Add ImpactSeg body

body/Body.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a4652d0c8139e2d3ce86247bf3e050671de369fe982a701bf6bf76304131c7a
+size 423981609

body/Evaluation.yml

@@ -0,0 +1,29 @@
+Evaluator:
+  metrics:
+    Output:
+      targets_criterions:
+        Reference;Mask:
+          criterions_loader:
+            DiceSaveMap:
+              labels: None
+              dataset: ./Evaluations/ImpactSeg-Body/Output:mha
+              group: Seg_MAE_map
+  Dataset:
+    groups_src:
+      Mask_0:
+        groups_dest:
+          Mask:
+            transforms: None
+      Volume_0:
+        groups_dest:
+          Output:
+            transforms: None
+      Reference_0:
+        groups_dest:
+          Reference:
+            transforms: None
+    subset: None
+    dataset_filenames:
+    - ./Dataset:a:mha
+    validation: None
+  train_name: ImpactSeg-Body

body/Model.py

@@ -0,0 +1,53 @@
+import torch
+from torch import nn
+
+from dynamic_network_architectures.architectures.unet import ResidualEncoderUNet
+from konfai.network import network
+
+
+def _replace_unpicklable_identities(module: nn.Module) -> None:
+    """Replace library lambdas such as ``lambda x: x`` with ``nn.Identity``."""
+    for child in module.modules():
+        if hasattr(child, "skip") and callable(child.skip) and not isinstance(child.skip, nn.Module):
+            child.skip = nn.Identity()
+        if hasattr(child, "nonlin2") and callable(child.nonlin2) and not isinstance(child.nonlin2, nn.Module):
+            child.nonlin2 = nn.Identity()
+    
+class ResEncMCT(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()},
+        in_channels: int = 5,
+        nb_class: int = 132,
+    ) -> None:
+        super().__init__(
+            in_channels=in_channels,
+            optimizer=optimizer,
+            schedulers=schedulers,
+            outputs_criterions=outputs_criterions,
+            dim=2,
+        )
+        self.add_module("DecoderOutputs", ResidualEncoderUNet(
+            input_channels=in_channels,
+            n_stages=7,
+            features_per_stage=(32, 64, 128, 256, 512, 512, 512),
+            conv_op=nn.Conv2d,
+            kernel_sizes=3,
+            strides=(1, 2, 2, 2, 2, 2, 2),
+            n_blocks_per_stage=(1, 3, 4, 6, 6, 6, 6),
+            num_classes=nb_class,
+            n_conv_per_stage_decoder=(1, 1, 1, 1, 1, 1),
+            conv_bias=True,
+            norm_op=nn.InstanceNorm2d,
+            norm_op_kwargs={"eps": 1e-5, "affine": True},
+            dropout_op=None,
+            dropout_op_kwargs=None,
+            nonlin=nn.LeakyReLU,
+            nonlin_kwargs={"inplace": True},
+            deep_supervision=False,
+        ))
+        _replace_unpicklable_identities(self.DecoderOutputs)

body/Prediction.yml

@@ -0,0 +1,78 @@
+Predictor:
+  Model:
+    classpath: Model:ResEncMCT
+    ResEncMCT:
+      in_channels: 5
+      nb_class: 12
+      outputs_criterions: None
+  Dataset:
+    groups_src:
+      Volume_0:
+        groups_dest:
+          Volume:
+            transforms:
+              TensorCast:
+                dtype: float32
+                inverse: false
+              ResampleToResolution:
+                spacing:
+                - 3
+                - 3
+                - 0
+                inverse: true
+              Standardize:
+                lazy: false
+                mean: None
+                std: None
+                mask: None
+                inverse: false
+            patch_transforms: None
+            is_input: true
+    augmentations: None
+    Patch:
+      patch_size:
+      - 1
+      - 192
+      - 192
+      overlap: None
+      mask: None
+      pad_value: -2
+      extend_slice: 4
+    subset: None
+    filter: None
+    dataset_filenames:
+    - ./Dataset/:mha
+    use_cache: false
+    batch_size: 32
+    num_workers: None
+    pin_memory: false
+    prefetch_factor: None
+    persistent_workers: None
+  outputs_dataset:
+    DecoderOutputs:
+      OutputDataset:
+        name_class: OutSameAsGroupDataset
+        before_reduction_transforms: None
+        after_reduction_transforms: None
+        final_transforms:
+          Softmax:
+            dim: 0
+          Argmax:
+            dim: 0
+          TensorCast:
+            dtype: uint8
+            inverse: true
+        dataset_filename: Output:mha
+        group: Output
+        same_as_group: Volume_0:Volume
+        patch_combine: None
+        inverse_transform: true
+        reduction: Mean
+        Mean: {}
+  train_name: ImpactSeg-Body
+  manual_seed: 42
+  gpu_checkpoints: None
+  images_log: None
+  combine: Mean
+  autocast: false
+  data_log: None

body/app.json

@@ -0,0 +1,51 @@
+{
+    "display_name": "Segmentation: IMPACTSeg CBCT/MR/CT",
+    "short_description": "<b>Description:</b><br><b>IMPACTSeg</b> is a multimodal anatomical segmentation model for <b>CBCT, MR, and CT</b> scans. It predicts <b>11 labels</b> spanning soft tissues, cavities, bones, and central structures, and was trained on <b>232 CBCT + 282 MR + 955 CT</b> cases.",
+    "description": "<b>Description:</b><br><b>IMPACTSeg</b> is a multimodal anatomical segmentation model packaged for inference with <b>KonfAI</b>. It is designed for <b>CBCT, MR, and CT</b> scans and produces a consistent set of <b>11 labels</b> across modalities.<br><br><b>Training cohort:</b><br><b>232 CBCT + 282 MR + 955 CT</b> cases.<br><br><b>Use case:</b><br>Automated multimodal segmentation for downstream analysis, quantitative workflows, and clinical research pipelines.",
+    "tta": 0,
+    "mc_dropout": 0,
+    "models": ["Body.pt"],
+    "inputs": {
+      "Volume": {
+        "display_name": "Input Volume",
+        "volume_type": "VOLUME",
+        "required": true
+      }
+    },
+    "outputs": {
+      "Segmentation": {
+        "display_name": "Segmentation",
+        "volume_type": "SEGMENTATION",
+        "required": true
+      }
+    },
+    "inputs_evaluations": {
+      "Image": {
+        "Evaluation.yml": {
+          "Segmentation": {
+            "display_name": "Output Segmentation",
+            "volume_type": "VOLUME",
+            "required": true
+          },
+          "GT_Segmentation": {
+            "display_name": "GT Segmentation",
+            "volume_type": "VOLUME",
+            "required": true
+          }
+        }
+      }
+    },
+    "terminology": {
+      "1": { "name": "subcutaneous_tissue", "color": "#F4A261" },
+      "2": { "name": "muscle", "color": "#E76F51" },
+      "3": { "name": "abdominal_cavity", "color": "#2A9D8F" },
+      "4": { "name": "thoracic_cavity", "color": "#264653" },
+      "5": { "name": "bones", "color": "#E9C46A" },
+      "6": { "name": "gland_structure", "color": "#8AB17D" },
+      "7": { "name": "pericardium", "color": "#C8553D" },
+      "8": { "name": "prosthetic_breast_implant", "color": "#B56576" },
+      "9": { "name": "mediastinum", "color": "#577590" },
+      "10": { "name": "spinal_cord", "color": "#6D597A" },
+      "11": { "name": "brain", "color": "#43AA8B" }
+    }
+}

body/requirements.txt

@@ -0,0 +1 @@
+dynamic_network_architectures