ClinGRAD - Clinically-Guided GNN for Dementia Diagnosis

ClinGRAD is a clinically-guided heterogeneous graph neural network for dementia diagnosis. It combines MRI-derived radiomics, gene-expression profiles, clinical features, structural brain connectivity priors, and gene co-expression edges to classify subjects as Control (CTL), Mild Cognitive Impairment (MCI), Alzheimer's Disease (AD), or Vascular Dementia (VaD).

The training code is available in this repository. The main entry points are:

• src/train.py for model training.
• src/infer.py for checkpoint evaluation/inference.
• src/data/dataset.py for the expected tabular input format.
• config/default.yaml for default paths and hyperparameters.

Paper: https://papers.miccai.org/miccai-2025/0148-Paper5205.html

Architecture

ClinGRAD models each subject as a heterogeneous graph with three node groups:

• patient nodes containing clinical/cognitive features;
• brain-structure nodes containing MRI radiomics features;
• gene nodes containing expression values.

The paper describes clinically informed patient, structure, and gene relationships. In the released implementation, the loader constructs parent-to-brain links, structure-to-structure links derived from anatomical distance and DWI connectivity priors, and gene-to-gene links derived from co-expression scores; patient, brain, and gene embeddings are then fused at the graph level.

Data Availability and Preprocessing

The model was developed using ANMerge. ANMerge access is controlled by Synapse, and this repository does not redistribute ANMerge data, derived masks, processed radiomics tables, gene-expression tables, or subject-level clinical data.

ANMerge does not include the 32 brain-structure masks used by ClinGRAD. As described in the paper, we generated these masks from the MRI scans with SynthSeg and then used the masks to extract radiomics features with PyRadiomics. These preprocessing tools are external to this repository because their dependencies and CUDA/GPU support change regularly and can introduce environment-specific installation issues.

Relevant preprocessing references:

• ClinGRAD paper: https://papers.miccai.org/miccai-2025/0148-Paper5205.html
• SynthSeg GitHub repository: https://github.com/BBillot/SynthSeg
• SynthSeg paper: https://doi.org/10.1016/j.media.2023.102789
• MATLAB SynthSeg example: https://www.mathworks.com/help/medical-imaging/ug/Brain-MRI-Segmentation-Using-Trained-3-D-U-Net.html
• PyRadiomics documentation: https://pyradiomics.readthedocs.io/
• PyRadiomics GitHub repository: https://github.com/Radiomics/pyradiomics
• PyRadiomics paper: https://doi.org/10.1158/0008-5472.CAN-17-0339
• ANMerge paper: https://doi.org/10.3233/JAD-200948
• ANMerge Synapse DOI: https://doi.org/10.7303/syn22252881
• DWI structural-connectivity matrices: https://doi.org/10.1038/s41597-022-01596-9
• DWI matrices and processing scripts on OSF: https://osf.io/yw5vf/
• BrainLat dataset paper: https://doi.org/10.1038/s41597-023-02806-8
• GeneMANIA co-expression reference: https://doi.org/10.1093/nar/gkq537
• Human Protein Atlas: https://www.proteinatlas.org/

Preprocessing Workflow

The released repository starts from harmonized CSV inputs. The preprocessing steps used to create those inputs are:

1. Generate Brain-Structure Masks

Start from the ANMerge MRI scans available through your approved Synapse access. ANMerge does not provide the 32 brain-structure masks required by ClinGRAD.

We generated those masks with SynthSeg:

• GitHub: https://github.com/BBillot/SynthSeg
• Paper: https://doi.org/10.1016/j.media.2023.102789
• MATLAB example using the pretrained SynthSeg network: https://www.mathworks.com/help/medical-imaging/ug/Brain-MRI-Segmentation-Using-Trained-3-D-U-Net.html

The SynthSeg repository provides command-line usage such as:

python ./scripts/commands/SynthSeg_predict.py \
  --i <input_mri_or_folder> \
  --o <output_segmentation_or_folder>

MATLAB users can follow the MathWorks SynthSeg example linked above. That example demonstrates loading the pretrained SynthSeg network and segmenting a brain MRI into 32 anatomical classes. This repository does not include the SynthSeg model, MATLAB files, or segmentation outputs.

The output of this step is one segmentation mask per MRI scan. Keep a consistent mapping between SynthSeg label IDs and the structure names used later in patient_structure.

2. Extract Radiomics with PyRadiomics

Use the MRI scan and its SynthSeg mask to extract radiomics features for each of the 32 structures.

