AUTOMATED DETECTION OF CHILDHOOD OBESITY IN ABDOMINOPELVIC REGION USING THERMAL IMAGING BASED ON DEEP LEARNING TECHNIQUES

Abstract

Childhood obesity is a preventable disorder which can reduce the risk of the comorbidities linked with an adult obesity. In order to improve the lifestyle of the obese children, early and accurate detection is required by using some non-invasive technique. Thermal imaging helps in evaluation of childhood obesity without injecting any form of harmful radiation in human body. The goal of this proposed research is to evaluate the body surface temperature in abdominopelvic and cervical regions and to evaluate which region is best for predicting childhood obesity using thermal imaging. Next, to customize the ResNet-18 and VGG-19 architecture using transfer learning approach and to obtain the best modified classifier and to study the classification accuracy between normal and obese children. The two-study region which was selected for this study was abdominopelvic and cervical region where the mean skin surface temperature was recorded. From the two selected body regions, abdominopelvic region has depicted highest temperature difference of 10.98% between normal and obese subjects. The proposed modified ResNet-18 model produced an overall accuracy of 94.2% than the modified VGG-19 model (86.5%) for the classification of obese and normal children. Thus, this study can be considered as a non-invasive and cost-effective way for pre-screening the obesity condition in children.