MHub / GC - Add GC nnUNet Pancreas model

models/gc_nnunet_pancreas/__init__.py

@@ -0,0 +1 @@
+from .utils import *

models/gc_nnunet_pancreas/config/default.yml

@@ -0,0 +1,44 @@
+general:
+  version: 1.0
+  data_base_dir: /app/data
+  description: base configuration for GC NNUnet Pancreas model (dicom to dicom, and json output)
+
+execute:
+- DicomImporter
+- MhaConverter
+- GCNNUnetPancreasRunner
+- DsegConverter
+- ReportExporter
+- DataOrganizer
+
+modules:
+  DicomImporter:
+    source_dir: input_data
+    import_dir: sorted_data
+    sort_data: true
+    meta: 
+      mod: '%Modality'
+
+  MhaConverter:
+    engine: panimg
+    targets: [dicom:mod=ct]
+
+  DsegConverter:
+    model_name: 'GC NNUnet Pancreas'
+    source_segs: ['mha:mod=seg:src=cleaned']
+    target_dicom: dicom:mod=ct
+    skip_empty_slices: True
+
+  ReportExporter:
+    format: compact
+    includes:
+      - data: prostate_cancer_likelihood
+        label: prostate_cancer_likelihood
+        value: value
+
+  DataOrganizer:
+    targets:
+    - mha:mod=heatmap-->[i:sid]/nnunet_pancreas_heatmap.mha
+    - mha:mod=seg:src=cleaned-->[i:sid]/nnunet_pancreas.seg.mha
+    - dicomseg:mod=seg-->[i:sid]/nnunet_pancreas.seg.dcm
+    - json:mod=report-->[i:sid]/nnunet_pancreas_case_level_likelihood.json

models/gc_nnunet_pancreas/dockerfiles/Dockerfile

@@ -0,0 +1,45 @@
+FROM mhubai/base:latest
+
+# Specify/override authors label
+LABEL authors="[email protected]"
+
+# Install PyTorch 2.0.1 (CUDA enabled)
+RUN pip3 install --no-cache-dir torch==2.0.1+cu118 -f https://download.pytorch.org/whl/torch_stable.html
+
+# Install git-lfs (required for downloading the model weights)
+RUN apt update && \
+    apt install -y --no-install-recommends git-lfs && \
+    rm -rf /var/lib/apt/lists/*
+
+# Install the model weights and the algorithm files
+#  * Pull algorithm from repo into /opt/algorithm (main branch, commit 15dd550beada43a8a55b81a32d9b3904a1cf8d30)
+#  * Remove .git folder to keep docker layer small
+RUN git clone https://github.com/DIAGNijmegen/CE-CT_PDAC_AutomaticDetection_nnUnet.git /opt/algorithm && \
+    cd /opt/algorithm && \
+    git reset --hard 15dd550beada43a8a55b81a32d9b3904a1cf8d30 && \
+    rm -rf /opt/algorithm/.git
+
+# Set this environment variable as a shortcut to avoid nnunet 1.7.0 crashing the build
+# by pulling sklearn instead of scikit-learn
+# N.B. this is a known issue:
+# https://github.com/MIC-DKFZ/nnUNet/issues/1281
+# https://github.com/MIC-DKFZ/nnUNet/pull/1209
+ENV SKLEARN_ALLOW_DEPRECATED_SKLEARN_PACKAGE_INSTALL=True
+
+# Install nnUNet 1.7.0 and other requirements
+RUN pip3 install --no-cache-dir evalutils==0.3.0 nnunet==1.7.0
+
+# Extend the nnUNet installation with custom trainers
+RUN SITE_PKG=`pip3 show nnunet | grep "Location:" | awk '{print $2}'` && \
+    mv /opt/algorithm/nnUNetTrainerV2_Loss_CE_checkpoints.py "$SITE_PKG/nnunet/training/network_training/nnUNetTrainerV2_Loss_CE_checkpoints.py"
+
+# Import the MHub model definiton
+ARG MHUB_MODELS_REPO
+RUN buildutils/import_mhub_model.sh gc_nnunet_pancreas ${MHUB_MODELS_REPO}
+
+# Add algorithm files to python path
+ENV PYTHONPATH=/opt/algorithm:/app
+
+# Configure main entrypoint
+ENTRYPOINT ["python3", "-m", "mhubio.run"]
+CMD ["--config", "/app/models/gc_nnunet_pancreas/config/default.yml"]

