Add config

CBCT/Model.py

@@ -0,0 +1,63 @@
+from konfai.network import network
+import segmentation_models_pytorch as smp
+import torch
+from konfai.predictor import Reduction
+from konfai.data.transform import Transform
+from konfai.utils.dataset import Attribute, Dataset, data_to_image
+
+class Head(network.ModuleArgsDict):
+
+    def __init__(self):
+        super().__init__()
+        self.add_module("Tanh", torch.nn.Tanh())
+
+class UNetpp(network.Network):
+    
+    def __init__(self,
+                optimizer : network.OptimizerLoader = network.OptimizerLoader(),
+                schedulers: dict[str, network.LRSchedulersLoader] = {
+                    "default:ReduceLROnPlateau": network.LRSchedulersLoader(0)
+                },
+                outputs_criterions: dict[str, network.TargetCriterionsLoader] = {"default" : network.TargetCriterionsLoader()},
+                pretrained: bool = False):
+        super().__init__(in_channels = 3, optimizer = optimizer, schedulers = schedulers, outputs_criterions = outputs_criterions, dim = 2)
+        self.add_module("model", smp.UnetPlusPlus(
+            encoder_name="resnet34",
+            encoder_weights=None if not pretrained else "imagenet", 
+            in_channels=3,
+            classes=1,  
+            activation=None  
+        ))
+        self.add_module("Head", Head())
+
+class Concat(Reduction):
+
+    def __init__(self):
+        pass
+
+    def __call__(self, tensors: list[torch.Tensor]) -> torch.Tensor:
+        return torch.cat(tensors, dim=1)
+        
+class Uncertainty(Transform):
+
+    def __init__(self):
+        pass
+
+    def __call__(self, name: str, tensors: torch.Tensor, cache_attribute: Attribute) -> torch.Tensor:
+        for i, tensor in enumerate(tensors):
+            dataset = Dataset("./Predictions/ImpactSynth/Dataset", "mha")
+            dataset.write(f"sCT_{i}", name, data_to_image(tensor.unsqueeze(0).numpy(), cache_attribute))
+        return tensors.mean(0).unsqueeze(0)
+    
+class UnNormalize(Transform):
+
+    def __init__(self) -> None:
+        super().__init__()
+        self.v_min = -1024
+        self.v_max = 3071
+
+    def __call__(self, name: str, input : torch.Tensor, cache_attribute: Attribute) -> torch.Tensor:
+        return (input + 1)/2*(self.v_max-self.v_min) + self.v_min
+    
+    def inverse(self, name: str, input : torch.Tensor, cache_attribute: Attribute) -> torch.Tensor:
+        pass

CBCT/Prediction.yml

@@ -0,0 +1,84 @@
+Predictor:
+  Model:
+    classpath: Model:UNetpp
+    UNetpp:
+      outputs_criterions: None
+      pretrained: false
+  Dataset:
+    groups_src:
+      Volume:
+        groups_dest:
+          Volume:
+            transforms:
+              Clip:
+                min_value: min
+                max_value: percentile:99.5
+                save_clip_min: false
+                save_clip_max: false
+                mask: None
+              Normalize:
+                lazy: true
+                channels: None
+                min_value: -1
+                max_value: 1
+                inverse: true
+            patch_transforms:
+              Normalize:
+                lazy: false
+                channels: None
+                min_value: -1
+                max_value: 1
+                inverse: true
+            is_input: true
+    augmentations: 
+      DataAugmentation_0:
+        data_augmentations:
+          Flip:
+            f_prob:
+            - 0
+            - 0.5
+            - 0.5
+            prob: 1
+        nb: 2
+    Patch:
+      patch_size:
+      - 1
+      - 512
+      - 512
+      overlap: None
+      mask: None
+      pad_value: -1
+      extend_slice: 2
+    subset: None
+    filter: None
+    dataset_filenames:
+    - ./Dataset:nii.gz
+    use_cache: false
+    batch_size: 8
+  outputs_dataset:
+    Head:Tanh:
+      OutputDataset:
+        name_class: OutSameAsGroupDataset
+        before_reduction_transforms: None
+        after_reduction_transforms: None
+        final_transforms:
+          Model:UnNormalize: {}
+          Model:Uncertainty: {}
+          TensorCast:
+            dtype: int16
+            inverse: true
+        dataset_filename: Dataset:mha
+        group: sCT
+        same_as_group: Volume:Volume
+        patch_combine: None
+        inverse_transform: false
+        reduction: Model:Concat
+        Model:Concat: {}
+  train_name: ImpactSynth
+  manual_seed: 32
+  gpu_checkpoints: None
+  images_log: None
+  combine: Model:Concat
+  autocast: false
+  data_log: None
+  Model:Concat: {}

