Abstract | Medical imaging data is typically 3D, causing scan sizes and databases to grow cubically with resolution, unlike the quadratic growth in standard computer vision tasks. Compressing scan dimensionality is essential for deep learning, as raw data often exceeds GPU memory limits. Autoencoders are commonly used for data-specific non-linear compression, balancing compactness and fidelity. However, they are limited to the resolution of the training data. Inspired by Neural Fields, we propose an autoencoder with a fully-connected network as its decoder, and train it on the UK Biobank abdominal MRI dataset. Beyond more fidelity in the reconstruction, our encoding is a continuous function of 3D coordinates rather than 3D rasters like the original data, which enables our architecture to be utilized in a variety of applications such as super-resolution, in-painting and extrapolation. We show that this change of paradigm in representation leads to higher and better compression, with better properties, and enables the use of such imaging databases for deep learning in their compressed state. |
---|