This PR enables building MONAI models from the config with meddlr.modeling.build_model

Example Usage

MONAI V-Net model

config.yaml

MODEL:
  MONAI:
    VNet:
      in_channels: 1
      out_channels: 1
      spatial_dims: 2
      dropout_dim: 2

Edge k-space is defined as the k-space along the edges of the matrix that is not sampled. An example of edge k-space for the Poisson Disc undersampling pattern with crop_corners=True is shown below

We often want to know the edge k-space for certain methods (SSDU, hard data consistency postprocessing, etc.). This PR provides utilities for dynamically computing the edge k-space based on the type of undersampling pattern used.

I encountered an exception when running import meddlr.ops as oF in an NGC PyTorch container.

The exception says,

>>> import meddlr.ops as oF
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/workspace/meddlr/meddlr/__init__.py", line 3, in <module>
    from meddlr.engine.model_zoo import get_model_from_zoo
  File "/workspace/meddlr/meddlr/engine/__init__.py", line 10, in <module>
    from .hooks import *  # noqa
  File "/workspace/meddlr/meddlr/engine/hooks.py", line 19, in <module>
    from meddlr.evaluation.testing import flatten_results_dict
  File "/workspace/meddlr/meddlr/evaluation/__init__.py", line 2, in <module>
    from .recon_evaluation import ReconEvaluator  # noqa: F401
  File "/workspace/meddlr/meddlr/evaluation/recon_evaluation.py", line 16, in <module>
    from meddlr.data.transforms.transform import build_normalizer
  File "/workspace/meddlr/meddlr/data/__init__.py", line 1, in <module>
    from meddlr.data import build, catalog, collate, data_utils, slice_dataset
  File "/workspace/meddlr/meddlr/data/build.py", line 13, in <module>
    from meddlr.data.slice_dataset import SliceData
  File "/workspace/meddlr/meddlr/data/slice_dataset.py", line 10, in <module>
    from meddlr.data.transforms.transform import DataTransform
  File "/workspace/meddlr/meddlr/data/transforms/transform.py", line 9, in <module>
    from meddlr.forward import SenseModel
  File "/workspace/meddlr/meddlr/forward/__init__.py", line 1, in <module>
    from meddlr.forward import mri
  File "/workspace/meddlr/meddlr/forward/mri.py", line 4, in <module>
    import meddlr.ops as oF
  File "/workspace/meddlr/meddlr/ops/__init__.py", line 1, in <module>
    from meddlr.ops import categorical, complex, fft, utils  # noqa: F401
  File "/workspace/meddlr/meddlr/ops/fft.py", line 10, in <module>
    if env.pt_version() >= [1, 6]:
  File "/workspace/meddlr/meddlr/utils/env.py", line 314, in pt_version
    version = [dtype(x) for x in version]
  File "/workspace/meddlr/meddlr/utils/env.py", line 314, in <listcomp>
    version = [dtype(x) for x in version]
ValueError: invalid literal for int() with base 10: '0a0'

I checked the related lines, the code assumes the PyTorch is using stable versions and installed from binary packs.

https://github.com/ad12/meddlr/blob/b2f055b482c534a2a6eb161c19e6abb6ea94c196/meddlr/utils/env.py#L314

However, the PyTorch in the NGC container is self-complied, and the version string contains a commit ID. It looks like this,

>>> import torch
>>> torch.__version__
'1.13.0a0+08820cb'

I suspect that's why it triggered the exception.

Would you please kindly modify the function to accept the version strings like this?

This PR refactors SimpleConvBlockNd block and GeneralizedUNet to use the LayerInfo abstraction.

NOTE: there is one implementation breaking PR. GeneralizedUNet's upsampling block will order the norm/activation based on the order in the original block. Previously the order was always convtranspose -> act -> norm. Order of act and norm will be determined by which comes first in the block_order

• deprecate 3.6 support in CI - seeing C407 errors in Py3.6 linting which are not supported in later versions
• tifffile<=2022.5.4 - tifffile==2022.7.28 causing issues with skimage data loading during testing

LayerInfo is a dataclass that allows for more modular control of the layers being built. This class provides two key functions:

• format: Creates a LayerInfo object from raw python types which are typically read from files (e.g. yaml/json)
• build: Builds the layer given the layer's init_kwargs

FAILED tests/utils/test_path.py::test_gdrive_handler - ImportError: cannot import name 'client' from 'gdown.download_folder' (/opt/hostedtoolcache/Python/3.7.15/x64/lib/python3.7/site-packages/gdown/download_folder.py)

Hello, I am new to MRI reconstruction and plan to learn to use fastMRI with singlecoil dataset and using meddlr to reproduce the VORTEX work. I just read the "VORTEX" paper and I think it was a great work. However, I go through the code and found most functions, such as meddlr.utils.transforms or datasets/format_fastmri.py is mainly used for multicoil data. Could you please consider adding singlecoil support?

Lookingforward to your reply. Thank you very much!

****On attempting to format mri data the scan 'b7d435a1-2421-48d2-946c-d1b3372f7c60.h5' is not found in the directory 'datasets/mridata_knee_2019/raw/ismrmrd/' and the issue is reported as 'file signature not found'.

Medical images have a heterogenous number of dimensions. For example, MRI volumes can be treated as 2D slices, 3D volumes, or nD tensors (e.g. multi-echo, time, etc.). It is important to have built-in modules be able to handle tensors of arbitrary dimensions with some (but little) opinion about the behavior of the model across these different dimensions.

Proposal

Currently, frameworks require the user to keep track of the order of dimensions for tensors. We may be able to simplify this by having a tensor dim_order value that is managed and manipulated by the meddlr ecosystem.

TODOs

• [ ] Update forward.mri.SenseModel as a sample use case for managing multiple dimensions

Calling sigpy.mri.poisson hangs during the data loading process, but copying the function seems to be running fine.

Things to check:

• [ ] Implementation differences between current sigpy implementation and internal copy
• [ ] Speed difference / hang time

Noise2Recon:

• [ ] Add experimental configs and pre-trained weights to google drive
• [ ] Colab tutorial for running experiments

VORTEX:

• [ ] Add experimental configs and pre-trained weights to google drive
• [ ] Colab tutorial for running experiments