BAMF MR Brain tumor segmentation

models/bamf_mr_brain_tumor/config/default.yml

@@ -0,0 +1,87 @@
+general:
+  data_base_dir: /app/data
+  version: 1.0
+  description: Default configuration for Bamf NNUnet Brain tumor segmentation on MR scans (dicom to dicom)
+
+execute:
+- FileStructureImporter
+- NiftiConverter
+- ReOrientationRunner
+- BiasCorrectionRunner
+- FLIRTRegistrationRunner
+- SkullStripRunner
+- StdRegistrationRunner
+- NNUnetRunnerV2
+- InverseStdRegistrationRunner
+- InverseRegistrationRunner
+- module: DsegConverter
+  target_dicom: dicom:mod=mr:type=t1
+  source_segs: nifti:mod=mr:type=t1:task=inverse
+- module: DsegConverter
+  target_dicom: dicom:mod=mr:type=t1ce
+  source_segs: nifti:mod=mr:type=t1ce:task=inverse
+- module: DsegConverter
+  target_dicom: dicom:mod=mr:type=t2
+  source_segs: nifti:mod=mr:type=t2:task=inverse
+- module: DsegConverter
+  target_dicom: dicom:mod=mr:type=flair
+  source_segs: nifti:mod=mr:type=flair:task=inverse
+- DataOrganizer
+
+modules:
+  FileStructureImporter:
+    input_dir: 'input_data'
+    structures:
+      - $patientID@instance/t1@dicom:mod=mr:type=t1
+      - $patientID/t1ce@dicom:mod=mr:type=t1ce
+      - $patientID/t2@dicom:mod=mr:type=t2
+      - $patientID/flair@dicom:mod=mr:type=flair
+    import_id: patientID
+
+  NiftiConverter:
+    allow_multi_input: True
+    out_datas: nifti:mod=mr:task=conversion
+    engine: dcm2niix
+
+  ReOrientationRunner:
+    in_datas: nifti:mod=mr
+
+  BiasCorrectionRunner:
+    in_datas: nifti:mod=mr:task=reorientation
+
+  FLIRTRegistrationRunner:
+    in_datas: nifti:mod=mr:task=bias_corrected
+    reference_data: nifti:mod=mr:task=bias_corrected:type=t1ce
+
+  SkullStripRunner:
+    in_datas: nifti:mod=mr:task=registration
+
+  StdRegistrationRunner:
+    in_datas: nifti:mod=mr:task=skull_stripped
+
+  NNUnetRunnerV2:
+    in_t1_data: nifti:mod=mr:task=std_registration:type=t1
+    in_t1ce_data: nifti:mod=mr:task=std_registration:type=t1ce
+    in_t2_data: nifti:mod=mr:task=std_registration:type=t2
+    in_flair_data: nifti:mod=mr:task=std_registration:type=flair
+
+  InverseStdRegistrationRunner:
+    in_seg_data: nifti:mod=seg:model=nnunet:nnunet_dataset=Dataset002_BRATS19
+    in_mat_datas: txt:task=std_registration_transform_mat
+    in_registration_datas: nifti:mod=mr:task=registration
+
+  InverseRegistrationRunner:
+    in_seg_datas: nifti:mod=mr:task=std_inverse
+    in_mat_datas: txt:task=registration_transform_mat
+    in_registration_datas: nifti:mod=mr:task=bias_corrected
+
+  DsegConverter:
+    model_name: BAMF Brain Tumor AI Segmentation
+    skip_empty_slices: True
+
+  DataOrganizer:
+    targets:
+    - dicomseg:type=t1ce-->[i:patientID]/t1ce.seg.dcm
+    - dicomseg:type=t1-->[i:patientID]/t1.seg.dcm
+    - dicomseg:type=t2-->[i:patientID]/t2.seg.dcm
+    - dicomseg:type=flair-->[i:patientID]/flair.seg.dcm

