Min Jong Kim1, Young Hun Choi2,3, Seul Bi Lee1,4, Yeon Jin Cho1,4, Seung Hyun Lee1,4, Chang Ho Shin5, Su-Mi Shin6, Jung-Eun Cheon1,4,7. 1. Department of Radiology, Seoul National University Hospital, 101 Daehangno, Jongno-gu, Seoul, 03080, Republic of Korea. 2. Department of Radiology, Seoul National University Hospital, 101 Daehangno, Jongno-gu, Seoul, 03080, Republic of Korea. iater@snu.ac.kr. 3. Department of Radiology, Seoul National University College of Medicine, Seoul, Republic of Korea. iater@snu.ac.kr. 4. Department of Radiology, Seoul National University College of Medicine, Seoul, Republic of Korea. 5. Division of Paediatric Orthopaedics, Seoul National University Children's Hospital, Seoul National University College of Medicine, Seoul, Republic of Korea. 6. Department of Radiology, SMG-SNU Boramae Medical Center, Seoul, Republic of Korea. 7. Institute of Radiation Medicine, Seoul National University Medical Research Center, Seoul, Republic of Korea.
Abstract
BACKGROUND: Leg length discrepancy (LLD) is a common problem that can cause long-term musculoskeletal problems. However, measuring LLD on radiography is time-consuming and labor intensive, despite being a simple task. OBJECTIVE: To develop and evaluate a deep-learning algorithm for measurement of LLD on radiographs. MATERIALS AND METHODS: In this Health Insurance Portability and Accountability Act (HIPAA)-compliant retrospective study, radiographs were obtained to develop a deep-learning algorithm. The algorithm developed with two U-Net models measures LLD using the difference between the bilateral iliac crest heights. For performance evaluation of the algorithm, 300 different radiographs were collected and LLD was measured by two radiologists, the algorithm alone and the model-assisting method. Statistical analysis was performed to compare the measurement differences with the measurement results of an experienced radiologist considered as the ground truth. The time spent on each measurement was then compared. RESULTS: Of the 300 cases, the deep-learning model successfully delineated both iliac crests in 284. All human measurements, the deep-learning model and the model-assisting method, showed a significant correlation with ground truth measurements, while Pearson correlation coefficients and interclass correlations (ICCs) decreased in the order listed. (Pearson correlation coefficients ranged from 0.880 to 0.996 and ICCs ranged from 0.914 to 0.997.) The mean absolute errors of the human measurement, deep-learning-assisting model and deep-learning-alone model were 0.7 ± 0.6 mm, 1.1 ± 1.1 mm and 2.3 ± 5.2 mm, respectively. The reading time was 7 h and 12 min on average for human reading, while the deep-learning measurement took 7 min and 26 s. The radiologist took 74 min to complete measurements in the deep-learning mode. CONCLUSION: A deep-learning U-Net model measuring the iliac crest height difference was possible on teleroentgenograms in children. LLD measurements assisted by the deep-learning algorithm saved time and labor while producing comparable results with human measurements.
BACKGROUND: Leg length discrepancy (LLD) is a common problem that can cause long-term musculoskeletal problems. However, measuring LLD on radiography is time-consuming and labor intensive, despite being a simple task. OBJECTIVE: To develop and evaluate a deep-learning algorithm for measurement of LLD on radiographs. MATERIALS AND METHODS: In this Health Insurance Portability and Accountability Act (HIPAA)-compliant retrospective study, radiographs were obtained to develop a deep-learning algorithm. The algorithm developed with two U-Net models measures LLD using the difference between the bilateral iliac crest heights. For performance evaluation of the algorithm, 300 different radiographs were collected and LLD was measured by two radiologists, the algorithm alone and the model-assisting method. Statistical analysis was performed to compare the measurement differences with the measurement results of an experienced radiologist considered as the ground truth. The time spent on each measurement was then compared. RESULTS: Of the 300 cases, the deep-learning model successfully delineated both iliac crests in 284. All human measurements, the deep-learning model and the model-assisting method, showed a significant correlation with ground truth measurements, while Pearson correlation coefficients and interclass correlations (ICCs) decreased in the order listed. (Pearson correlation coefficients ranged from 0.880 to 0.996 and ICCs ranged from 0.914 to 0.997.) The mean absolute errors of the human measurement, deep-learning-assisting model and deep-learning-alone model were 0.7 ± 0.6 mm, 1.1 ± 1.1 mm and 2.3 ± 5.2 mm, respectively. The reading time was 7 h and 12 min on average for human reading, while the deep-learning measurement took 7 min and 26 s. The radiologist took 74 min to complete measurements in the deep-learning mode. CONCLUSION: A deep-learning U-Net model measuring the iliac crest height difference was possible on teleroentgenograms in children. LLD measurements assisted by the deep-learning algorithm saved time and labor while producing comparable results with human measurements.