prostate-multimodal-survival-mri

MRI-Only Prostate Cancer Survival Model

This repository contains outputs from the MRI-only survival prediction model.
Only imaging data (T2, ADC, HBV) were used—no clinical variables were included.

A 3D ResNet-18 encoder was used to extract voxel-level imaging representations, followed by a Cox survival prediction head.

Contents

• mri_only_outputs.zip
  • models/ — best checkpoints for 5 folds
  • predictions_fold*.csv — per-fold risk 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

Performance

• Mean C-index: 0.6470 ± 0.0759

Notes

This model demonstrates the standalone predictive utility of MRI.
Performance improves substantially when clinical variables are added, supporting the value of multimodal fusion.

For full training code, please refer to:
https://github.com/AliEngineering/multimodal-prostate-survival