Adolescent idiopathic scoliosis (AIS) is an extremely common three-dimensional (3-D) deformity of the spine, affecting the population between 10 and 18 years of age. Early detection of AIS is critical, as the earlier that the spinal deformity can be identified, the more likely it is that curve progression can be minimized or arrested using conservative treatment options. However, today much of the responsibility for detecting AIS falls on untrained parents. Therefore, the purpose of this thesis project is to use images taken with a smartphone containing a depth sensor to create a simple and effective machine learning (ML) algorithm that can detect the absence or presence of scoliosis. Secondarily, this thesis project aims to 1) provide a proof of concept for a regression-based ML algorithm that can predict the main curvature of the scoliotic spine and 2) to determine if the depth information from the smartphone contains additional features or information that can improve the performance of the ML algorithm when compared to regular red-green-blue (RGB) images. Thirty-three participants (28 AIS; 5 Control) were recruited from the Children’s Hospital of Eastern Ontario (CHEO). Images of the unclothed backs of participants were taken with a smartphone (Samsung S20 Ultra 5G) containing a depth sensor, while participants assumed two positions: an upright standing posterior-anterior (PA; mirroring the position participants are in when getting an EOS scan) and a forward bending position. A convolutional-neural-network (CNN)-backed decision tree algorithm was developed and trained using three different data streams: a red-green-blue-depth (RGB-D), a Colourized depth map, and an RGB data stream. It was determined that the model trained with the Colourized forward bending images had the highest overall accuracy. The CNN backed decision tree was able to classify images of participants in a forward bend posture with an accuracy of 93%, specificity of 75%, and a sensitivity of 99%. Additionally, it was found that all algorithms trained with the varying data streams were able to predict the Cobb angle of the spine within 16° of the ground truth Cobb angles. The lowest root mean square error (RMSE) values were obtained from the RGB images when the participants were in the PA position. The PA RGB dataset had RMSE values of 7.17° between the ground truth and predicted Cobb angles. Inter-rater reliability errors typically range between 5-7° for manually measured Cobb angles. Therefore, given the calculated RMSE for the PA RGB dataset were close to this range, there is the potential to use this smartphone technology to screen for scoliosis and predict the curvature of the spine (Morrison et al., 2015). While these results are promising, the dataset is small compared to other studies; therefore, this thesis provides a proof of concept, and more work needs to be done to increase the robustness of the model and to further improve the ability of the model to predict the Cobb angle of the spine.
| Identifer | oai:union.ndltd.org:uottawa.ca/oai:ruor.uottawa.ca:10393/44069 |
| Date | 19 September 2022 |
| Creators | Wenghofer, Jessica |
| Contributors | Graham, Ryan |
| Publisher | Université d'Ottawa / University of Ottawa |
| Source Sets | Université d’Ottawa |
| Language | English |
| Detected Language | English |
| Type | Thesis |
| Format | application/pdf |
Page generated in 0.0018 seconds