models/gc_nnunet_pancreas/meta.json

@@ -0,0 +1,149 @@
+{
+  "id": "bf7ae4bb-c6f5-4b1e-89aa-a8de246def57",
+  "name": "gc_nnunet_pancreas",
+  "title": "Pancreatic Ductal Adenocarcinoma Detection in CT",
+  "summary": {
+    "description": "This algorithm produces a tumor likelihood heatmap for the presence of pancreatic ductal adenocarcinoma (PDAC) in an input venous-phase contrast-enhanced computed tomography scan (CECT). Additionally, the algorithm provides the segmentation of multiple surrounding anatomical structures such as the pancreatic duct, common bile duct, veins and arteries. The heatmap and segmentations are resampled to the same spatial resolution and physical dimensions as the input CECT image for easier visualisation.",
+    "inputs": [
+      {
+        "label": "Venous phase CT scan",
+        "description": "A contrast-enhanced CT scan in the venous phase and axial reconstruction",
+        "format": "DICOM",
+        "modality": "CT",
+        "bodypartexamined": "Abdomen",
+        "slicethickness": "2.5mm",
+        "non-contrast": false,
+        "contrast": false
+      }
+    ],
+    "outputs": [
+      {
+        "type": "Prediction",
+        "valueType": "Likelihood map",
+        "label": "Pancreatic tumor likelihood heatmap",
+        "description": "Pancreatic tumor likelihood heatmap, where each voxel represents a floating point in range [0,1].",
+        "classes": []
+      },
+      {
+        "type": "Prediction",
+        "valueType": "Likelihood",
+        "label": "Pancreatic tumor likelihood",
+        "description": "Case-level pancreatic tumor likelihood value with a value in range [0,1].",
+        "classes": []
+      },
+      {
+        "type": "Segmentation",
+        "label": "Pancreas segmentation",
+        "description": "Segmentation of pancreas related tissues, these segmentation classes were not thoroughly validated, use them on your own risk!",
+        "classes": [
+          "veins",
+          "arteries",
+          "pancreas",
+          "pancreatic duct",
+          "bile duct"
+        ]
+      }
+    ],
+    "model": {
+      "architecture": "nnUnet ",
+      "training": "supervised",
+      "cmpapproach": "3D"
+    },
+    "data": {
+      "training": {
+        "vol_samples": 242
+      },
+      "evaluation": {
+        "vol_samples": 361
+      },
+      "public": true,
+      "external": false
+    }
+  },
+  "details": {
+    "name": "Fully Automatic Deep Learning Framework for Pancreatic Ductal Adenocarcinoma Detection on Computed Tomography",
+    "version": "0.1.0",
+    "devteam": "DIAGNijmegen (Diagnostic Image Analysis Group, Radboud UMC, The Netherlands)",
+    "type": "The models were developed using nnUnet. All models employed a 3D U-Net as the base architecture and were trained for 250.000 training steps with five-fold cross-validation.",
+    "date": {
+      "weights": "2023-06-28",
+      "code": "2022-07-19",
+      "pub": "2022-01-13"
+    },
+    "cite": "Alves N, Schuurmans M, Litjens G, Bosma JS, Hermans J, Huisman H. Fully Automatic Deep Learning Framework for Pancreatic Ductal Adenocarcinoma Detection on Computed Tomography. Cancers (Basel). 2022 Jan 13;14(2):376. doi: 10.3390/cancers14020376. PMID: 35053538; PMCID: PMC8774174.",
+    "license": {
+      "code": "Apache 2.0",
+      "weights": "Apache 2.0"
+    },
+    "publications": [
+      {
+        "title": "Fully Automatic Deep Learning Framework for Pancreatic Ductal Adenocarcinoma Detection on Computed Tomography ",
+        "uri": "https://www.mdpi.com/2072-6694/14/2/376"
+      }
+    ],
+    "github": "https://github.com/DIAGNijmegen/CE-CT_PDAC_AutomaticDetection_nnUnet",
+    "zenodo": "",
+    "colab": "",
+    "slicer": false
+  },
+  "info": {
+    "use": {
+      "title": "Intended Use",
+      "text": "This algorithm is intended to be used only on venous-phase CECT examinations of patients with clinical suspicion of PDAC. This algorithm should not be used in different patient demographics.",
+      "references": [],
+      "tables": []
+    },
+    "analyses": {
+      "title": "Analysis",
+      "text": "The study evaluated a medical model's performance for tumor detection by analyzing receiver operating characteristic (ROC) and free-response receiver operating characteristic (FROC) curves, assessing both tumor presence and lesion localization, and compared three configurations using statistical tests and ensemble modeling. The table below lists the model's performance on an external evaluation dataset of 361 cases. Additional analysis details and results can be found in the original paper [1].",
+      "references": [
+        {
+          "label": "Fully Automatic Deep Learning Framework for Pancreatic Ductal Adenocarcinoma Detection on Computed Tomography",
+          "uri": "https://www.mdpi.com/2072-6694/14/2/376"
+        }
+      ],
+      "tables": [
+        {
+          "label": "Evaluation results of the nnUnet_MS model on the external test set of 361 cases.",
+          "entries": {
+              "Mean AUC-ROC (95% CI)": "0.991 (0.970-1.0)",
+              "Mean pAUC-FROC (95% CI)": "3.996 (3.027-4.965)"
+          }
+        }
+      ]
+    },
+    "evaluation": {
+      "title": "Evaluation Data",
+      "text": "This framework was tested in an independent, external cohort consisting of two publicly available datasets of respectively 281 and 80 patients each. The Medical Segmentation Decathlon pancreas dataset (training portion) [1] consisting of 281 patients with pancreatic malignancies (including lesions in the head, neck, body, and tail of the pancreas) and voxel-level annotations for the pancreas and lesion. The Cancer Imaging Archive dataset from the US National Institutes of Health Clinical Center [2], containing 80 patients with normal pancreas and respective voxel-level annotations.",
+      "references": [
+        {
+          "label": "The Medical Segmentation Decathlon pancreas dataset (training portion)",
+          "uri": "http://medicaldecathlon.com/"
+        },
+        {
+          "label": "The Cancer Imaging Archive dataset from the US National Institutes of Health Clinical Center",
+          "uri": "https://wiki.cancerimagingarchive.net/display/Public/Pancreas-CT"
+        }
+      ],
+      "tables": []
+    },
+    "training": {
+      "title": "Training data",
+      "text": "CE-CT scans in the portal venous phase from 119 patients with pathology-proven PDAC in the pancreatic head (PDAC cohort) and 123 patients with normal pancreas (non-PDAC cohort), acquired between 1 January 2013 and 1 June 2020, were selected for model development.",
+      "references": [],
+      "tables": []
+    },
+    "ethics": {
+      "title": "",
+      "text": "",
+      "references": [],
+      "tables": []
+    },
+    "limitations": {
+      "title": "Before using this model",
+      "text": "Test the model retrospectively and prospectively on a diagnostic cohort that reflects the target population that the model will be used upon to confirm the validity of the model within a local setting.",
+      "references": [],
+      "tables": []
+    }
+  }
+}

