Abstract
Translational brain research using Magnetic Resonance Imaging (MRI) is becoming increasingly popular as animal models are an essential part of scientific studies and ultra-high-field scanners become more available. Some drawbacks of MRI are MRI scanner availability, and the time needed to perform a full scanning session (it usually takes over 30 minutes). Data protection laws and 3R ethical rule also make it difficult to create large data sets for training Deep Learning models. Generative Adversarial Networks (GAN) have been shown capable of performing data augmentation with higher quality than other techniques. In this work, the alpha-GAN architecture is used to test its ability to generate realistic 3D MRI scans of the rat brain. As far as the authors are aware, this is the first time an approach based on GANs is used for data augmentation in preclinical data. The generated scans are evaluated using various qualitative and quantitative metrics. A Turing test performed by 4 experts has shown that the generated scans can trick almost any expert. The generated scans were also used to evaluate their impact on the performance of an existing deep learning model developed for rat brain segmentation of white matter, grey matter, and cerebrospinal fluid. The models were compared using the Dice score. The best results for the segmentation of whole brain and white matter were achieved when 174 real scans and 348 synthetic ones were used, with improvements of 0.0172 and 0.0129. The use of 174 real scans and 87 synthetic ones led to improvements of 0.0038 and 0.0764 of grey matter and cerebrospinal fluid segmentation. Thus, by using the proposed new normalisation layer and loss functions, it was possible to improve the realism of the generated rat MRI scans and it was demonstrated that using the data generated improved the segmentation model more than using conventional data augmentation.
Abstract (translated)
URL
https://arxiv.org/abs/2112.13626