Detection of coronary lesions in Kawasaki disease by Scaled-YOLOv4 with HarDNet backbone

Ho-Chang Kuo; Shih-Hsin Chen; Yi-Hui Chen; Yu-Chi Lin; Chih-Yung Chang; Yun-Cheng Wu; Tzai-Der Wang; Ling-Sai Chang; I-Hsin Tai; Kai-Sheng Hsieh

doi:10.3389/fcvm.2022.1000374

Detection of coronary lesions in Kawasaki disease by Scaled-YOLOv4 with HarDNet backbone

Front Cardiovasc Med. 2023 Jan 20:9:1000374. doi: 10.3389/fcvm.2022.1000374. eCollection 2022.

Authors

Ho-Chang Kuo¹, Shih-Hsin Chen², Yi-Hui Chen^{1

3}, Yu-Chi Lin¹, Chih-Yung Chang², Yun-Cheng Wu², Tzai-Der Wang⁴, Ling-Sai Chang¹, I-Hsin Tai^{5

6}, Kai-Sheng Hsieh⁷

Affiliations

¹ Department of Pediatrics, Kawasaki Disease Center, Kaohsiung Chang Gung Memorial Hospital, College of Medicine, Chang Gung University, Kaohsiung, Taiwan.
² Department of Computer Science and Information Engineering, Tamkang University, New Taipei City, Taiwan.
³ Department of Information Management, Chang Gung University, Kaohsiung, Taiwan.
⁴ Department of E-Sport Technology Management, Cheng Shiu University, Kaohsiung, Taiwan.
⁵ Department of Medicine, College of Medicine, China Medical University, Taichung, Taiwan.
⁶ Department of Pediatric Cardiology, China Medical University Children's Hospital, China Medical University, Taichung, Taiwan.
⁷ Center of Structure and Congenital Heart Disease/Ultrasound and Department of Cardiology, Children's Hospital, China Medical University, Taichung, Taiwan.

Abstract

Introduction: Kawasaki disease (KD) may increase the risk of myocardial infarction or sudden death. In children, delayed KD diagnosis and treatment can increase coronary lesions (CLs) incidence by 25% and mortality by approximately 1%. This study focuses on the use of deep learning algorithm-based KD detection from cardiac ultrasound images.

Methods: Specifically, object detection for the identification of coronary artery dilatation and brightness of left and right coronary artery is proposed and different AI algorithms were compared. In infants and young children, a dilated coronary artery is only 1-2 mm in diameter than a normal one, and its ultrasound images demonstrate a large amount of noise background-this can be a considerable challenge for image recognition. This study proposes a framework, named Scaled-YOLOv4-HarDNet, integrating the recent Scaled-YOLOv4 but with the CSPDarkNet backbone replaced by the CSPHarDNet framework.

Results: The experimental result demonstrated that the mean average precision (mAP) of Scaled-YOLOv4-HarDNet was 72.63%, higher than that of Scaled YOLOv4 and YOLOv5 (70.05% and 69.79% respectively). In addition, it could detect small objects significantly better than Scaled-YOLOv4 and YOLOv5.

Conclusions: Scaled-YOLOv4-HarDNet may aid physicians in detecting KD and determining the treatment approach. Because relatively few artificial intelligence solutions about images for KD detection have been reported thus far, this paper is expected to make a substantial academic and clinical contribution.

Keywords: HarDNet; Kawasaki disease; Scaled-YOLOv4; coronary dilatation and brightness; deep learning; echocardiography; object detection.

Grants and funding

This project was supported by the National Science and Technology Council with ID (NSTC 109-2221-E-032-042 and NSTC 109-2221-E-230-006) and Kaohsiung Chang Gung Memorial Hospital (Project Nos. CMRPD3M0011, CORPG8L0271, and CORPG8M0091).