Implements vSHARP model (2D)

README.rst

@@ -25,8 +25,11 @@
 DIRECT: Deep Image REConstruction Toolkit
 =========================================
 
-``DIRECT`` is a Python, end-to-end pipeline for solving Inverse Problems emerging in Imaging Processing. It is built with PyTorch and stores state-of-the-art Deep Learning imaging inverse problem solvers such as denoising, dealiasing and reconstruction. By defining a base forward linear or non-linear operator, ``DIRECT`` can be used for training models for recovering images such as MRIs from partially observed or noisy input data.
-``DIRECT`` stores inverse problem solvers such as the Learned Primal Dual algorithm, Recurrent Inference Machine and Recurrent Variational Network, which were part of the winning solution in Facebook & NYUs FastMRI challenge in 2019 and the Calgary-Campinas MRI reconstruction challenge at MIDL 2020. For a full list of the baselines currently implemented in DIRECT see `here <#baselines-and-trained-models>`_.
+``DIRECT`` is a Python, end-to-end pipeline for solving Inverse Problems emerging in Imaging Processing.
+It is built with PyTorch and stores state-of-the-art Deep Learning imaging inverse problem solvers such as denoising, dealiasing and reconstruction.
+By defining a base forward linear or non-linear operator, ``DIRECT`` can be used for training models for recovering images such as MRIs from partially observed or noisy input data.
+``DIRECT`` stores inverse problem solvers such as the vSHARP, Learned Primal Dual algorithm, Recurrent Inference Machine and Recurrent Variational Network, which were part of the winning solutions in Facebook & NYUs FastMRI challenge in 2019, the Calgary-Campinas MRI reconstruction challenge at MIDL 2020 and the CMRxRecon challenge 2023.
+For a full list of the baselines currently implemented in DIRECT see `here <#baselines-and-trained-models>`_.
 
 .. raw:: html
 
@@ -49,7 +52,7 @@ In the `projects <https://github.com/NKI-AI/direct/tree/main/projects>`_ folder
 Baselines and trained models
 ----------------------------
 
-We provide a set of baseline results and trained models in the `DIRECT Model Zoo <https://docs.aiforoncology.nl/direct/model_zoo.html>`_. Baselines and trained models include the `Recurrent Variational Network (RecurrentVarNet) <https://arxiv.org/abs/2111.09639>`_, the `Recurrent Inference Machine (RIM) <https://www.sciencedirect.com/science/article/abs/pii/S1361841518306078>`_, the `End-to-end Variational Network (VarNet) <https://arxiv.org/pdf/2004.06688.pdf>`_, the `Learned Primal Dual Network (LDPNet) <https://arxiv.org/abs/1707.06474>`_, the `X-Primal Dual Network (XPDNet) <https://arxiv.org/abs/2010.07290>`_, the `KIKI-Net <https://pubmed.ncbi.nlm.nih.gov/29624729/>`_, the `U-Net <https://arxiv.org/abs/1811.08839>`_, the `Joint-ICNet <https://openaccess.thecvf.com/content/CVPR2021/papers/Jun_Joint_Deep_Model-Based_MR_Image_and_Coil_Sensitivity_Reconstruction_Network_CVPR_2021_paper.pdf>`_, and the `AIRS Medical fastmri model (MultiDomainNet) <https://arxiv.org/pdf/2012.06318.pdf>`_.
+We provide a set of baseline results and trained models in the `DIRECT Model Zoo <https://docs.aiforoncology.nl/direct/model_zoo.html>`_. Baselines and trained models include the `vSHARP <https://arxiv.org/abs/2309.09954>`_, `Recurrent Variational Network (RecurrentVarNet) <https://arxiv.org/abs/2111.09639>`_, the `Recurrent Inference Machine (RIM) <https://www.sciencedirect.com/science/article/abs/pii/S1361841518306078>`_, the `End-to-end Variational Network (VarNet) <https://arxiv.org/pdf/2004.06688.pdf>`_, the `Learned Primal Dual Network (LDPNet) <https://arxiv.org/abs/1707.06474>`_, the `X-Primal Dual Network (XPDNet) <https://arxiv.org/abs/2010.07290>`_, the `KIKI-Net <https://pubmed.ncbi.nlm.nih.gov/29624729/>`_, the `U-Net <https://arxiv.org/abs/1811.08839>`_, the `Joint-ICNet <https://openaccess.thecvf.com/content/CVPR2021/papers/Jun_Joint_Deep_Model-Based_MR_Image_and_Coil_Sensitivity_Reconstruction_Network_CVPR_2021_paper.pdf>`_, and the `AIRS Medical fastmri model (MultiDomainNet) <https://arxiv.org/pdf/2012.06318.pdf>`_.
 
