Medical Image Processing at Native Resolutions with Spatially Continuous Deep Learning Models

Medical images come in all shapes and sizes. They vary in resolution, noise, geometric distortion, and more, depending on scanner hardware, acquisition parameters, and patient pathology. This variability clashes with many deep learning models in medical image processing, which require resampling images out of their native resolution to a fixed size. However, this discards important information and moves images further away from what clinicians see in the real world. To address this shortcoming, we must stop thinking of medical images as just fixed grids of pixels and instead view them as (imperfect) samples of continuous physical anatomy. This dissertation presents my work on developing spatially continuous deep learning models for medical image processing, building upon previous implicit neural representations (INRs) and applying them to several medical imaging tasks. First, I will present my work on super-resolution in diffusion magnetic resonance imaging (dMRI) with FENRI (Fiber orientations from Explicit Neural RepresentatIons), a novel method for accurately reconstructing neural fibers in low-quality dMRIs. I then present MedIL (MEDical images from Implicit Latent spaces), a spatially continuous autoencoder for generating medical volumes at native resolutions. Finally, I present EPINR, an INR-based deformable image registration model for correcting geometric distortion in dMRIs without auxiliary images. Together, these models demonstrate how spatially continuous approaches allow deep learning models to see images as clinicians do.