saieswarreddy/glioma-progression-prediction

🧠 Glioma Tumor Progression Prediction using 3D CNN-LSTM

Spatiotemporal Modeling of Post-Treatment MRI for Tumor Progression Risk Estimation

Overview

This project predicts patient-specific probabilities of glioma tumor progression within predefined temporal boundaries (6 and 12 months) by modeling the tumor's spatiotemporal evolution from longitudinal post-treatment MRI scans.

Architecture

Longitudinal MRI Scans (T1, T1c, T2, FLAIR) + Tumor Masks
           │
           ▼
      Preprocessing
 (Z-score Normalization, Resampling to 128³,
  RAS+ Reorientation, Channel Stacking)
           │
           ▼  [Per Timepoint]
     3D ResNet-18
  Spatial Feature Extraction → 512-dim embedding
           │
           ▼  [Sequence of Timepoints]
     Bidirectional LSTM
  Temporal Tumor Evolution Modeling
           │
           ▼
   Concatenate with Clinical Features
           │
           ▼
     FC Prediction Head
 (Dropout → Linear → ReLU → Linear)
           │
           ▼
 Tumor Progression Probability
 P(progression ≤ time horizon)

Dataset

MU-Glioma-Post from The Cancer Imaging Archive (TCIA)

• 203 glioma patients with 596 longitudinal MRI timepoints
• Up to 6 follow-up scans per patient
• 4 MRI modalities: T1 native, T1 contrast (T1c), T2-weighted, FLAIR
• nnU-Net segmentation masks with 4 tumor sub-regions:
  • Label 1: Non-Enhancing Tumor Core (NETC)
  • Label 2: Surrounding Non-enhancing FLAIR Hyperintensity (SNFH)
  • Label 3: Enhancing Tissue (ET)
  • Label 4: Resection Cavity (RC)
• Clinical metadata (demographics, molecular markers, treatment details)

Source: TCIA Collection | HuggingFace Mirror

Notebook Contents

Section	Description
1. Setup	Install dependencies, GPU check
2. Data Download	Download from HuggingFace Hub (11.1 GB)
3. Dataset Exploration	Directory structure analysis, timepoint distributions
4. Clinical Analysis	Demographics, tumor grades, molecular markers
5. MRI Visualization	Multi-modal views, tumor segmentation overlays, temporal evolution
6. Preprocessing	Z-score normalization, resampling, channel stacking
7. Feature Engineering	Volumetric features, Pearson correlation (threshold=0.75), RF importance, PCA
8. Model Architecture	3D ResNet-18 + Bidirectional LSTM definition
9. Training	5-fold patient-level CV, weighted loss, SMOTE, early stopping
10. Evaluation	ROC curves, confusion matrix, classification report
11. Grad-CAM	3D Grad-CAM for model interpretability

Training Configuration

Parameter	Value	Source
Learning Rate	1e-4	arXiv:2504.18268
Weight Decay	0.01	arXiv:2502.03999
Batch Size	1 (effective 4 via accumulation)	GPU memory constraint
Early Stopping	Patience=10	arXiv:2504.18268
LR Schedule	ReduceLROnPlateau (factor=0.1, patience=5)	Standard practice
Loss	CrossEntropy with inverse-prevalence class weights	Class imbalance
Cross-Validation	5-fold stratified, patient-level	Standard for medical imaging
Input Resolution	96×96×96 or 128×128×128	Memory-dependent

Key References

1. Jang et al. (2018) — CNN-LSTM for pseudoprogression vs progression, Scientific Reports, DOI: 10.1038/s41598-018-31007-2
2. MU-Glioma-Post dataset — Nature Scientific Data 2025, DOI: 10.1038/s41597-025-06011-7
3. Self-supervised ViT (SOTA) — arXiv:2502.03999, AUC=0.753
4. LUMIERE/DenseNet — arXiv:2504.18268, RANO classification benchmark
5. Grad-CAM — Selvaraju et al., arXiv:1610.02391
6. 2D-Mamba+CNN hybrid — PMC:12292999, best accuracy-efficiency tradeoff

Requirements

nibabel
huggingface_hub
openpyxl
monai
scikit-learn
seaborn
imbalanced-learn
scipy
torch
torchvision

License

This project uses the MU-Glioma-Post dataset which is available under the TCIA Data Usage Policy. The code is provided for research purposes.