A Deep-Learning Aided Diagnostic System in Assessing Developmental Dysplasia of the Hip on Pediatric Pelvic Radiographs

• Published in: Frontiers in pediatrics Vol. 9; p. 785480
• Main Authors: Xu, Weize, Shu, Liqi, Gong, Ping, Huang, Chencui, Xu, Jingxu, Zhao, Jingjiao, Shu, Qiang, Zhu, Ming, Qi, Guoqiang, Zhao, Guoqiang, Yu, Gang
• Format: Journal Article
• Language: English
• Published: Switzerland Frontiers Media S.A 08.03.2022
• Subjects: aided diagnostic system; deep learning; developmental dysplasia of the hip; high-resolution network; Pediatrics; three-stage pipeline; deep learning; high-resolution network; developmental dysplasia of the hip; aided diagnostic system; three-stage pipeline
• ISSN: 2296-2360; 2296-2360
• DOI: 10.3389/fped.2021.785480

Developmental dysplasia of the hip (DDH) is a common orthopedic disease in children. In clinical surgery, it is essential to quickly and accurately locate the exact position of the lesion, and there are still some controversies relating to DDH status. We adopt artificial intelligence (AI) to solve the above problems. In this paper, automatic DDH measurements and classifications were achieved using a three-stage pipeline. In the first stage, we used Mask-RCNN to detect the local features of the image and segment the bony pelvis, including the ilium, pubis, ischium, and femoral heads. For the second stage, local image patches focused on semantically related areas for DDH landmarks were extracted by high-resolution network (HRNet). In the third stage, some radiographic results are obtained. In the above process, we used 1,265 patient x-ray samples as the training set and 133 samples from two other medical institutions as the verification set. The results of AI were compared with three orthopedic surgeons for reliability and time consumption. AI-aided diagnostic system's Tönnis and International Hip Dysplasia Institute (IHDI) classification accuracies for both hips ranged from 0.86 to 0.95. The measurements of numerical indices showed that there was no statistically significant difference between surgeons and AI. Tönnis and IHDI indicators were similar across the AI system, intermediate surgeon, and junior surgeon. Among some objective interpretation indicators, such as acetabular index and CE angle, there were good stability and consistency among the four observers. Intraclass consistency of acetabular index and CE angle among surgeons was 0.79-0.98, while AI was 1.00. The measurement time required by AI was significantly less than that of the doctors. The AI-aided diagnosis system can quickly and automatically measure important parameters and improve the quality of clinical diagnosis and screening referral process with a convenient and efficient way.