3,675 patients
NKI3,449
Radboudumc226
NKI: 1,379 training · 1,034 validation · 1,036 internal test
V-LINCINDIVIDUALIZED BREAST CANCER PROGNOSIS
BREAST IMAGING GROUP · MULTIMODAL AI
Breast cancer prognosis.
Breast cancer prognosis.
A more complete picture.
V-LINC integrates MRI, clinical information and radiology reports for breast cancer prognosis. Studied in 5,234 patients across Europe (the Netherlands), the United States and China.
DCE-MRIClinical profileRadiology report
5,234Patients in the study
1,785External-cohort patients
6Cohorts
3Countries · 3 continents
01 / INTERNATIONAL STUDY
Across Europe, the United States and China.
Developed and evaluated in geographically diverse breast cancer cohorts, with NKI as the development and internal-testing cohort and five external cohorts.
1,090 patients
Duke922
I-SPY1168
Two independent external cohorts
469 patients
Sun Yat-Sen Memorial Hospital423
Binzhou Medical University Hospital46
Two independent external cohorts
View the full cohort breakdown
| Cohort / split | Country | Patients |
|---|---|---|
| NKI training | Netherlands | 1,379 |
| NKI validation | Netherlands | 1,034 |
| NKI internal test | Netherlands | 1,036 |
| Duke | United States | 922 |
| I-SPY1 | United States | 168 |
| Sun Yat-Sen Memorial Hospital | China | 423 |
| Binzhou Medical University Hospital | China | 46 |
| Radboudumc (RUMC) | Netherlands | 226 |
| Total | 3 countries | 5,234 |
02 / OVERALL SURVIVAL
Performance across cohorts.
Key internal-test and external-validation results for overall survival (OS).
NKI INTERNAL TEST0.737C-index · 95% CI 0.692–0.780
DUKE EXTERNAL COHORT0.744C-index · 95% CI 0.681–0.803
Model-defined risk groups were evaluated alongside discrimination, with hazard ratios reported for high-risk versus low-risk patients.
| Cohort | Cohort size | Hazard ratio (95% CI) |
|---|---|---|
| NKI internal testInternal | 1,036 | 3.98(2.48–6.39) |
| DukeExternal · United States | 922 | 4.13(2.36–7.22) |
| I-SPY1External · United States | 168 | 3.10(1.49–6.42) |
| Sun Yat-Sen Memorial HospitalExternal · China | 423 | 2.97(1.18–7.45) |
| Radboudumc (RUMC)External · Netherlands | 226 | 4.58(1.38–15.26) |
HR: high-risk versus low-risk group. CI: confidence interval. The study total includes all cohorts; only reported OS test results are displayed here.
03 / MODEL EXTENSIONS
More information.
More information.
A broader prognostic view.
Additional study analyses include genomic integration, distant metastasis–free survival and disease-free survival.
Disease-free survival
Disease-free survival analysis across cohorts, with hazard ratios comparing model-defined high- and low-risk groups.
HR range across cohorts2.08–4.11Minimum–maximum HR
Adding genomic information
Genomic information extends the multimodal model for overall-survival analysis in the NKI test cohort.
C-index0.75995% CI 0.725–0.793
Hazard ratio4.7595% CI 2.83–7.96
Beyond overall survival
Distant metastasis–free survival analysis in the NKI internal test cohort showed separation between model-defined high- and low-risk groups.
Hazard ratio5.2995% CI 2.86–9.77
EndpointDistant metastasis–free survival
These are study-level extension results. The current online demo uses the original MRI, clinical and report-based OS model; genomic inputs and separate distant-metastasis or DFS prediction modes are not enabled in this demo.
04 / EXPLORE THE MODEL
From images to an integrated view.
Prepare the images
Load the MRI pair and tumour mask. Inspect the slices and the interactive 3D tumour surface.
Complete the profile
Bring together clinical staging, tumour receptors, treatment information and the radiology report.
Explore the result
Review the original model score percentile, risk group and survival display.
Build the patient profile.
Upload the MRI volumes, review the clinical information and run the model.
The existing survival display uses a percentile-based mapping; it is not a validated, calibrated individual survival probability.
Understand the output.
A concise guide to the information shown by the existing V-LINC workflow.
Score percentile
Places the patient's model score within the original reference score distribution. A percentile is not an event probability.
Risk group
Uses the existing percentile cut-offs to assign a low-, medium- or high-risk group. The model's original calculation is unchanged.
Survival display
The original 3-, 5- and 10-year values and curve are derived from a predefined percentile mapping, not a fitted and validated baseline survival model.
MODEL CONTINUITYA new interface.
A new interface.
The same scoring.
Model weights, image normalisation, tumour-mask weighting, text processing, clinical encodings and score calculations remain those of the source Space. No recurrence-model results have been substituted.
Research contacts
Dr. Tianyu Zhang · Tianyu.Zhang@radboudumc.nl
Prof. Jakob Nikolas Kather · jakob.kather@med.uni-heidelberg.de
Prof. Ritse Mann · Ritse.Mann@radboudumc.nl