ScolioSIM project

This project is aimed to develop and evaluate deep learning models for predicting spinal curvature from 3D back surface scans as a radiation-free alternative to traditional X-ray methods. Two primary approaches were investigated: a Convolutional Neural Network using depth maps and a Point Cloud Transformer using 3D point clouds. The models were optimized and compared against optical data from standard medical system and gold standard data from X-ray measurements. Both models outperformed the current optical standard accuracy provided with the standard medical system, suggesting the potential applicability of the models as a supplement to current clinical methods. Extensive dataset variations were tested, including augmented data, balanced data, armless data, anatomical landmark points data, and combined data sets. These variations yielded mixed results in terms of accuracy improvement and transferability. Further research could explore improvements in data preprocessing and model architecture, the developed models already show promise in advancing the field of non-invasive scoliosis diagnosis and monitoring. The most significant limitations were the lack of sufficient gold standard data and the accuracy of the gold standard data. As technology continues to evolve, these models may contribute to reducing radiation exposure and improving the accuracy and efficiency of scoliosis assessments.