Abstract
BACKGROUD: The use of deep learning algorithms in medical imaging has increased rapidly. This study aimed to develop an automated, hybrid approach combining a deep learning architecture and a conventional computer-aided diagnostic method to detect anatomical landmarks and measure radiological parameters in wrist radiography. METHODS: Overall, 487 wrist radiographs were randomly sampled for training and validation, and 100 radiographs collected from 2 institutions were used as the test set. Anatomical landmarks for 4 commonly used parameters, namely, radial inclination (RI), radial length (RL), volar tilt (VT), and ulnar variance (UV), were identified and labeled. A 2-step hybrid method combining a deep learning model with conventional computer-aided diagnosis was developed to measure radiological parameters using these anatomical landmarks. Measurements were obtained from the test set. The mean value of each parameter determined by 2 hand surgeons served as the reference standard. The performance of the deep learning algorithm was evaluated using the successful detection rate (SDR), mean absolute error (MAE), intraclass correlation coefficient (ICC), and Pearson correlation coefficient (r). RESULTS: The SDR of the model ranged from 97% to 98% at the 1 mm threshold and 99% at the 2 mm threshold. The RI, RL, VT, and UV generated by the model were 26.41° ± 4.01°, 12.63 ± 2.72 mm, 14.01° ± 8.02°, and 2.09 ± 2.76 mm, respectively. The overall MAEs for RI, RL, VT, and UV between the manually and automatically measured parameter values were 1.62° ± 1.26°, 1.56 ± 1.23 mm, 1.88° ± 1.68°, and 0.43 ± 0.41 mm, respectively. In the correlation analysis, good or high reliability was observed for RI, VT, and UV (ICC: 0.86, 0.95, and 0.98; r = 0.86, r = 0.95, and r = 0.98, respectively), and moderate reliability was observed for RL (ICC: 0.75, r = 0.78). CONCLUSIONS: This novel automated hybrid method can accurately identify landmarks on wrist radiographs and automatically generate the radiological parameters of the distal radius. This method saves time and reduces human labor in creating datasets for training segmentation models and developing image processing algorithms.