adrianarnaiz / brain-mri-autoencoder

Deep Convolutional Autoencoders for reconstructing magnetic resonance images of the healthy brain

ELLIS PhD student in Alorithmic Fairness and Graph Neural Networks @ ELLIS Alicante

• Personal Web Page
• Google Scholar

brain-mri-autoencoder's Issues

Keras Data Loader Development

Create a Data Loader which loads images into model. It could make some preprocessing or data aumentation.

Posibilities:

• ImageDataGenerator I
  • How to use it (Towards Data Science)
• ImageDataGenerator II
• Data Generator

Make PPT slides

could you share checkpoints with us?

Improve loaders, architecture and environment to be faster

• Cloud

  • Colab
  • Kaggle GPU Notebooks
  • Downside: How do I load all the 25 GB npy files for training?

• On-Premise

  • Improve workflow: keras and tf libraries to improve pipeline
  • Get optimize NVIDIA configuration
  • Downside: my pc and its capability

• Alpha

  • Better Hardware
  • Downside: Configure all project on tmux and data load.

Post-Hoc Analysis

Set up environment

Set up environment

Install CUDA, cuDNN and tensorflow-gpu. Must be compatible.

• CUDA 10.1
• cuDNN 7.6.5 for CUDA 10.1
• Tensorflow-gpu 2.1 or 2.3.1

Consider cloud computing options for further stages: Google Colaboratory, Kaggle, Alpha, etc...

Experiments without data augmentation

We create 3 base models without data augmentation.

• FCN + skip connections: Unet
• Resnet-like Autoencoder
• Resnet + Skip connections

Stratified Train Test splitting

We will Explore the metadata of the volumes and stratify them by Age, Sex and Ethnic.

Select Autoencoder architectures to compare

We should select a set of architectures in order to make the experiments and compare them.

• Select layers and activation function:
  • original
  • resnet-like
  • unet-like
  • dense-like
  • VGG-like
  • ...
• Transfer learning
  • Frozen weights
  • Initial weights
• Select normalization techniques:
  • L1, L2
  • Dropout
  • Batch Normalization
  • Group Normalization
  • Early-sttopping
  • ...

Experiments with DataAugmentation

• Archs:

  • Shallow Res-encoder
  • Myronenko encoder
  • CAE+skip-con
  • ¿? Res+Skip conn

• Loss

  • MSE
  • dSSIM

Select Loss functions to compare

Select loss functions to be compared for each architecture: i.e. MSE, cross-covariance, dssim, KL divergenci (if VAE), cross-entropy (if semi-supervised), etc

Research state of art

Make the first steps in state of the art research

• Main dataset: IXI Dataset: research about other projects using this dataset.
• Main papers:
  • Review
  • [Pinaya et al, 2018]: Identify abnormal structures with an autoencoder (semi-supervised).
  • 4 papers from Review:
    • [Gondara et al, 2016]: Denoising autoencoder.
    • [Bermudez et al, 2018]: Autoencoder with skip connections.
    • [XJ Mao et al, 2016]: code available, DCA with skip connections.
    • [Chauhan et al, 2019]: DCA with fuzzy logic.
• More Recent papers like:
  • Jose V Manjon, 2020: Brain Lesion Inpainting with DL
  • Search more
• GAN-based
• How to deal with MRI images and first approach to Deep Convolutional Autoencoder

Finish documentación

Data Exploration - Relevant Slice Selection

• Select view of volume slices: sagittal, axial or coronal.
• Select slices with relevant information: n-fixed, outer software or OpenCV code.