FURI | Spring 2022

Using Batched Rays to Enable Higher Resolution Reconstruction of Computed Tomography Images

In our work, we tried to enable the reconstruction of high-resolution computed tomography (CT) images using implicit neural networks which usually require access to large amounts of graphics processing unit (GPU) memory. We propose a new ray-sampling technique that enables high-resolution reconstruction using INRs. Our approach uses a random batch of rays in each training iteration and therefore enables the computation to be done on a single GPU, and consequently democratizes the need for access to a big cluster of GPUs. In future work, the method can be expanded to 3D applications of computed tomography.