• GitHub: https://github.com/Radiomics/pyradiomics
• Documentation: https://pyradiomics.readthedocs.io/
• Paper: https://doi.org/10.1158/0008-5472.CAN-17-0339

The paper uses 107 PyRadiomics features per structure. Store the output in long format, with one row per patient and structure, then build the patient_structure field as:

<patient_id>_<structure_name>

This radiomics table becomes radiomics.csv.

3. Build DWI Structure-Connectivity Priors

The DWI prior referenced in the paper comes from:

• Skoch et al., "Human brain structural connectivity matrices-ready for modelling": https://doi.org/10.1038/s41597-022-01596-9
• OSF data and scripts: https://osf.io/yw5vf/

That resource provides DWI-derived structural-connectivity matrices as both individual CSV files and a MATLAB-readable .mat file. The OSF project also contains the processing scripts used by the original DWI paper, including the steps for DWI preprocessing, registration, tractography, and matrix creation. The script sequence described by the DWI resource is:

1_dti_preprocess
2_dti_reg
3_dti_track2
4_dti_get_conn2

If using the MATLAB-readable file from OSF, the conversion step is:

S = load("SCmatrices88healthy.mat");
% Select or average the structural-connectivity matrices from S.
% Map or aggregate the source atlas regions to the 32 SynthSeg structures.
% Store the resulting 32 x 32 matrix.
writematrix(dwi32, "DWI_Matrix.csv")

For ClinGRAD, convert the DWI connectivity resource into a 32 x 32 matrix that matches the same 32 SynthSeg structures used for radiomics. The original DWI resource uses its own atlas/parcellation, so this step requires mapping or aggregating the available regions to your 32-structure ordering. Save the final matrix as DWI_Matrix.csv.

The paper also reports validation/refinement using BrainLat DWI scans:

• BrainLat dataset paper: https://doi.org/10.1038/s41597-023-02806-8

4. Prepare Clinical and Gene Tables

Prepare the remaining model inputs from ANMerge and external gene resources:

• patient-level clinical, demographic, and cognitive-assessment features for patient_data.csv; in our experiments this was derived from ANMerge_clinical_under_90.csv;
• expression values for the 75 AD-associated genes selected from the Human Protein Atlas for gene_expression.csv; in our experiments this was derived from ANMerge_gene_expression_normalized_under_90.csv, filtered to the panel listed in resources/gene_panel_75.txt;
• GeneMANIA-derived co-expression edges for co_expression.csv.

Expected Input Files

The default file locations are defined in config/default.yaml:

paths:
  output_dir: ./outputs
  patient_csv: ./data/patient_data.csv
  radiomics_csv: ./data/radiomics.csv
  dwi_csv: ./data/DWI_Matrix.csv
  distance_csv: ./data/structure_distance.csv
  gene_expression_csv: ./data/gene_expression.csv
  coexpression_csv: ./data/co_expression.csv

The released code expects harmonized intermediate CSV files, not the raw ANMerge release tables. The current loader accepts the following schemas.

Patient and Clinical Features

patient_data.csv aligns patients across modalities. It must include a Key column. If DIAGNOSIS, Subtype, diagnosis, or subtype is present, the loader uses it as the label column. All columns after the first two columns are encoded as patient-level clinical features. Numeric columns are z-scored and categorical columns are one-hot encoded.

The paper uses patient nodes for clinical markers, including cognitive scores. For ANMerge-derived experiments, these features should come from the available clinical, demographic, and cognitive-assessment fields after applying the appropriate cohort harmonization and missing-value handling. In our experiments, patient_data.csv was derived from ANMerge_clinical_under_90.csv.

Minimal harmonized example. Age, Sex, MMSE, CDR, Global_Deterioration_Scale, and APOE are illustrative feature names, not a complete list of raw ANMerge clinical columns:

Key,DIAGNOSIS,Age,Sex,MMSE,CDR,Global_Deterioration_Scale,APOE
PRGCTL025,0,73,F,29,0,1,E3/E3
PRGAD014,2,78,M,19,1,4,E3/E4

Alternatively, labels may be supplied in a separate labels file with:

PTID,DIAGNOSIS
PRGCTL025,0
PRGAD014,2

MRI Radiomics Features

radiomics.csv is a long table with one row per patient and brain structure. It must include patient_structure. The value is parsed as:

<patient_id>_<structure_name>

The paper extracts 107 PyRadiomics features per structure from the SynthSeg masks. All columns except patient_structure and optional Label are treated as radiomics features.

Example:

patient_structure,original_firstorder_Mean,original_firstorder_Variance,original_shape_MeshVolume
PRGCTL025_left_hippocampus,42.3,18.5,3210.0
PRGCTL025_right_hippocampus,41.9,17.8,3188.0

If your patient IDs contain underscores, adapt src/data/dataset.py or convert IDs before running the loader, because the current parser splits patient_structure at the first underscore.

