Tuya E1, Rile Nai1, Xiang Liu1, Cen Wang2, Jing Liu1, Shijia Li3, Jiahao Huang3, Junhua Yu3, Yaofeng Zhang3, Weipeng Liu3, Xiaodong Zhang1, Xiaoying Wang4. 1. Department of Radiology, Peking University First Hospital, No.8 Xishiku Street, Xicheng District, Beijing, 100034, China. 2. Department of Radiology, Beijing Nuclear Industry Hospital, No.2, Lane 4, Sanlihe South, Xicheng District, 100045, Beijing, China. 3. Beijing Smart Tree Medical Technology Co. Ltd., No.24 Huangsi Street, Xicheng District, Beijing, 100011, China. 4. Department of Radiology, Peking University First Hospital, No.8 Xishiku Street, Xicheng District, Beijing, 100034, China. wangxiaoying@bjmu.edu.cn.
Abstract
OBJECTIVES: To explore the performance of a deep learning-based algorithm for automatic patellofemoral joint (PFJ) parameter measurements from the Laurin view. METHODS: A total of 1431 consecutive Laurin views of the PFJ were retrospectively collected and divided into two parts: (1) the model development dataset (dataset 1, n = 1230) and (2) the hold-out test set (dataset 2, n = 201). Dataset 1 was used to develop the U-shaped fully convolutional network (U-Net) model to segment the landmarks of the PFJ. Based on the predicted landmarks, the PFJ parameters were calculated, including the sulcus angle (SA), congruence angle (CA), patellofemoral ratio (PFR), and lateral patellar tilt (LPT). Dataset 2 was used to assess the model performance. The mean of three radiologists who independently measured the PFJ parameters was defined as the reference standard. Model performance was assessed by the intraclass correlation coefficient (ICC), mean absolute difference (MAD), and root mean square (RMS) compared to the reference standard. Ninety-five percent limits of agreement (95% LoA) were calculated pairwise for each radiologist, reference standard, and model. RESULTS: Compared with the reference standard, U-Net showed good performance for predicting SA, CA, PFR, and LPT, with ICC = 0.85-0.97, MAD = 0.06-5.09, and RMS = 0.09-6.90 in the hold-out test set. Except for the PFR, the remaining parameters measured between the reference standard and the model were within the 95% LoA in the hold-out test dataset. CONCLUSIONS: The U-Net-based deep learning approach had a relatively high model performance in automatically measuring SA, CA, PFR, and LPT. KEY POINTS: • The U-Net model could be used to segment the landmarks of the PFJ and calculate the SA, CA, PFR, and LPT, which could be used to evaluate the patellar instability. • In the hold-out test, the automatic measurement model yielded comparable performance with reference standard. • The automatic measurement model could still accurately predict SA, CA, PFR, and LPT in patients with PI and/or PFOA.
OBJECTIVES: To explore the performance of a deep learning-based algorithm for automatic patellofemoral joint (PFJ) parameter measurements from the Laurin view. METHODS: A total of 1431 consecutive Laurin views of the PFJ were retrospectively collected and divided into two parts: (1) the model development dataset (dataset 1, n = 1230) and (2) the hold-out test set (dataset 2, n = 201). Dataset 1 was used to develop the U-shaped fully convolutional network (U-Net) model to segment the landmarks of the PFJ. Based on the predicted landmarks, the PFJ parameters were calculated, including the sulcus angle (SA), congruence angle (CA), patellofemoral ratio (PFR), and lateral patellar tilt (LPT). Dataset 2 was used to assess the model performance. The mean of three radiologists who independently measured the PFJ parameters was defined as the reference standard. Model performance was assessed by the intraclass correlation coefficient (ICC), mean absolute difference (MAD), and root mean square (RMS) compared to the reference standard. Ninety-five percent limits of agreement (95% LoA) were calculated pairwise for each radiologist, reference standard, and model. RESULTS: Compared with the reference standard, U-Net showed good performance for predicting SA, CA, PFR, and LPT, with ICC = 0.85-0.97, MAD = 0.06-5.09, and RMS = 0.09-6.90 in the hold-out test set. Except for the PFR, the remaining parameters measured between the reference standard and the model were within the 95% LoA in the hold-out test dataset. CONCLUSIONS: The U-Net-based deep learning approach had a relatively high model performance in automatically measuring SA, CA, PFR, and LPT. KEY POINTS: • The U-Net model could be used to segment the landmarks of the PFJ and calculate the SA, CA, PFR, and LPT, which could be used to evaluate the patellar instability. • In the hold-out test, the automatic measurement model yielded comparable performance with reference standard. • The automatic measurement model could still accurately predict SA, CA, PFR, and LPT in patients with PI and/or PFOA.