Performance Evaluation of Convolutional Segmentation Models with Human Hand Thermal Images (H2TI) Dataset

Abstract

Disturbances such as inflammation, edema, and neural activation disorders in the human body lead to local thermal asymmetries. With the recently developed thermal imaging systems, it is possible to detect these asymmetries in the human body. Since the hand performs many functions in human daily activities, it may be subject to abrasion due to overuse. These overuse-related disturbances give a neural and circulatory reaction. Thermal imaging is used to detect these disturbances. In the observation of thermal asymmetries, the regions of interest in the hand should be segmented from the background. When this segmentation process is performed by experts, it may cause problems such as both waste of time and different evaluations. In this study, Human Hand Thermal Image “H2TI” dataset is introduced. Commonly used segmentation models were trained with the H2TI dataset. As a result of the training process, the highest success was achieved by the “Linknet_efficientnet” model with a mIoU (Mean Intersection Over Union) value of 84.5%. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.