Gene Expression Features

gene_expression.csv is a wide table with one row per patient/sample. The ID column must be named one of ptid, patient_id, sample, id, or key case-insensitively. All other columns are interpreted as gene-expression values.

The paper uses 75 AD-associated genes selected from the Human Protein Atlas. The released loader does not enforce a fixed gene list; it uses the non-ID columns present in this CSV as gene nodes. For paper-aligned experiments, use the panel in resources/gene_panel_75.txt and preserve gene names consistently with the co-expression table. In our experiments, gene_expression.csv was derived from ANMerge_gene_expression_normalized_under_90.csv, not the raw expression table.

Example:

patient_id,APP,MAPT,PSEN1,APOE
PRGCTL025,8.42,5.17,3.91,6.28
PRGAD014,9.10,5.88,4.02,7.45

Gene Co-expression Edges

co_expression.csv defines gene-to-gene edges. It must include Gene1, Gene2, and Weight. In the paper, co-expression relationships are derived from GeneMANIA.

Example:

Gene1,Gene2,Weight
APP,PSEN1,0.83
APOE,MAPT,0.57

DWI Connectivity and Structure Distance

DWI_Matrix.csv and structure_distance.csv provide structure-to-structure edge priors for the 32 SynthSeg-derived brain structures.

The DWI prior in the paper is based on published DWI structural-connectivity matrices and was validated/refined with BrainLat DWI scans. The Scientific Data structural-connectivity resource provides MATLAB-readable matrices and processing scripts on OSF. Users should map or aggregate the available connectivity matrix to the 32 structures used in their SynthSeg/radiomics pipeline before passing it to ClinGRAD.

The loader accepts either:

• global 32 x 32 numeric CSV matrices for both DWI_Matrix.csv and structure_distance.csv; or
• a long structure_distance.csv with patient_id, structure_1, structure_2, and distance columns.

Example global matrix shape shown truncated:

0,0.12,0.03
0.12,0,0.08
0.03,0.08,0

Example long distance table:

patient_id,structure_1,structure_2,distance
PRGCTL025,left_hippocampus,right_hippocampus,41.8
PRGCTL025,left_hippocampus,left_amygdala,12.6

Installation

Install the Python dependencies for model training:

pip install -r requirements.txt

SynthSeg, PyRadiomics, MATLAB, FSL, and DWI preprocessing environments are not installed by this repository. Use the linked project documentation for those preprocessing stages.

Training

Edit config/default.yaml so each path points to your harmonized CSV files, or pass paths directly on the command line:

python src/train.py --config config/default.yaml

Example with explicit paths:

python src/train.py \
  --patient_csv data/patient_data.csv \
  --radiomics_csv data/radiomics.csv \
  --dwi_csv data/DWI_Matrix.csv \
  --distance_csv data/structure_distance.csv \
  --gene_expression_csv data/gene_expression.csv \
  --coexpression_csv data/co_expression.csv \
  --num_classes 4

The best checkpoint is saved under:

outputs/checkpoints/best.pt

Notes for Reproduction

• Raw ANMerge files must be converted into the intermediate schemas above before training.
• ANMerge MRI scans do not include the required 32 brain-structure masks; those masks must be generated externally with SynthSeg or an equivalent segmentation workflow.
• Radiomics features should be extracted per structure from each subject's MRI and corresponding segmentation mask.
• Gene-expression and co-expression tables must use matching gene symbols.
• Structure names must be consistent across radiomics rows and any per-patient structure-distance table.
• Data splits and hyperparameters should be matched to the paper for exact reproduction.

Performance Reported in the Paper

• AD vs CTL: 98.75% accuracy
• AD vs MCI: 94.25% accuracy
• MCI vs CTL: 89.66% accuracy
• AD vs VaD: 89.45% accuracy
• Four-class AD/MCI/CTL/VaD: 93.15% accuracy

Citation

If you use ClinGRAD, please cite the MICCAI 2025 paper and the preprocessing resources relevant to your pipeline, including SynthSeg, PyRadiomics, ANMerge, the DWI connectivity resource, BrainLat if used, GeneMANIA, and the Human Protein Atlas.

@InProceedings{HasSal_ClinGRAD_MICCAI2025,
  author = {Hassan, Salma and Salem, Mostafa and Papineni, Vijay Ram Kumar and Elsayed, Ayman and Yaqub, Mohammad},
  title = {ClinGRAD: Clinically-Guided Genomics and Radiomics Interpretable GNN for Dementia Diagnosis},
  booktitle = {Proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
  year = {2025},
  publisher = {Springer Nature Switzerland},
  volume = {LNCS 15971},
  pages = {204--214}
}