Free-Form Deformation Approach for Registration of Visible and Infrared Facial Images in Fever Screening

作     者：Chenna, Yedukondala Narendra Dwith Ghassemi, Pejhman Pfefer, T. Joshua Casamento, Jon Wang, Quanzeng 

作者机构：US FDA Ctr Devices & Radiol Hlth Silver Spring MD 20993 USA Univ Maryland Dept Elect & Comp Engn College Pk MD 20740 USA 

出 版 物：《SENSORS》 (传感器)

年 卷 期：2018年第18卷第1期

页      面：125-125页

核心收录：

学科分类：0710[理学-生物学] 071010[理学-生物化学与分子生物学] 0808[工学-电气工程] 07[理学] 0804[工学-仪器科学与技术] 0703[理学-化学] 

基　　金：U.S. Food and Drug Administration's Medical Countermeasures Initiative (MCMi) Regulatory Science Program [16ECDRH407] 

主　　题：thermal imaging fever screening temperature measurement canthi detection multi-modality image registration free form deformation Demons algorithm cubic B-spline algorithm 

摘      要：Fever screening based on infrared (IR) thermographs (IRTs) is an approach that has been implemented during infectious disease pandemics, such as Ebola and Severe Acute Respiratory Syndrome. A recently published international standard indicates that regions medially adjacent to the inner canthi provide accurate estimates of core body temperature and are preferred sites for fever screening. Therefore, rapid, automated identification of the canthi regions within facial IR images may greatly facilitate rapid fever screening of asymptomatic travelers. However, it is more difficult to accurately identify the canthi regions from IR images than from visible images that are rich with exploitable features. In this study, we developed and evaluated techniques for multi-modality image registration (MMIR) of simultaneously captured visible and IR facial images for fever screening. We used free form deformation (FFD) models based on edge maps to improve registration accuracy after an affine transformation. Two widely used FFD models in medical image registration based on the Demons and cubic B-spline algorithms were qualitatively compared. The results showed that the Demons algorithm outperformed the cubic B-spline algorithm, likely due to overfitting of outliers by the latter method. The quantitative measure of registration accuracy, obtained through selected control point correspondence, was within 2.8 +/- 1.2 mm, which enables accurate and automatic localization of canthi regions in the IR images for temperature measurement.