Room: AAPM ePoster Library
Many non-rotating x-ray computed tomography (CT) frameworks are being developed, with discrete sources, including a system previously reported by our group (Cramer et al AAPM 2018, 2019; Cramer et al Sci Rep 2018). For these discrete-source systems, the number of x-ray sources is directly linked to the number of projection angles.
If a CT image could be reconstructed with fewer projections, it could mean not only less dose to the patient, but also ease the engineering constraints as each individual source can occupy a larger arc angle.
We adapted AUTOMAP (AUTOmated transform by Manifold APproximation) for CT image reconstruction. AUTOMAP is an end-to-end generalized reconstruction framework, implemented with a deep neural network architecture composed of two fully connected layers followed by a sparse convolutional autoencoder. We have previously shown its utility in reconstructing MRI data, particularly in images with missing data (Zhu et al, Nature, 2018).
We trained the network on a large database of head CT images from an RNSA stroke detection grand challenge dataset. For each image in the dataset, their corresponding sinograms were generated using the forward Radon transform.
We demonstrated a tenfold reduction of reconstruction error over the conventional filtered back projection (FBP) algorithm in downsampled CT head images with 56 projections. This reconstruction approach was also validated on generic IMAGENET and MNIST datasets (with appropriate training data). We discuss how this image reconstruction approach can be accomplished on specialized CPUs and large model support capable GPU clusters for highly-resolved CT images.
We have developed a machine learning (ML) algorithm that can improve image quality for inverse Radon reconstruction problems with fewer projections. We discuss two distinct computing hardware frameworks for addressing the memory requirements of ML architectures with fully connected layers for the reconstruction of images with high resolution.