and diabetic foot patients. The research methodology was constituted by 3 phases (Analysis and Design, Development and Evaluation) composed of 5 steps that comply with the proposed objective. Several convolutional neural network (CNN) models were developed: CNN built from scratch, AlexNet, CNN with data augmentation technique, FE-VGG16, FE-ResNet50 and FT-VGG16. These models were evaluated using a set of metrics derived from the confusion matrix, the Receiver Operating Characteristic curve (ROC) of each model and the value corresponding to the area under the curve (AUC). The best performance corresponded to FT-VGG16 model that fuses VGG-16 pretrained model with a block of fully connected layers. Finally, satisfactory results are reported and allow us to conclude that the application of Transfer Learning models for the classification of diabetic foot images constitutes a viable tool for the non-invasive diagnosis of T2D.