Neuromorphologicaly-preserving Volumetric data encoding using VQ-VAE

Petru-Daniel Tudosiu, Thomas Varsavsky, Richard Shaw, Mark Graham, Parashkev Nachev, Sebastien Ourselin, Carole H. Sudre, M. Jorge Cardoso

Abstract: Due to recent advancements in both hardware and software, Deep Learning applied to medical imaging has become feasible at higher resolutions. Even so, due to the sheer size of a single image the models and the convergence speeds are being hindered. Recently the Vector-Quantisation Variational Autoencoder shown promising results in generating realistic images while compressing them to ~2% of their original size. Here, we show that a VQ-VAE inspired network can be used to compress the data to ~3% while maintaining reconstructed images that adhere to the same morphological and tissue statistics as the original data. Furthermore, we show that one can use one of our models that was trained on widely available neurologically healthy patients and fine-tune on pathological ones, thus allowing faster training times.

Paper Type: both

TL;DR: We show that VQ-VAE can reconstruct morphologically-correct volumetric imaging data at full resolution, while compressing it to <1% of the original size.

Track: full conference paper

Keywords: 3D, MRI, Morphology, Encoding, VQ-VAE

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/neuromorphologicaly-preserving-volumetric/code)