prostate-multimodal-survival-main

Multimodal Prostate Cancer Survival Model (MRI + Clinical)

This repository contains outputs from the main multimodal survival model, which integrates both imaging and structured clinical variables.
The model combines:

• T2-weighted MRI
• ADC MRI
• HBV MRI
• Structured clinical variables

MRI features are extracted using a 3D ResNet-18 encoder, while clinical features are processed via a multilayer perceptron (MLP).
The fused representation is optimized with a Cox proportional hazards survival loss.

Contents

• main_model_outputs.zip
  • models/ — best checkpoints for all 5 folds
  • predictions_fold*.csv — per-fold survival predictions

Training Settings

• 5-fold cross-validation
• Epochs: 30
• Batch size: 2
• Learning rate: 3e-5
• Weight decay: 1e-5
• Pad size: 40
• MRI encoder frozen: Yes
• Loss: Cox proportional hazards
• Optimizer: Adam

Performance

• Mean C-index: 0.8152 ± 0.1189

Notes

This model achieved the highest overall performance across all tested configurations.
It serves as the primary model in the accompanying technical report, demonstrating the benefit of integrating MRI with clinical variables.

For complete training code and documentation, visit:
https://github.com/AliEngineering/multimodal-prostate-survival