Pyraug Case Study 1

Use of Pyraug software to perform MRI classification

Case Study 1: Classification on 3D MRI (ADNI & AIBL)

Introduction

A Riemannian Hamiltonian VAE model and Pyraug software were used to perform Data Augmentation in the High Dimensional Low Sample Size Setting on 3D MRI neuroimaging data from ADNI database. The model was used to try to enhance the classification task consisting in finding Alzheimer's disease patients (AD) from Cognitively Normal participants (CN) using T1-weighted MR images . such augmentation.

Classification set up

Data Splitting

The ADNI data set was split into 3 sets: train, validation and test. First, the test set was created using 100 randomly chosen participants for each diagnostic label (i.e. 100 CN, 100 AD). The rest of the data set was split such that 80% is allocated from training and 20% for validation. The authors ensured that age, sex and site distributions between the three sets were not significantly different. The train set is referred to as train-full in the following. In addition, a smaller training set (denoted as train-50) was extracted from train-full. This set comprised only 50 images per diagnostic label, instead of 243 CN and 210 AD for train-full. It was ensured that age and sex distributions between train-50 and train-full were not significantly different. This was not done for the site distribution as there are more than 50 sites in the ADNI data set (so they could not all be represented in this smaller training set). The AIBL data was never used for training or hyperparameter tuning and was only used as an independent test set.

Data Split for the classification task: Alzheimer Disease (AD) vs. Cognitively Normal (CN)

Data Processing

All the data was processed as follows:

  • Raw data are converted to the BIDS standard .
  • Bias field correction is applied using N4ITK .
  • T1w images are linearly registered to the MNI standard space with ANTS and cropped. This produced images of size 169x208x179 with 1mm3 isotropic voxels.
  • An automatic quality check is performed using an open-source pretrained network . All images passed the quality check.
  • NIfTI files are converted to tensor format.
  • (Optional) Images are down-sampled using a trilinear interpolation, leading to an image size of 84x104x89.
  • Intensity rescaling between the minimum and maximum values of each image is performed.

    • Classifier

    To perform such classification task a CNN was used with two different paradigms to choose the architecture. First, the authors reused the same architecture as in which was obtained by optimizing manually the networks on the ADNI data set for the same task (AD vs CN). A slight adaption was done for the down-sampled images, which consisted in resizing the number of nodes in the fully-connected layers to keep the same ratio between the input and output feature maps in all layers. This architecture is denoted baseline. Secondly, a random search was launched allowing to explore different hyperperameter values. The hyperparameters explored for the architecture were the number of convolutional blocks, of filters in the first layer and of convolutional layers in a block, the number of fully-connected layers and the dropout rate. Other hyperparameters such as the learning rate and the weight decay were also part of the search. 100 different random architectures were trained on the 5-fold cross-validation done on train-full. For each input, the selected architecture is the one that obtained the best mean balanced accuracy across the validation sets of the cross-validation. This architecture is referred to as optimized.

    CNN architectures: left: The baseline net. right: The optimized one using a random search across100 architectures.

    Augmentation Set up

    On the meantime, a RHVAE was trained on each class of the train sets (train-50 or train-full) to be able to generate new synthetic data. Noteworthy is the fact that the VAE and the CNN shared the same training set and no augmentation was performed on the validation set or the test set.

    Data Augmentation scheme with a VAE.

    Then the baseline (resp. optimized) CNN networks were then trained for 100 (resp. 50) epochs using the cross entropy loss for training and validation losses. Balanced accuracy was also computed at the end of each epoch. The models were trained on either 1) only the real images; 2) only the synthetic samples created by the RHVAE or 3) the augmented training set (real + synthetic) on 20 independent runs for each experiment. The final model was chosen as the one that obtained the highest validation balanced accuracy during training.

    Results

    Below are presented some of the main results obtained in this case study. We refer the reader to for the full results of the study.

    Augmentation results with the baseline CNN network.
    Augmentation results with the optimized CNN network.

    Pyraug software allowed for a significant gain in the model classification results even when the CNN was optimized on the real data only (random search not performed for augmented data set) and even though small size data sets were considered along with very challenging high dimensional data.