Ivadomed: Adaptative Loader

[AnBucquet]

To reproduce the Pseudo-Curriculum Training Procedure from the HeMIS paper, we need a loader that can change its distribution across time. To do so, I came up with the idea of adding a Dataframe in the Dataset class. (see example below)

Why

• Dataframes are easy to manipulate and perfectly supported by the loader of Pytorch.
• No need to recreate a Dataset every epoch (which is very long) - We can just re-initialize the loader. (shouldn't be very long)
• Add new features

New Features

• files can be loaded/ partially-loaded (later) into RAM or read "on the fly"
• Dataframe can be saved and used later
• possibility to load only the slices we want
• Allows curriculum learning procedures (including HeMIS)
• Columns can be added to create to deal with new strategies such as active learning (_Difficulty_)

WIP

• Create a class Dataframe - need to be more modulable
• Create new loader that uses that class

TODO

• change Medical-torch such that files are not necessarily loaded into RAM
• implement strategies

Questions

• This loader seems to be very specific to our problem, but maybe it should be added directly to MedicalTorch?
• Any suggestions/ problems?

_Example of Dataframe_

[charleygros]

Like really like your ideas @AnBucquet !

As said during meeting:

• start by implementing the new loader within ivadomed repo'. If too much duplication with medicaltorch, then we will consider of moving it there.
• use HDF5 file instead of files can be loaded/ partially-loaded (later) into RAM or read "on the fly"

[AnBucquet]

About the use of HDF5 file:

I'm currently working on a function to generate an HDF5 file from a Bids dataset (SpineGeneric in my case). After discussing with @charleygros, we agree to use the following architecture in the HDF5 file:

The idea is to create a GROUP for each patient. Each of this GROUP has at most 3 sub-groups:

• Inputs: contains the different modalities as Datasets
• gt: idem
• roi idem

As those datasets are like _Numpy array_, we would be able to read only some slices (to replace the slice filter).

[jcohenadad]

looks good! could you detail what would be in each of the metadata block?

[AnBucquet]

Yes! I'm currently working on that. I will keep you informed!

[AnBucquet]

About the metadata block, I've started listing metadata in each block:

• HDF5 block:

  • list of patients
  • Canonical reordering of the volume axes or not
  • slice axis
  • slice filter
  • metadata choice

• Group block (Patient):

  • Different modalities
  • Useful slices (i.e. slices with GT)

• Sub-GROUP (inputs, gt, roi):

  • datasets inside (i.e.: different modailities, gt, ...)

• Dataset (MRI and labels):

  • all parameters in JSON file (_FlipAngle_, _TE_, _TR_, ... )

This list is not fixed for now.