V-LINC: Individualized Breast Cancer Prognosis

An interactive platform is available at: https://huggingface.co/spaces/zhang0319/Multimodal_Surv

This repository contains the preprocessing, training, validation, and evaluation code for V-LINC, a multimodal survival prognostication system for individualized breast cancer prognosis. The system integrates breast DCE-MRI, radiology reports, structured clinical variables, treatment information, and optional mutation features.

V-LINC system

Note: Study setting and V-LINC system. a, Clinical context of pretreatment breast MRI acquisition and subsequent treatment pathways, including neoadjuvant therapy and surgery-first management. b, Geographic distribution of the in-house cohort and external validation cohorts. c, Data sources used for prognostic evaluation, including DCE-MRI and radiology reports from radiological evaluation, clinicopathological variables from pathological evaluation and longitudinal follow-up for survival endpoints. d, Overview of the V-LINC system.

Repository structure

.
├── configs/
│   └── default.yaml
├── data/
│   ├── internal_survival.example.json
│   ├── duke_survival.json
│   └── ispy1_survival.json
├── preprocessing/
│   └── preprocess_mri.py
├── RadioLOGIC/
├── src/
│   ├── config.py
│   ├── dataset.py
│   ├── features.py
│   ├── losses.py
│   ├── metrics.py
│   ├── models.py
│   └── utils.py
├── validate_setup.py
├── train.py
├── test.py
└── requirements.txt

Installation

pip install -r requirements.txt

RadioLOGIC

Place the pretrained report encoder and tokenizer files in:

RadioLOGIC/

Contains:

RadioLOGIC/
├── config.json
├── vocab.json
├── merges.txt
└── pytorch_model.bin

MRI segmentation and preprocessing

Tumor masks were obtained using our breast MRI segmentation tool:

https://huggingface.co/spaces/zhang0319/breast-mri-seg

The preprocessing script expects one directory per case:

CASE_ID/
├── dce1.nii.gz
├── dce2.nii.gz
└── seg.nii.gz

Run preprocessing with:

python preprocessing/preprocess_mri.py \
  --input-root /path/to/segmented_cases \
  --output-root /path/to/preprocessed_cases

The script selects the tumor-containing breast, applies the configured crop, and center-crops or pads the volumes to 160 × 160 × 160.

Metadata format

Example metadata files and the required data structure are provided in the data/ directory.

Configuration

Edit:

configs/default.yaml

Set the following before running the code:

• metadata paths
• MRI directories
• RadioLOGIC directory
• output and checkpoint directories
• endpoint (os or dfs)
• batch size, learning rate, and number of epochs

Validate the setup

Check metadata and configuration:

python validate_setup.py --config configs/default.yaml

Also verify RadioLOGIC:

python validate_setup.py \
  --config configs/default.yaml \
  --check-radiologic

Verify all MRI files:

python validate_setup.py \
  --config configs/default.yaml \
  --check-radiologic \
  --check-images

Training

python train.py --config configs/default.yaml

The best checkpoint is selected using validation C-index and saved in the configured checkpoint directory.

Evaluation

Internal evaluation:

python test.py --config configs/default.yaml --cohort internal

External evaluation:

python test.py --config configs/default.yaml --cohort duke
python test.py --config configs/default.yaml --cohort ispy1

Results

• Overall survival Note: Overall survival stratification across internal and external cohorts. a, Distribution of survival status across the in-house and external cohorts used for overall survival analysis. b–d, Kaplan–Meier curves for V-LINC-defined low-risk and high-risk groups in the NKI training, validation and test sets. e–h, Kaplan–Meier curves for the ISPY1, DUKE, Sun Yat-Sen and RUMC external test sets.

• Interpretation Note: System interpretability and time-dependent performance for overall survival. a, Representative DCE-MRI examples with lesion masks and Grad-CAM maps before and after contrast enhancement. b, Token-level attention maps from radiology reports, with higher-intensity tokens indicating greater contribution to the text-derived representation. c, Gradient-based attribution of structured clinical variables across patients. d, Decision curve analysis for 5-year overall survival. e, Volcano plot showing feature differences between V-LINC system-defined high-risk and low-risk groups across imaging, report, clinical and treatment-prompt features. f, Time-dependent receiver operating characteristic curves for overall survival from 1 to 10 years.

Citation

Citation information will be added after publication.

Contact details

If you have any questions please contact us.

Email: tianyu.zhang@radboudumc.nl / t.zhang@nki.nl (Dr. Tianyu Zhang); jakob_nikolas.kather@tu-dresden.de (Prof. Jakob Nikolas Kather); ritse.mann@radboudumc.nl (Prof. Ritse M. Mann)