This repository contains the official implementation of our hybrid multimodal survival prediction framework for Non-Small Cell Lung Cancer (NSCLC), which — for the first time — jointly leverages:
-
3D Masked Autoencoder (MAE) — a self-supervised Vision Transformer pretrained on pretreatment chest CT without manual annotation, producing deep imaging embeddings that capture latent morphological and textural patterns beyond hand-crafted radiomic features.
-
Tumor Growth Simulation Framework — a patient-specific simulator whose proliferation and necrosis parameters are calibrated directly from interpretable radiomic descriptors (voxel intensity entropy and surface sphericity), yielding biologically grounded summary markers not observable from a single static scan.
These complementary signals are fused with conventional radiomic features and clinical covariates through Cox-based, Neural Cox, and ensemble survival learners for NSCLC prognosis.
| Metric | SOTA (Ferretti & Corino) | Ours (Primary) | Ours (Exploratory) |
|---|---|---|---|
| C-index | 0.631 | 0.641 | 0.662 |
| Integrated AUC | 0.592 | 0.731 | 0.748 |
| Hazard Ratio (HR) | — | 1.95 | 2.21 |
| Log-rank p-value | 0.031 | < 0.001 | < 0.001 |
## Installation
```bash
git clone https://github.com/your-username/nsclc-survival-fusion.git
cd nsclc-survival-fusion
pip install -r requirements.txt
torch>=1.13.0
numpy
pandas
pydicom
pyradiomics
lifelines
scikit-learn
scipy
scikit-image
optuna
joblib
tqdm
matplotlib
plotly
This study uses the NSCLC-Radiomics (Lung1) dataset, publicly available through The Cancer Imaging Archive (TCIA):
To download the dataset, follow the instructions on the TCIA website.
Expected directory structure:
data/
└── NSCLC-Radiomics/
└── LUNG1-001/
└── <study_date>/
├── <CT_series>/
│ ├── *.dcm
└── 300.<SEG_series>/
└── *.dcm
This project is licensed under the MIT License. See LICENSE for details.