prostate_cancer_recurrence_prediction

This repository contains several trained model variants used for prostate cancer recurrence prediction. Please see GitHub link for further details:

Github

Each model was trained for 10 epochs per fold in a 5-fold cross-validation setup.

For every fold, the best-performing checkpoint based on validation C-index was saved.

All model files follow this naming pattern: <model_name>_foldX.pt, where X denotes the fold (0-4)

Available Models

1. Clinical Linear Model

Description: Regularized Cox proportional hazards model using clinical data

Prefix: clinical_linear_model

2. Clinical MLP

Description: Fully connected neural network using clinical data. Learning rate 1e-3

Prefix: clinical_mlp_model

3. MRI CNN

Description: 3D CNN trained on full MRI volumes (ADC, HBV, T2W + mask channel). Preprocessing includes mask as a seperate channel

Prefix: mri_cnn_model

4. Full Multimodal Model

Description: Multimodal model integrating Clinical MLP backbone to process clinical data and MRI CNN backbone to process MRI volumes. Feature embeddings from each branch passed through a risk prediction MLP

Prefix: multimodal_model_full

5. Multimodal Model — MRI Ablated

Description: Multimodal model with MRI CNN branch ablated

Prefix: mutlimodal_model_mri_ablated

6. Multimodal Model — Clinical Ablated

Description: Multimodal model with Clinical MLP branch ablated

Prefix: mutlimodal_model_clinical_ablated

7. Clinical MLP - Learning Rate 1e-4

Description: Fully connected neural network using clinical data (Same as #2) but with a Learning rate 1e-4

Prefix: clinical_mlp_model_lr1e_4

8. MRI CNN - Precropped

Description: 3D CNN trained on full MRI volumes (ADC, HBV, T2W). However, channels cropped using T2W-mask prior to concatinating into 1 tensor.

Prefix: mri_cnn_model_precropped

⭐ 9. Full Multimodal Model (Dual Learning Rate Optimizer)

Description: Same as #4 but uses a dual learning-rate optimizer (higher LR for the clinical MLP, lower LR for the MRI CNN) to balance learning between modalities.

The strongest-performing variant across all folds

prefix: multimodal_model_full_diff_lr

---

license: mit