Contour detection in synthetic bi-planar X-ray images of the scapula: Towards improved 3D reconstruction using deep learning

Three-dimensional (3D) reconstruction of two-dimensional (2D) X-ray images using statistical shape models provides a cost-effective way of increasing diagnostic X-ray utility, especially in low-resource settings. Feature-constrained model fitting is one way to obtain patient-specific models from a statistical model. This approach requires an accurate selection of corresponding features, usually landmarks, from the biplanar X-ray images. However, super-positioned structures in 2DX-ray images confound this approach. This paper reports on the use of a deep learning algorithm to detect the contour of the scapula in synthetic bi-planar X-ray images to address this limitation. Two independent U-net models were trained to learn the mapping between synthetic bi-planar X-ray images of the scapula. The first model was trained to learn the mapping of the scapula contour in the lateral image view given the anterior-posterior image view and the second to learn the mapping of the scapula contour in the anterior-posterior image view given the lateral image view. The predicted images, when compared to the ground-truth, gave Dice coefficient values of 0.93 and 0.96 for the first and second models, respectively. These results are comparable to those in literature. However, these trained models did not generalize to real data, suggesting a need for further training with manually segmented images.