models/bamf_mr_brain_tumor/dockerfiles/Dockerfile

@@ -0,0 +1,78 @@
+FROM mhubai/base:latest
+
+ENV DEBIAN_FRONTEND "noninteractive"
+ENV LANG            "en_GB.UTF-8"
+
+# 1. Install gcc-14 and build ants
+RUN apt-get update && apt-get install -y cmake make ninja-build git bzip2 flex manpages-dev g++ wget unzip file
+RUN wget https://github.com/gcc-mirror/gcc/archive/refs/tags/releases/gcc-11.4.0.zip \
+    && unzip gcc-11.4.0.zip && cd gcc-releases-gcc-11.4.0 && ./contrib/download_prerequisites \
+    && mkdir /app/gcc-build && cd /app/gcc-build \
+    && ../gcc-releases-gcc-11.4.0/configure -v --target=x86_64-linux-gnu --prefix=/usr/local/gcc-11.4.0 --enable-checking=release --enable-languages=c,c++ --disable-multilib --program-suffix=-11.4 \
+    && make -j$(nproc) && make install-strip \
+    && update-alternatives --install /usr/bin/gcc gcc /usr/local/gcc-11.4.0/bin/gcc-11.4 11 --slave /usr/bin/g++ g++ /usr/local/gcc-11.4.0/bin/g++-11.4 \
+    && rm -rf /app/gcc-*
+
+ENV PATH=/usr/local/gcc-11.4.0/bin:$PATH
+ENV LD_LIBRARY_PATH="/usr/local/gcc-11.4.0/lib64:$LD_LIBRARY_PATH"
+
+RUN git clone https://github.com/ANTsX/ANTs.git /usr/local/src/ants \
+    && mkdir /app/build && cd /app/build && cmake -DBUILD_TESTING=ON \
+    -DRUN_LONG_TESTS=OFF \
+    -DRUN_SHORT_TESTS=ON \
+    -DBUILD_SHARED_LIBS=OFF \
+    -DCMAKE_INSTALL_PREFIX=/opt/ants \
+    /usr/local/src/ants \
+    && cmake --build . --parallel \
+    && cd /app/build/ANTS-build && cmake --build . --target test \
+    && cmake --install . && rm -rf /app/build
+
+# 2. Install mri_convert
+ENV FREESURFER_HOME="/freesurfer"
+RUN git clone https://github.com/freesurfer/freesurfer.git && cd freesurfer/mri_synthstrip \
+    git+https://github.com/freesurfer/surfa.git@0d83332351083b33c4da221e9d10a63a93ae7f52 \
+    && mkdir -p $FREESURFER_HOME/models/ \
+    && git remote add datasrc https://surfer.nmr.mgh.harvard.edu/pub/dist/freesurfer/repo/annex.git \
+    && apt-get update -y && apt-get install -y git-annex \
+    && git fetch datasrc && git-annex get . \
+    && cp mri_synthstrip $FREESURFER_HOME \
+    && cp synthstrip.*.pt $FREESURFER_HOME/models/
+
+# 3. Install fsl
+RUN wget https://fsl.fmrib.ox.ac.uk/fsldownloads/fslconda/releases/fslinstaller.py && \
+    python3 ./fslinstaller.py -V 6.0.7.11 -d /usr/local/fsl/ && rm fslinstaller.py
+
+ENV PATH="/opt/ants/bin:$PATH:/$FREESURFER_HOME:/usr/local/fsl/bin" \
+LD_LIBRARY_PATH="/opt/ants/lib:$LD_LIBRARY_PATH"
+
+# 3. Install nnunet
+# FIXME: set this environment variable as a shortcut to avoid nnunet 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
+
+# 4. Install nnunet and surfa
+RUN uv pip install torch==2.4.1 nnunetv2==2.4.1 surfa==0.6.1
+RUN apt-get update && apt-get install -y bc
+
+# Clone the main branch of MHubAI/models
+ARG MHUB_MODELS_REPO
+RUN buildutils/import_mhub_model.sh bamf_mr_brain_tumor ${MHUB_MODELS_REPO}
+
+# Pull nnUNet model weights into the container
+ENV WEIGHTS_DIR=/root/.nnunet/nnUNet_models/
+RUN mkdir -p $WEIGHTS_DIR
+ENV WEIGHTS_FN=Dataset002_BRATS19.zip
+ENV WEIGHTS_URL=https://zenodo.org/records/11582627/files/$WEIGHTS_FN
+RUN wget --directory-prefix ${WEIGHTS_DIR} ${WEIGHTS_URL}
+RUN unzip ${WEIGHTS_DIR}${WEIGHTS_FN} -d ${WEIGHTS_DIR}
+RUN rm ${WEIGHTS_DIR}${WEIGHTS_FN}
+
+# specify nnunet specific environment variables
+ENV WEIGHTS_FOLDER=$WEIGHTS_DIR
+
+# Default run script
+ENTRYPOINT ["mhub.run"]
+CMD ["--config", "/app/models/bamf_mr_brain_tumor/config/default.yml"]