models/gc_nnunet_pancreas/utils/GCNNUnetPancreasRunner.py

@@ -0,0 +1,85 @@
+"""
+-----------------------------------------------------------
+GC / MHub - Run Module for the GC NNUnet Pancreas Algorithm
+-----------------------------------------------------------
+
+-----------------------------------------------------------
+Author: Sil van de Leemput
+Email:  [email protected]
+-----------------------------------------------------------
+"""
+
+from mhubio.core import Module, Instance, InstanceData, DataType, Meta, IO, ValueOutput
+
+from pathlib import Path
+import SimpleITK
+import sys
+
+
+CLI_PATH = Path(__file__).parent / "cli.py"
+
+
+@ValueOutput.Name('prostate_cancer_likelihood')
+@ValueOutput.Label('ProstateCancerLikelihood')
+@ValueOutput.Meta(Meta(min=0.0, max=1.0, type="likelihood"))
+@ValueOutput.Type(float)
+@ValueOutput.Description('Likelihood of case-level prostate cancer.')
+class ProstateCancerLikelihood(ValueOutput):
+    pass
+
+
+class GCNNUnetPancreasRunner(Module):
+    @IO.Instance()
+    @IO.Input('in_data', 'mha:mod=ct', the="input data")
+    @IO.Output('heatmap', 'heatmap.mha', 'mha:mod=heatmap:model=GCNNUnetPancreas', data="in_data",
+               the="raw heatmap of the pancreatic tumor likelihood (not masked with any pancreas segmentations).")
+    @IO.Output('segmentation_raw', 'segmentation_raw.mha', 'mha:mod=seg:src=original:model=GCNNUnetPancreas:roi=VEIN,ARTERY,PANCREAS,PANCREATIC_DUCT,BILE_DUCT,PANCREAS+CYST,RENAL_VEIN', data="in_data",
+               the="original segmentation of the pancreas, with the following classes: "
+                   "0-background, 1-veins, 2-arteries, 3-pancreas, 4-pancreatic duct, 5-bile duct, 6-cysts, 7-renal vein")
+    @IO.Output('segmentation', 'segmentation.mha', 'mha:mod=seg:src=cleaned:model=GCNNUnetPancreas:roi=VEIN,ARTERY,PANCREAS,PANCREATIC_DUCT,BILE_DUCT', data="in_data",
+               the="cleaned segmentation of the pancreas, with the following classes: "
+                   "0-background, 1-veins, 2-arteries, 3-pancreas, 4-pancreatic duct, 5-bile duct")
+    @IO.OutputData('cancer_likelihood', ProstateCancerLikelihood, the='Case-level pancreatic tumor likelihood. This is equivalent to the maximum of the pancreatic tumor likelihood heatmap.')
+    def task(self, instance: Instance, in_data: InstanceData, heatmap: InstanceData, segmentation_raw: InstanceData, segmentation: InstanceData, cancer_likelihood: ProstateCancerLikelihood, **kwargs) -> None:
+        # Call the PDAC CLI
+        # A CLI was used here to ensure the mhub framework properly captures the nnUNet stdout output
+        cmd = [
+            sys.executable,
+            str(CLI_PATH),
+            in_data.abspath,
+            heatmap.abspath,
+            segmentation_raw.abspath
+        ]
+        self.subprocess(cmd, text=True)
+
+        # Remove cysts and renal vein classes from the original segmentation.
+        # Insufficient training samples were present in the training data for these classes.
+        # Hence, these classes should be omitted from the final output, since these are not
+        # expected to produce reliable segmentations.
+        self.clean_segementation(
+            segmentation_in=segmentation_raw,
+            segmentation_out=segmentation
+        )
+
+        # Extract case-level cancer likelihood
+        cancer_likelihood.value = self.extract_case_level_cancer_likelihood(
+            heatmap=heatmap
+        )
+
+    def clean_segementation(self, segmentation_in: InstanceData, segmentation_out: InstanceData):
+        self.log("Cleaning output segmentation", level="NOTICE")
+        seg_sitk = SimpleITK.ReadImage(segmentation_in.abspath)
+        seg_numpy = SimpleITK.GetArrayFromImage(seg_sitk)
+        seg_numpy[seg_numpy >= 6] = 0  # remove cysts and renal vein segmentation from original segmentation
+        remapped_sitk = SimpleITK.GetImageFromArray(seg_numpy)
+        remapped_sitk.CopyInformation(seg_sitk)
+        SimpleITK.WriteImage(remapped_sitk, segmentation_out.abspath, True)
+
+    def extract_case_level_cancer_likelihood(self, heatmap: InstanceData):
+        self.log("Extracting case-level cancer likelihood", level="NOTICE")
+        heatmap_sitk = SimpleITK.ReadImage(heatmap.abspath)
+        f = SimpleITK.MinimumMaximumImageFilter()
+        f.Execute(heatmap_sitk)
+        cancer_likelihood = f.GetMaximum()
+        assert 0.0 <= cancer_likelihood <= 1.0, "Cancer likelihood value must be in range [0.0, 1.0]"
+        return cancer_likelihood

