42. An algorithm for using deep learning convolutional neural networks with three-dimensional depth sensor imaging in scoliosis detection in order to avoid the detection of extremely mild patients and false positive cases

Adolescent idiopathic scoliosis (AIS) is one of the most frequent pediatric spinal diseases. Early detection of AIS is essential because timely intervention, such as brace treatment, is needed for growing individuals with moderate AIS. We developed a device for the detection of AIS using a three-dimensional (3D) depth sensor and an algorithm installed on a laptop computer. This device calculates the asymmetry index, which had a correlation with the actual Cobb angle of 0.85 (P < 0.01). However, unnecessary radiation exposure can result from detecting patients with Cobb angle of 10˚ to 15˚in school screenings, because they do not require brace treatment. Furthermore, additional examinations for false positive cases and extremely mild AIS patients create a burden on medical resources. This study aimed to create a deep learning algorithm (DLA) to identify moderate or severe AIS patients requiring secondary screening using data from subjects detected in school screening. Retrospectively study Data from 334 subjects detected using the 3D depth sensor system in school screenings from April 2021 to March 2023 were obtained. The area under the curve (AUC) derived from the receiver operating characteristic (ROC) curve, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, positive likelihood ratio (PLR), and negative likelihood ratio (NLR) were calculated. The DLA with convolutional neural networks was constructed, utilizing the 3D depth sensor images as input data. Binary classification was used as the output layer, where images with Cobb angles of < 12˚ were labeled as 0 and images with Cobb angles of ≥ 12˚ were labled as 1 based on the mean actual Cobb angle of 12.0 ˚to prevent an imbalanced dataset in internal validation. The 3D images of 334 subjects were randomly divided into an internal validation dataset with 250 images and an external validation dataset with 84 images. Internal validation was performed to evaluate the probability for Cobb angles of ≥ 12˚ using five-fold cross-validation. To identify the second screening targets, the minimum predicted probability in subjects with Cobb angle of ≥ 15˚ (MPP15) was defined as the cut-off value. In the external validation, 84 images were assessed with the DLA trained in the internal validation and decisions for requiring secondary screening were made based on MPP15. The range of Cobb angles was 0˚ to 34˚ and 0˚ to 32˚ in the internal and external validation, respec