models/bamf_mr_brain_tumor/meta.json

@@ -0,0 +1,211 @@
+{
+  "id": "",
+  "name": "bamf_mr_brain_tumor",
+  "title": "BAMF MR Brain Tumor",
+  "summary": {
+    "description": "An nnU-Net based model to segment brain tumors using four MR modalities Fluid Attenuated Inversion Recovery (FLAIR), native (T1), post-contrast T1-weighted (T1Gd), and T2-weighted (T2)",
+    "inputs": [
+      {
+        "label": "Input Image",
+        "description": "native (T1) MRI scan of a patient.",
+        "format": "DICOM",
+        "modality": "MRI",
+        "bodypartexamined": "BRAIN",
+        "slicethickness": "1mm",
+        "non-contrast": true,
+        "contrast": false
+      },
+      {
+        "label": "Input Image",
+        "description": "post-contrast T1-weighted MRI scan of a patient.",
+        "format": "DICOM",
+        "modality": "MRI",
+        "bodypartexamined": "BRAIN",
+        "slicethickness": "1mm",
+        "non-contrast": false,
+        "contrast": true
+      },
+      {
+        "label": "Input Image",
+        "description": "T2-weighted MRI scan of a patient.",
+        "format": "DICOM",
+        "modality": "MRI",
+        "bodypartexamined": "BRAIN",
+        "slicethickness": "1mm",
+        "non-contrast": true,
+        "contrast": false
+      },
+      {
+        "label": "Input Image",
+        "description": "T2 Fluid Attenuated Inversion Recovery (T2-FLAIR) MRI scan of a patient.",
+        "format": "DICOM",
+        "modality": "MRI",
+        "bodypartexamined": "BRAIN",
+        "slicethickness": "1mm",
+        "non-contrast": true,
+        "contrast": false
+      }
+    ],
+    "outputs": [
+      {
+        "label": "T1 Segmentation",
+        "type": "Segmentation",
+        "description": "Tumor Segmentation of T1 MRI scan",
+        "classes": [
+          "NECROSIS",
+          "EDEMA",
+          "ENHANCING_LESION"
+        ]
+      },
+      {
+        "label": "T1ce Segmentation",
+        "type": "Segmentation",
+        "description": "Tumor Segmentation of T1ce MRI scan",
+        "classes": [
+          "NECROSIS",
+          "EDEMA",
+          "ENHANCING_LESION"
+        ]
+      },
+      {
+        "label": "T2 Segmentation",
+        "type": "Segmentation",
+        "description": "Tumor Segmentation of T2 MRI scan",
+        "classes": [
+          "NECROSIS",
+          "EDEMA",
+          "ENHANCING_LESION"
+        ]
+      },
+      {
+        "label": "FLAIR Segmentation",
+        "type": "Segmentation",
+        "description": "Tumor Segmentation of FLAIR MRI scan",
+        "classes": [
+          "NECROSIS",
+          "EDEMA",
+          "ENHANCING_LESION"
+        ]
+      }
+    ],
+    "model": {
+      "architecture": "U-net",
+      "training": "supervised",
+      "cmpapproach": "3D"
+    },
+    "data": {
+      "training": {
+        "vol_samples": 1251
+      },
+      "evaluation": {
+        "vol_samples": 45
+      },
+      "public": true,
+      "external": true
+    }
+  },
+  "details": {
+    "name": "AIMI MR Brain Tumor",
+    "version": "1.0.0",
+    "devteam": "BAMF Health",
+    "authors": [
+      "Soni, Rahul",
+      "McCrumb, Diana",
+      "Murugesan, Gowtham Krishnan",
+      "Van Oss, Jeff",
+      "Jithendra Kumar"
+    ],
+    "type": "nnU-Net (U-Net structure, optimized by data-driven heuristics)",
+    "date": {
+      "code": "30.08.2024",
+      "weights": "11.06.2024",
+      "pub": "30.08.2024"
+    },
+    "cite": "Gowtham Krishnan Murugesan, Diana McCrumb, Rahul Soni, Jithendra Kumar, Leonard Nuernberg, Linmin Pei, Ulrike Wagner, Sutton Granger, Andrey Y. Fedorov, Stephen Moore, Jeff Van Oss. AI generated annotations for Breast, Brain, Liver, Lungs and Prostate cancer collections in National Cancer Institute Imaging Data Commons. arXiv:2409.20342  (2024).",
+    "license": {
+      "code": "MIT",
+      "weights": "CC BY-NC 4.0"
+    },
+    "publications": [
+      {
+        "title": "AI generated annotations for Breast, Brain, Liver, Lungs and Prostate cancer collections in National Cancer Institute Imaging Data Commons",
+        "uri": "https://arxiv.org/abs/2409.20342"
+      }
+    ],
+    "github": "https://github.com/bamf-health/aimi-brain-mr"
+  },
+  "info": {
+    "use": {
+      "title": "Intended Use",
