An exploratory radiogenomics study for predicting IDH mutation status in glioma by integrating multi-modal brain MRI and somatic mutation profiles with a heterogeneous graph neural network.
MRI 영상 표현과 체세포 변이 정보를 이종 그래프로 통합해 교종의 IDH 변이 상태를 탐색적으로 예측한 의료 AI 연구 포트폴리오입니다.
This public repository is intentionally documentation-first. Research code, patient-level data, learned weights, and raw prediction tables are kept private.
| Item | Description |
|---|---|
| Task | Exploratory binary prediction of IDH mutation status |
| Cohort | 59 matched imaging–genomics subjects |
| Inputs | T1, T1ce, T2, FLAIR MRI and 104 non-target mutation features |
| Model | Heterogeneous graph neural network with typed relations |
| Evaluation | Stratified 5-fold cross-validation with multi-seed ablations |
| Public scope | Methodology, architecture, aggregate results, and limitations |
- Constructed the matched imaging–genomics study design and target-leakage exclusions.
- Designed patient, tumor-region, and gene node types with anatomical, genomic, and cross-modal relations.
- Compared directional message passing, gene encodings, attention mechanisms, and residual MRI relations.
- Applied fold-specific preprocessing and multi-seed cross-validation for the principal ablations.
- Summarized aggregate performance and limitations without exposing patient-level data.
Can an anatomy-aware graph represent interactions between tumor regions and somatic mutations more effectively than direct multimodal feature concatenation?
The model organizes the problem into:
- Patient nodes for subject-level context
- Tumor-region nodes for core, edema, and enhancing tumor
- Gene nodes for 104 non-target somatic mutation features
- Typed relations for anatomical, genomic, and cross-modal interactions
MRI representations use T1, T1ce, T2, and FLAIR modalities. The target is the
binary IDH_label; IDH-related and clinical target columns are excluded from
the input features.
- 59 matched imaging-genomics subjects
- Stratified 5-fold cross-validation
- Multi-seed evaluation for the principal ablations
- Fold-specific standardization and dimensionality reduction
- Class-balanced training objectives
- Comparisons across directional fusion, gene encoding, attention, and residual-relation variants
| Configuration | Mean ROC-AUC | Evaluation |
|---|---|---|
| Learned 32-d gene embedding + mutation rate | 0.912 ± 0.196 | 10 seeds × 5 folds |
| Gene-conditioned region summary fusion | 0.884 | 10 seeds × 5 folds |
| MRI→gene directional exchange | 0.882 | 3 seeds × 5 folds |
| Learnable residual MRI edges | 0.852 | 10 seeds × 5 folds |
The comparison suggests that learned gene representations and directional MRI-to-gene information flow may be useful hypotheses for multimodal graph design. Because the cohort is small and fold-level variance is substantial, these results are not estimates of clinical deployment performance.
| Experiment | Purpose |
|---|---|
| Regional cross-modal HGT | Represent tumor regions and genes as typed nodes |
| Directional message passing | Compare MRI→gene, gene→MRI, bidirectional, and no exchange |
| Gene representation ablation | Compare one-hot and learned gene embeddings |
| Mutation-rate ablation | Test cohort mutation rate as a gene-node feature |
| Gene-conditioned attention | Learn genomic-context-dependent regional weighting |
| Residual MRI relations | Compare learned and fixed residual regional edges |
| Raw MRI / MedicalNet | Explore 3D CNN, regional, patch, and supervoxel encoders |
| Pathway-conditioned HGT | Route mutations through exploratory pathway groups |
The study uses locally prepared matched imaging-genomics tables derived from brain MRI and somatic mutation sources. Source and derived subject-level data are not redistributed due to access conditions, licensing, and research-data privacy considerations.
- The matched cohort contains only 59 subjects.
- External validation has not yet been performed.
- Raw-MRI variants remain computational research prototypes.
- The exploratory pathway mapping requires validation against a cited pathway database before publication.
- Aggregate results require reproduction in an independent clean environment.
.
├── README.md
├── figures/
│ ├── architecture.png
│ └── results.png
├── LICENSE
└── .gitignore
Additional implementation details can be discussed for academic collaboration. No code or data in this project should be used for clinical decision-making.
The public documentation is released under the MIT License. Dataset licenses and third-party model licenses are not covered by this repository.