A slide. A score. A spatial view.
Explore an AI-predicted HRD score, model attention and estimated tumor regions from an H&E whole-slide image.
Input & analysis
- Upload
- Encode & predict
- Spatial outputs
Submit de-identified research slides only.
Whole-slide viewer
Spatial outputs are displayed only when supplied by the analysis. No simulated attention or tumor regions are shown.
Methods & sampling coverage
Analysis results
The ≥42 label uses this same continuous prediction. The display scale does not constrain model output.
HRR14 panel
Separate classifierAt least one qualifying nonsynonymous mutation or deep deletion defines panel positivity. An image score does not determine individual gene alterations.
- Result source
- No result yet
- Model version
- R14 · Two-seed ensemble
- Analysis time
- —
- Sampled patches
- —
From tissue to a patient-level prediction
Your analysis workspace.
Tasks are scoped to this browser session. Open a completed task to inspect its score and spatial outputs.
| Case identifier | Input | Status | HRD AI score | HRR14 score | Created | Results |
|---|
Your first slide starts here
Open the TCGA example or submit a whole-slide image.
Go to slide analysis →Evidence you can inspect.
Results by model version, evaluation cohort and original label availability.
Current runtime · Full-cohort two-seed ensemble
Fixed validation patients: 1299 TCGA and 156 CPTAC. Each endpoint uses only its known labels; unknown labels are not treated as negative.
| Endpoint | TCGA | CPTAC | Known labels (TCGA / CPTAC) |
|---|---|---|---|
| HRD Spearman ρSame continuous HRD output | 0.710 | 0.692 | 1268 / 146 |
| HRD≥42 AUCAUC of the same continuous output | 0.821 | 0.892 | 1268 / 146 |
| HRR14 AUCSeparate classifier · Full-cohort multiscale model | 0.673 | 0.698 | 1289 / 112 |
Three-route fusion · Fixed-subset experiment
A fixed fusion of multiscale, denser sampling and encoder-adaptation routes yielded these HRR14 AUCs on a 256-patient validation subset:
This separate subset experiment is not the runtime model used for uploaded tasks.
What these numbers tell us
Spearman ρ measures rank agreement between predicted and measured HRD scores. AUC measures discrimination between positive and negative cases.
These estimates do not establish clinical utility or treatment benefit. Interpretation requires formal pathology and molecular testing.
A transparent model and workflow.
Input formats, output definitions and spatial interpretation.
Analyze a slide
- Inspect the TCGA example
Open a real TCGA PRAD slide with precomputed model outputs and recorded provenance. Values are actual computations when available.
- Upload a whole-slide image
Choose a supported SVS, TIFF or NDPI file. Large WSIs use sequential chunk uploads. The service performs tissue sampling, registered image encoding and model prediction.
- Inspect spatial outputs
Use the HRD attention and tumor-estimate switches independently. Attention is the model's relative weighting of sampled tissue, not a tumor mask or local molecular measurement.
- Review and export
Download the result report and available spatial images, or print the current view. Task status remains available in analysis history.
Advanced input: registered feature NPZ
uni2 [S, 128, 1536] conch [S, 2, 16, 512] scale_valid [S, 2] · bool
S is the number of slides owned by one patient, at most 32. The two CONCH branches use 0.5 and 1.0 μm/px. Encoder and preprocessing compatibility are required. Feature-only inputs do not produce spatial outputs without coordinates and a matching image. Public NPZ uploads are limited to 64 MiB.
HRR14 panel definition
The panel is positive if at least one of the 14 genes has a qualifying nonsynonymous mutation or an explicit deep deletion (−2). Two or more altered genes also define a positive panel.
A negative label requires complete assay evidence when no alteration is observed; otherwise it remains unknown. The site does not infer individual gene alterations from the slide.
Continuous HRD and ≥42
The HRD AI score is the continuous output of a separate regression model. The ≥42 label is derived directly from this output, not from the HRR14 score.
Read the layers independently
High attention does not establish the presence of tumor, a mutation or local HRD. A tumor estimate is not a pathologist annotation. Its method and spatial coverage accompany the result; unavailable layers remain disabled.
Version and intended use
Model output is not a clinically calibrated diagnostic probability and should not independently determine treatment. Actual model and visualization availability is shown for each task.
Interaction references: RCC Lens Lab and KIPAN demo. HRD_Path is an independent implementation.