CBCT/metadata.json

@@ -0,0 +1,7 @@
+{
+    "display_name": "CBCT",
+    "short_description": "<b>Description:</b><br>Supervised CBCT synthesis model developed as part of the <a href=\"https://synthrad2025.grand-challenge.org/\">SynthRAD 2025 Challenge (Task 2)</a>.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
+    "description": "<b>Description:</b><br>Supervised CBCT synthesis model developed as part of the <a href=\"https://synthrad2025.grand-challenge.org/\">SynthRAD 2025 Challenge (Task 2)</a>.<br><br><b>Architecture:</b><br>Based on a 2.5D UNet++ with a ResNet34 encoder, the model was optimized using the <b>IMPACT-Synth loss</b>, a perceptual loss leveraging semantic priors from <b>SAM 2.1-s</b>. Training was conducted with the <b>KonfAI</b> deep learning framework.<br><br><b>Training data:</b><br>Paired CT–CBCT volumes from the <a href=\"https://synthrad2025.grand-challenge.org/\">SynthRAD 2025 Challenge (Task 2)</a>, <b>aligned using IMPACT-based registration</b>. Corresponding B-spline deformation fields are available in the <a href=\"https://huggingface.co/datasets/VBoussot/synthrad2025-impact-registration\">SynthRAD2025-IMPACT (aligned)</a> dataset repository.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
+    "tta": 2,
+    "mc_dropout": false
+}

CBCT/requirements.txt

@@ -0,0 +1 @@
+segmentation_models_pytorch

MR/Model.py

@@ -0,0 +1,63 @@
+from konfai.network import network
+import segmentation_models_pytorch as smp
+import torch
+from konfai.predictor import Reduction
+from konfai.data.transform import Transform
+from konfai.utils.dataset import Attribute, Dataset, data_to_image
+
+class Head(network.ModuleArgsDict):
+
+    def __init__(self):
+        super().__init__()
+        self.add_module("Tanh", torch.nn.Tanh())
+
+class UNetpp(network.Network):
+    
+    def __init__(self,
+                optimizer : network.OptimizerLoader = network.OptimizerLoader(),
+                schedulers: dict[str, network.LRSchedulersLoader] = {
+                    "default:ReduceLROnPlateau": network.LRSchedulersLoader(0)
+                },
+                outputs_criterions: dict[str, network.TargetCriterionsLoader] = {"default" : network.TargetCriterionsLoader()},
+                pretrained: bool = False):
+        super().__init__(in_channels = 3, optimizer = optimizer, schedulers = schedulers, outputs_criterions = outputs_criterions, dim = 2)
+        self.add_module("model", smp.UnetPlusPlus(
+            encoder_name="resnet34",
+            encoder_weights=None if not pretrained else "imagenet", 
+            in_channels=3,
+            classes=1,  
+            activation=None  
+        ))
+        self.add_module("Head", Head())
+
+class Concat(Reduction):
+
+    def __init__(self):
+        pass
+
+    def __call__(self, tensors: list[torch.Tensor]) -> torch.Tensor:
+        return torch.cat(tensors, dim=1)
+        
+class Uncertainty(Transform):
+
+    def __init__(self):
+        pass
+
+    def __call__(self, name: str, tensors: torch.Tensor, cache_attribute: Attribute) -> torch.Tensor:
+        for i, tensor in enumerate(tensors):
+            dataset = Dataset("./Predictions/ImpactSynth/Dataset", "mha")
+            dataset.write(f"sCT_{i}", name, data_to_image(tensor.unsqueeze(0).numpy(), cache_attribute))
+        return tensors.mean(0).unsqueeze(0)
+    
+class UnNormalize(Transform):
+
+    def __init__(self) -> None:
+        super().__init__()
+        self.v_min = -1024
+        self.v_max = 3071
+
+    def __call__(self, name: str, input : torch.Tensor, cache_attribute: Attribute) -> torch.Tensor:
+        return (input + 1)/2*(self.v_max-self.v_min) + self.v_min
+    
+    def inverse(self, name: str, input : torch.Tensor, cache_attribute: Attribute) -> torch.Tensor:
+        pass

