UC Santa Cruz

Medical image representation has long been dominated by voxel-grid matrices. While their inherent structure and order work efficiently for various linear transformations and provide a seamless visualization method on monitors, they fail to preserve the topology of the scan and to encode sparse information in a memory-efficient way.The recent emergence of machine learning-based continuous coordinate-based scene representations such as neural radiance fields and Gaussian splatting has provided alternative representation techniques. These approaches overfit the weights of a model by iterative differentiable rendering and have been shown to be more compact than grid representations. Moreover, they are then able to perform novel view synthesis from any given camera pose.Off-grid representations translate directly to Cone Beam Computed Tomography sparse-view acquisitions, where streaking and quantum noise artifacts are dominant. Using differentiable rendering, a continuous representation is achieved, with interpolation providing a path to recover some of the lost signal.In this dissertation, we apply a variety of methodologies, including Gaussian splatting, implicit occupancy fields, and Neural Attenuation Fields regularized with an anatomic prior, to Cone Beam Computed Tomography reconstruction, and evaluate their performance across a range of anatomic datasets. Our models show that learned gridless representations achieve substantial memory reduction, recover signal under extreme view sparsity, and preserve scene topology.