Diabetes Diagnosis Using Plantar Thermogram Based on DenseNet

Abstract

In Japan, the number of diabetics is rapidly increasing due to changes in lifestyle and social environment. In the early stages of diabetes, patients have few subjective symptoms and the disease may be left untreated for a long period of time. However, if metabolic abnormalities in diabetes persist over a long period of time, the likelihood of developing complications increases. Therefore, it is important to complete the diagnosis of diabetes as early as possible. The use of plantar thermography images is expanding as one way to determine diabetes. However, conventional techniques have not been evaluated to take into account the difficulties of acquiring images in actual use environments, such as out-of-focus or low-resolution cameras. This evaluation is essential in a practical diabetes detector. In this method, we created various simulated images assuming realistic usage environments and devices, and evaluated their impact on diabetes determination accuracy using Recall, Precision, and F-measure. The diabetes determination method uses DenseNet201, a convolutional neural network specifically designed for image classification. As a model, training is performed using only the source image, either single-foot images or both-foot images. The dataset consists of plantar thermogram images of 122 diabetic and 45 non-diabetic patients published by Hernandez et al. Due to the small amount of training data, the training was augmented with image processing such as rotation and reduction. For the original image, the Recall and F-measure for the single-foot image were 96.4% and 87.1% for the original image, and 100.0% and 78.9% for both-foot images, respectively. Considering the F-measure, the classification with a single foot as input data is relatively more accurate. Furthermore, even at 87.5% reduction, there was no effect of reduced resolution on the accuracy of diabetes determination, indicating that focusing has a significant effect on the accuracy of plantar thermography images.

Keywords: Convolutional Neural Network, Diabetes Mellitus, DenseNet201

DOI: 10.54941/ahfe1004088