MR/Prediction.yml

@@ -0,0 +1,72 @@
+Predictor:
+  Model:
+    classpath: Model:UNetpp
+    UNetpp:
+      outputs_criterions: None
+      pretrained: false
+  Dataset:
+    groups_src:
+      Volume:
+        groups_dest:
+          Volume:
+            transforms:
+              Standardize:
+                lazy: false
+                mean: None
+                std: None
+                mask: None
+                inverse: true
+            patch_transforms: None
+            is_input: true
+    augmentations:
+      DataAugmentation_0:
+        data_augmentations:
+          Flip:
+            f_prob:
+            - 0
+            - 0.5
+            - 0.5
+            prob: 1
+        nb: 2
+    Patch:
+      patch_size:
+      - 1
+      - 512
+      - 512
+      overlap: None
+      mask: None
+      pad_value: -1
+      extend_slice: 2
+    subset: None
+    filter: None
+    dataset_filenames:
+    - ./Dataset:nii.gz
+    use_cache: false
+    batch_size: 8
+  outputs_dataset:
+    Head:Tanh:
+      OutputDataset:
+        name_class: OutSameAsGroupDataset
+        before_reduction_transforms: None
+        after_reduction_transforms: None
+        final_transforms:
+          Model:UnNormalize: {}
+          Model:Uncertainty: {}
+          TensorCast:
+            dtype: int16
+            inverse: true
+        dataset_filename: Dataset:mha
+        group: sCT
+        same_as_group: Volume:Volume
+        patch_combine: None
+        inverse_transform: false
+        reduction: Model:Concat
+        Model:Concat: {}
+  train_name: ImpactSynth
+  manual_seed: 32
+  gpu_checkpoints: None
+  images_log: None
+  combine: Model:Concat
+  autocast: false
+  data_log: None
+  Model:Concat: {}

MR/metadata.json

@@ -0,0 +1,7 @@
+{
+    "display_name": "MR",
+    "short_description": "<b>Description:</b><br>Supervised MRI (T1-weighted) synthesis model developed as part of the <a href=\"https://synthrad2025.grand-challenge.org/\">SynthRAD 2025 Challenge (Task 1)</a>.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
+    "description": "<b>Description:</b><br>Supervised MRI (T1-weighted) synthesis model developed as part of the <a href=\"https://synthrad2025.grand-challenge.org/\">SynthRAD 2025 Challenge (Task 1)</a>.<br><br><b>Architecture:</b><br>Based on a 2.5D UNet++ with a ResNet34 encoder, the model was optimized using the <b>IMPACT-Synth loss</b>, a perceptual loss leveraging semantic priors from <b>SAM 2.1-s</b>. Training was conducted with the <b>KonfAI</b> deep learning framework.<br><br><b>Training data:</b><br>Paired CT–MRI (T1-weighted) volumes from the <a href=\"https://synthrad2025.grand-challenge.org/\">SynthRAD 2025 Challenge (Task 1)</a>, <b>aligned using IMPACT-based registration</b>. Corresponding B-spline deformation fields are available in the <a href=\"https://huggingface.co/datasets/VBoussot/synthrad2025-impact-registration\">SynthRAD2025-IMPACT (aligned)</a> dataset repository.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
+    "tta": 2,
+    "mc_dropout": false
+}

MR/requirements.txt

@@ -0,0 +1 @@
+segmentation_models_pytorch