+      "text": "To perform tumor segmentation in Brain MRI scans, you need four MRI modalities—T1-weighted, T1 with contrast enhancement (T1ce), T2-weighted, and FLAIR—each providing critical information about brain anatomy and tumor characteristics. The data should be organized in separated directories per instance (e.g., named by a unique sample or patient identifier) with one subfolder for each of the four input modalities, containing the respective DICOM files (namely 't1', 't1ce', 't2', and 'flair'). The model processes these inputs to produce three segmentation maps for necrosis, edema, and enhancing tumor regions. These outputs are vital for clinical applications like treatment planning, monitoring tumor progression, and prognosis, as well as for research projects focused on developing new algorithms, identifying imaging biomarkers, and integrating automated segmentation into clinical workflows"
+    },
+    "analyses": {
+      "title": "Quantitative Analyses",
+      "text": "The model's performance was assessed using the Dice Coefficient and Normalized Surface Distance (NSD) with tolerance 7mm. The analysis is published here [2]",
+      "tables": [
+        {
+          "label": "Label-wise metrics Dice score between AI derived and manually corrected annotations",
+          "entries": {
+            "Whole Tumor": "0.98±0.07",
+            "Enhancing Tumor": "0.95±0.13",
+            "Non enhancing Tumor": "0.97±0.08"
+          }
+        },
+        {
+          "label": "Label-wise metrics 95% Hausdorff Distance between AI derived and manually corrected annotations",
+          "entries": {
+            "Whole Tumor": "6.88±0.34",
+            "Enhancing Tumor": "6.57±0.24",
+            "Non enhancing Tumor": "0.42±1.05"
+          }
+        },
+        {
+          "label": "Label-wise metrics NSD score between AI derived and manually corrected annotations",
+          "entries": {
+            "Whole Tumor": "0.98±0.04",
+            "Enhancing Tumor": "0.99±0.03",
+            "Non enhancing Tumor": "0.97±0.09"
+          }
+        }
+      ],
+      "references": [
+        {
+          "label": "UPENN-GBM",
+          "uri": "https://www.cancerimagingarchive.net/collection/upenn-gbm/"
+        },
+        {
+          "label": "AI generated annotations for Breast, Brain, Liver, Lungs and Prostate cancer collections in National Cancer Institute Imaging Data Commons",
+          "uri": "https://arxiv.org/abs/2409.20342"
+        }
+      ]
+    },
+    "evaluation": {
+      "title": "Evaluation Data",
+      "text": "The model was used to segment 541 cases from the IDC [1] collection UPENN-GBM [1]. Approximately 10% of these annotations, 45 cases were randomly selected to be reviewed and corrected by a board-certified radiologist. Quality metrics such as Dice coefficient, normalized surface distance (NSD), and detection accuracy were reported",
+      "references": [
+        {
+          "label": "Imaging Data Collections (IDC)",
+          "uri": "https://datacommons.cancer.gov/repository/imaging-data-commons"
+        },
+        {
+          "label": "UPENN-GBM",
+          "uri": "https://www.cancerimagingarchive.net/collection/upenn-gbm/"
+        }
+      ]
+    },
+    "training": {
+      "title": "Training Data",
+      "text": "The training dataset used from the BraTS21 challenge consists of 1,251 brain mpMRI scans along with segmentation annotations of tumorous regions. The 3D volumes were skull-stripped and resampled to 1 mm isotropic resolution, with dimensions of (240, 240, 155) voxels. For each example, four modalities were given: Fluid Attenuated Inversion Recovery (FLAIR), native (T1), post-contrast T1-weighted (T1Gd), and T2-weighted (T2). Annotations consisted of four classes: 1 for necrotic tumor core (NCR), 2 for peritumoral edematous tissue (ED), 4 for enhancing tumor (ET), and 0 for background (voxels that are not part of the tumor).",
+      "references": [
+        {
+          "label": "Brain Tumor Segmentation (BraTS)",
+          "uri": "http://braintumorsegmentation.org"
+        }
+      ]
+    },
+    "limitations": {
+      "title": "Limitations",
+      "text": "The model has been trained and tested on scans acquired during clinical care of patients, so it might not be suited for a healthy population. The generalization capabilities of the model on a range of ages, genders, and ethnicities are unknown."
+    }
+  }
+}

models/bamf_mr_brain_tumor/mhub.toml

@@ -0,0 +1,3 @@
+
+[model.deployment]
+test = "https://zenodo.org/records/13880663/files/bamf_mr_brain_tumor.test.zip?download=1"