Deep Learning-based Reconstruction of Volumetric CT Images of Vertebrae from a Single View X-Ray Image

Abstract

Computed tomography is often used in medical fields today because it creates more detailed information for doctors than regular X-ray images. However, one major side effect is that patients may be exposed to a large dose of radiation because it takes hundreds of X-ray images to compute a CT scan. Another shortcoming is that patients are required to lay down on the CT machine for the scan, which is usually not the ideal position when diagnosing spine related issues such as cervical spondylosis and lumbar disc herniation. The prime motivation for this study is to reconstruct CT images using only one or a few X-ray images by using deep learning models trained to map projection radiographs to the corresponding 3D anatomy. My work demonstrates the feasibility of the approach with 20 Dicom sets of human vertebrae. The training set of the deep learning model consists of pairs of information, where each pair is made up of a 3D volume and a manually generated radiograph. The deep learning model for this study is CNN (Convolutional Neural Network) based encoder-decoder framework. The encoder converts high-dimensional data into embedded feature maps whereas the decoder reconstructs high-dimensional 3D output we desire. After training, the network can take in single or multiple 2D x-ray images and output an array of intensity values that represent a 3D CT image. MATLAB 3D viewer is used to visualize the result. We performed 50 experiments, averaging 3 model training for each experiment. The results generated by the model have an acceptable accuracy but there is a lot of room for improvement. The best PSNR (Peak Signal-to-Noise Ratio) value we obtain is 17.34 dB. While a state-of-the-art 3D reconstruction usually has a PSNR value above 30 dB. In addition, this paper summarizes the challenges and limitations that my teammates and I faced. I will also introduce methods that the team used to overcome these barriers. Since this is still an ongoing research project, the team will continue the work on improving the result. The end goal is to apply this study on real medical cases.