 License and usage
 -----------------
@@ -63,15 +66,15 @@ If you use DIRECT in your own research, or want to refer to baseline results pub
 
 .. code-block:: BibTeX
 
-   @misc{DIRECTTOOLKIT,
-     doi = {10.21105/joss.04278},
-     url = {https://doi.org/10.21105/joss.04278},
-     year = {2022},
-     publisher = {The Open Journal},
-     volume = {7},
-     number = {73},
-     pages = {4278},
-     author = {George Yiasemis and Nikita Moriakov and Dimitrios Karkalousos and Matthan Caan and Jonas Teuwen},
-     title = {DIRECT: Deep Image REConstruction Toolkit},
-     journal = {Journal of Open Source Software}
-   }
+    @article{DIRECTTOOLKIT,
+        doi = {10.21105/joss.04278},
+        url = {https://doi.org/10.21105/joss.04278},
+        year = {2022},
+        publisher = {The Open Journal},
+        volume = {7},
+        number = {73},
+        pages = {4278},
+        author = {George Yiasemis and Nikita Moriakov and Dimitrios Karkalousos and Matthan Caan and Jonas Teuwen},
+        title = {DIRECT: Deep Image REConstruction Toolkit},
+        journal = {Journal of Open Source Software}
+    }

direct/data/datasets_config.py

@@ -46,7 +46,7 @@ class NormalizationTransformConfig(BaseConfig):
 
 @dataclass
 class TransformsConfig(BaseConfig):
-    masking: MaskingConfig = MaskingConfig()
+    masking: Optional[MaskingConfig] = MaskingConfig()
     cropping: CropTransformConfig = CropTransformConfig()
     random_augmentations: RandomAugmentationTransformsConfig = RandomAugmentationTransformsConfig()
     padding_eps: float = 0.001

direct/nn/vsharp/__init__.py

@@ -0,0 +1 @@
+# Copyright (c) DIRECT Contributors

direct/nn/vsharp/config.py

@@ -0,0 +1,36 @@
+# Copyright (c) DIRECT Contributors
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from direct.config.defaults import ModelConfig
+from direct.nn.types import ActivationType, InitType, ModelName
+
+
+@dataclass
+class VSharpNetConfig(ModelConfig):
+    num_steps: int = 10
+    num_steps_dc_gd: int = 8
+    image_init: InitType = InitType.SENSE
+    no_parameter_sharing: bool = True
+    auxiliary_steps: int = 0
+    image_model_architecture: ModelName = ModelName.UNET
+    initializer_channels: tuple[int, ...] = (32, 32, 64, 64)
+    initializer_dilations: tuple[int, ...] = (1, 1, 2, 4)
+    initializer_multiscale: int = 1
+    initializer_activation: ActivationType = ActivationType.PRELU
+    image_resnet_hidden_channels: int = 128
+    image_resnet_num_blocks: int = 15
+    image_resnet_batchnorm: bool = True
+    image_resnet_scale: float = 0.1
+    image_unet_num_filters: int = 32
+    image_unet_num_pool_layers: int = 4
+    image_unet_dropout: float = 0.0
+    image_didn_hidden_channels: int = 16
+    image_didn_num_dubs: int = 6
+    image_didn_num_convs_recon: int = 9
+    image_conv_hidden_channels: int = 64
+    image_conv_n_convs: int = 15
+    image_conv_activation: str = ActivationType.RELU
+    image_conv_batchnorm: bool = False