Reconstructing CT Volumes from Single and Multiplanar Chest Radiographs Using Deep Learning and Synthetic X-Ray Projections

Reconstructing CT volumes in 3D using convolutional encoder-decoder representation frameworks and generative adversarial networks. Initial results can be accessed in the 'IPYNB' folder via Google Colaboratory or JupyterLab. Training is done on a single Tesla V100 GPU hosted on Amazon Web Services (AWS).

Synthetic X-Ray Dataset

Beer's Law and ray-tracing via the Siddon-Jacobs algorithm are implemented in MATLAB to generate realistic chest radiographs from a CT input.

CT Input	Radiograph Projections

Single View Encoder-Decoder Results

MSE: 0.0011900706052539642 (avg) Lung DICE: 0.9411571828756816 (avg) SSIM: 0.8068173427439007 (avg)

Double View Encoder-Decoder Results

MSE: 0.000952999815192659 (avg) Lung DICE: 0.971042315 (avg) SSIM: 0.83141169669 (avg)