Our Final Year Project is proposed by Department of Computer Science and Department of Orthopedics at the University of Hong Kong, and the final goal of this project is to automate the reconstruction of a 3d spine model from patient’s X-ray images. We are going to utilize computer vision and deep learning to model the spine.
1. Background
Due to the many types of spinal deformity and the diversity of human bodies, the shape and alignment of spines can vary in many ways and thus it requires doctors to treat patients on a case-by-case basis. The most efficient way is to construct a 3D spine model which will help doctors to fully comprehend the structure and the shape of a patient’s spine before making a treatment plan.
However, the 3D construction procedure has always been tricky because currently there are not many assistant tools available. The best tool so far we have learnt from the Department of Orthopedics at the University of Hong Kong is EOS which builds the model based on a standard spine of a French teenager. This tool works as follows: to begin with, patients’ X-ray images which are taken from the front view and the lateral view simultaneously are input into EOS. Then EOS will generate a rough match of the images and the model. In the end, it requires doctors to manually adjust to make the model fit better. However, EOS causes issues in clinical practice. Firstly, it requires tedious manual adjustments on each vertebra from both points of views to fit the 3D model with the X-ray images, which could take 30 minutes or longer. Secondly, the manual adjustment introduces error up to 20 degrees, which is intolerable in medical diagnosis, especially when a patient is diagnosed with scoliosis (a type of abnormal spinal alignment) when the lateral curvature is larger than 10 degrees. Thirdly, the French teenager model is not very appropriate for diagnosis and treatment in local hospitals, where Asian patients form the majority. To address the inefficiencies and inaccuracies with EOS, we propose a new tool that will automate the whole process of diagnosis which will save the doctor’s time and improve on the accuracy.
2. Related studies
This project is supported by the AI Spine team of Department of Orthopedics and the Computer Vision team of Department of Computer Science. They have both made some major progress in locating and segmenting the spine on X-ray images with decent accuracy and their work will serve as a foundation of our project. AI Spine team will provide images for our later training and testing purposes as well.