models/gc_nnunet_pancreas/utils/__init__.py

@@ -0,0 +1 @@
+from .GCNNUnetPancreasRunner import *

models/gc_nnunet_pancreas/utils/cli.py

@@ -0,0 +1,60 @@
+"""
+-------------------------------------------------------------
+GC / MHub - CLI for the GC nnUnet Pancreas Algorithm
+  The model algorith was wrapped in a CLI to ensure
+  the mhub framework is able to properly capture the nnUNet
+  stdout/stderr outputs
+-------------------------------------------------------------
+
+-------------------------------------------------------------
+Author: Sil van de Leemput
+Email:  [email protected]
+-------------------------------------------------------------
+"""
+import argparse
+from pathlib import Path
+
+# Import the algorithm pipeline class from the CE-CT_PDAC_AutomaticDetection_nnUnet repository
+from process import PDACDetectionContainer
+
+
+def run_pdac_detection(
+    input_ct_image: Path, output_heatmap: Path, output_segmentation: Path
+):
+    # Configure the algorithm pipeline class and run it
+    algorithm = PDACDetectionContainer(output_raw_heatmap=True)
+    algorithm.ct_image = input_ct_image
+    algorithm.heatmap_raw = output_heatmap
+    algorithm.segmentation = output_segmentation
+    algorithm.process()
+
+
+def run_pdac_detection_cli():
+    parser = argparse.ArgumentParser("CLI for the GC nnUNet Pancreas Algorithm")
+    parser.add_argument(
+        "input_ct_image",
+        type=str,
+        help="input CT scan (MHA)"
+    )
+    parser.add_argument(
+        "output_heatmap",
+        type=str,
+        help="raw heatmap of the pancreatic tumor likelihood (MHA)",
+    )
+    parser.add_argument(
+        "output_segmentation",
+        type=str,
+        help="segmentation map of the pancreas (MHA), with the following classes: "
+        "0-background, 1-veins, 2-arteries, 3-pancreas, 4-pancreatic duct, 5-bile duct, "
+        "6-cysts, 7-renal vein",
+    )
+    args = parser.parse_args()
+    run_pdac_detection(
+        input_ct_image=Path(args.input_ct_image),
+        output_heatmap=Path(args.output_heatmap),
+        output_segmentation=Path(args.output_segmentation),
+    )
+
+
+if __name__ == "__main__":
+    run_pdac_detection_cli()