Deep learning convolutional neural networks with three dimensional depth sensor imaging in scoliosis detection --An algorithm to reduce the detection of false porisive cases and mild cases--.

Introduction: Adolescent idiopathic scoliosis (AIS) is the most ordinary pediatric spinal disease. Timely intervention in growing individuals, such as brace treatment, relies on early detection of AIS. We developed a system consisting of a 3D depth sensor and an algorithm installed in a laptop computer. In this system, the correlation between the actual Cobb angle and the predicted Cobb angle calculated from the asymmetry index was 0.85 (P < 0.01). The purpose of this study is to create a deep learning algorithm (DLA) to identify moderate or severe AIS patients requiring the secondary screening using data of subjects detected in the school screening. Materials and Methods: We included 334 subjects detected using the 3D depth sensor system in school screening. The 3D images from the 3D depth sensor system were used as input data for the DLA with Convolutional neural networks. We randomly separated the 334 subjects into an internal validation data of 250 and an external validation data of 84. Binary classification was performed as 0 for images with Cobb angle of < 12° and 1 for images with Cobb angle of ≥ 12° based on the average actual Cobb angle of 12.0°. Five-fold cross validation was conducted to evaluate the probability for Cobb angle of ≥ 12°. The minimum predicted probability in subjects with Cobb angle of ≥ 15° was configured as the cut-off value to detect the second screening targets. In the external validation, 84 images were evaluated utilizing trained DLA in the internal validation, and decide to require secondary screening, based on the cu-off value. Results: In internal validation, the five-fold cross validation showed that the dataset 3 had the highest predicted performance. The minimum predicted probability in subjects with Cobb angle of ≥ 15° was 0.47 in dataset 3. In the external validation, the number of subjects with Cobb angle of < 10° and < 15° were 36 and 62, respectively. Based on a cut-off value of 0.47, 39 (63%) subjects with Cobb angle of < 15° were judged as unnecessary for the second screening. There was only one false negative case with Cobb angle of 19°. Conclusions: This DLA reduced the number of extremely mild AIS patient and false positive cases in the external validation, indicating that this DLA can reduce the unnecessary medical care expenditures and the unnecessary radiation exposure for children and adolescents. [ABSTRACT FROM AUTHOR]