7. An algorithm for using deep learning convolutional neural networks with 3-dimensional depth sensor imaging in scoliosis detection

Abstract

<h3>BACKGROUND CONTEXT</h3> Timely intervention in growing individuals, such as brace treatment, relies on early detection of adolescent idiopathic scoliosis (AIS). To this end, several screening methods have been implemented. However, these methods have limitations in predicting the Cobb angle. <h3>PURPOSE</h3> This study aimed to evaluate the performance of a 3-dimensional depth sensor imaging system with a deep learning algorithm (DLA), in predicting the Cobb angle in AIS. <h3>STUDY DESIGN/SETTING</h3> Retrospective analysis of prospectively collected, consecutive, nonrandomized series of patients at 5 scoliosis centers in Japan. <h3>PATIENT SAMPLE</h3> One hundred sixty human subjects suspected to have AIS were included. <h3>OUTCOME MEASURES</h3> Patient demographics, radiographic measurements, and predicted Cobb angle derived from the DLA were the outcome measures for this study. <h3>METHODS</h3> One hundred sixty data files were shuffled into 5 datasets with 32 data files at random (dataset 1, 2, 3, 4 and 5) and 5-fold cross validation was performed. The relationships between the actual and predicted Cobb angles were calculated using Pearson's correlation coefficient analyses. The prediction performances of the network models were evaluated using mean absolute error (MAE) and root mean square error (RMSE) between the actual and predicted Cobb angles. The shuffling into five datasets and five-fold cross validation was conducted ten times. There were no study-specific biases related to conflicts of interest. <h3>RESULTS</h3> The mean age was 14.7 ± 2.4 years, and the average Cobb angle was 30° (range: 0˚ to 64˚). The range of correlation coefficients was 0.87 to 0.89 in the five-fold cross validation with 10 repeats. The correlation between the actual Cobb angle and the mean predicted Cobb angle with 10 repeats was 0.91. The range of MAEs was 4.4˚ to 4.7˚in the 5-fold cross validation with 10 repeats. The range of RMSEs was 5.8˚ to 6.3˚. The MAE and RMSE between the actual Cobb angle and the mean predicted Cobb angle with 10 repeats was 4.0˚ and 5.4˚ respectively. At a Cobb angle of 10˚, which confirms a diagnosis of scoliosis, the mean predicted Cobb angle with 10 repeats showed a sensitivity of 0.99, a specificity of 0.42, PPV of 0.95, NPV of 0.71, accuracy of 0.94, PLR of 1.69, and an NLR of 0.03. <h3>CONCLUSIONS</h3> The 3D depth sensor imaging system with its newly innovated CNN for regression is objective and has significant ability to predict the Cobb angle in children and adolescents. This system is expected to be used for screening scoliosis in clinics or physical examination at schools. In addition, this is also a possible surrogate for radiographs to monitor curve progression, preventing unnecessary X-rays for mild case of scoliosis. <h3>FDA DEVICE/DRUG STATUS</h3> This abstract does not discuss or include any applicable devices or drugs.