Consideration of Convolutional Neural Networks for Image Processing of Capillaries

Abstract

The Convolutional Neural Network (CNN) is an effective algorithm in deep learning and the performance which the CNN brings in life problem is recognized worthily. Tobacco is one of the biggest public health threats and results in 8 million deaths every year through cardiovascular diseases, lung disorders, cancers, diabetes, and hypertension. There are several methods used in hospitals for inspecting their own health, however, they are difficult to use in daily life because all inspecting devices are large-scale and complex. Thus, the purpose of this study was to propose a new method to self-check the effect of smoking on capillaries and surface skin in daily life, then evaluate the usefulness of the proposed method. The dataset was collected from the 26 human subjects through the capillaroscopy; 13 subjects were the smoker and the 13 were the non-smoker. Through all of the results for the recognition of the difference between smokers and non-smokers, it was confirmed that conventional methods to extract featured points from the edge or corner points such as ssim (structural similarity) and sift (scale-invariant feature transform) was not so good for the image processing of capillaries. However, it was found that CNN worked well with over 80% accuracy. It was discussed that efficientnet with the compound scaling was so good for the small dataset with the comparison of resnet50, vgg16, densenet121 with one scaling factor, although COVID-19 virus affected the dataset making procedure measured from human subjects directly.