Find the patients whose tumours were never tested.
Many people with colorectal or stomach cancer never get the MSI test that decides whether immunotherapy could work. Their slide is already in the lab. Histofinder reads it and flags who should be tested next.
Live: colorectal and stomach cancer. Coming next: endometrial, pancreatic and prostate cancer.
of people with metastatic colorectal cancer in a 41,061-patient US study had no MSI test on record.
ADLM, 2025 ↗responded to immunotherapy in an early study when tumours were MSI-high vs when they were not.
CURE ↗What the model sees on whole slides
Real colorectal and stomach cancer slides from patients the model never saw in training. Switch to the AI view to see which regions the model weighted most when scoring the slide. A pathologist can open the slide and check those regions.
One missed test affects three groups of people
A treatment that might work much better
MSI-high tumours are covered in DNA errors, so the immune system can recognise them. Immunotherapy helps it do that. Without the test, it is never considered.
A chance to prevent cancer, not just treat it
Some MSI-high cancers come from Lynch syndrome. Relatives who learn they carry it can start colonoscopy early and have polyps removed before they turn into cancer.
Trials that can find their patients
Trials of new drugs for MSI-high cancer need these patients. Untested patients are invisible to them.
The difference is whether anyone looks again
Today The test slips through
- Diagnosis and surgeryAn H&E slide is made, as for every patient.
- MSI test never orderedThe order is missed, the biopsy runs out, or the clinic is stretched.
- Standard chemotherapy onlyImmunotherapy is never considered.
- Family never toldRelatives keep a hidden, preventable risk.
With Histofinder The slide speaks up
- Diagnosis and surgeryThe same slide is scanned.
- Flagged for testingHistofinder adds the patient to the lab's worklist.
- Confirmed by IHC or PCRA standard test makes the call, not the AI.
- Immunotherapy consideredThe oncologist can weigh a drug matched to the tumour.
- Family protectedIf it is Lynch syndrome, relatives are offered testing and early screening.
Five steps, no new sample
Scan
The lab digitises the H&E slide it already made.
Tile
Tissue is cut into thousands of small tiles at 0.5 µm per pixel.
Read
A pathology foundation model describes each tile.
Weigh
Attention learning finds the regions that signal MSI.
Flag
Untested patients are ranked; the top ones go to testing.
What would it find at your hospital?
Set your numbers. Defaults come from published figures; the model's hit rate comes from the Demo page.
Run Histofinder on your own slides
We are looking for pathology labs and cancer centres for a first research pilot on de-identified slides.
Ask if your tumour was tested. If it was MSI-high, tell your family.
An MSI or MMR test on a colorectal or stomach tumour can change your treatment. Sometimes it also reveals Lynch syndrome, an inherited condition your relatives can act on before cancer starts.
people have Lynch syndrome. Up to 90% don't know they have it.
chance that each brother, sister or child of a carrier has the same gene change.
colonoscopy from age 25 is recommended for relatives who test positive.
One result can protect a whole family
Lynch syndrome is passed down in families. When one person is found to carry it, close relatives can take a simple blood or saliva test for the same gene change. This is called cascade testing.
Relatives who test negative go back to average risk. Relatives who test positive can start colonoscopy early, when polyps can be removed before they become cancer.
On average a person has about four first-degree relatives (parents, siblings, children), and about half are likely to share the family's gene change (Washington State DOH).
From one result to a protected family
- Get the patient's genetic report
It names the gene and the exact change, and confirms it causes disease.
- See a genetic counsellor
They review family history and explain what each result would mean.
- Relatives give a blood or saliva sample
The lab looks only for the family's known change, which is quicker and cheaper.
- Act on the result
Negative: normal screening. Positive: colonoscopy every 1–2 years from age 25, plus screening for other Lynch cancers.
- Pass it on
Each relative who tests positive tells their own children and siblings.
Bring these to your next appointment
- Was my tumour tested for MSI or mismatch repair (MMR)?
- If not, can it be tested on the tissue you already have?
- If it is MSI-high or MMR-deficient, could immunotherapy be an option for me?
- Should I have genetic counselling or testing for Lynch syndrome?
- Which of my relatives should be tested, and how do I tell them?
General information, not medical advice. Talk to your doctor or a genetic counsellor about your own situation. Sources: Moffitt Cancer Center, Washington State DOH, CION Cancer Clinics, CURE.
A background safety net for MMR/MSI testing
Histofinder runs on slides your lab already scans and produces a ranked worklist of untested patients. It never replaces IHC or PCR. It makes sure the right patients reach them.
An MMR-deficient tumour without BRAF V600E or MLH1 promoter methylation suggests a hereditary cause and warrants genetic counselling (CURE). Follow your institution's protocol.
It can only add tests, never remove them
| Rule-out pre-screen (e.g. MSIntuit) | Histofinder | |
|---|---|---|
| Runs on | Patients already headed for testing | Patients never tested |
| Goal | Skip tests for likely MSS cases | Send likely MSI-high cases to testing |
| If the AI misses | A case could be cleared by mistake | No worse than today |
MSIntuit: Saillard et al., Nature Communications 2023.
Fits the lab you already run
- Slides from your scanner
Standard whole-slide formats (.svs, .tiff, .ndpi).
- A ranked worklist
Untested patients ordered by MSI priority, with a cut-off you choose.
- Evidence for review
The original slide, the AI attention map and the highest-attention regions.
- A test order, not a diagnosis
A short report recommending MMR IHC or MSI PCR.
Research use
Retrospective screening of slide archives with partner labs. Labelled research use only.
Lab-validated test
A CLIA-certified partner lab validates and runs it as part of its own test.
FDA clearance
Multi-site validation supports clearance for use in any hospital.
About 60% of academic centres and 50% of reference labs already use whole-slide imaging clinically (L.E.K., 2026).
FAQ
Does Histofinder replace MMR immunohistochemistry or MSI PCR?
No. It decides who should be tested among patients who were not. Every flag is confirmed with a standard test before any treatment or genetic decision.
Which specimens does it work on?
The research models are trained on diagnostic (FFPE) H&E slides of colorectal and stomach cancers from TCGA. Biopsies and other scanners need separate validation, which is part of the pilot plan.
How accurate is it?
See the performance section of the Demo page. Those are held-out test-set numbers from one public dataset (TCGA colorectal and stomach); validation on your own slides is the first step of any pilot.
Does patient data leave the hospital?
For a pilot, Histofinder can run inside your environment on de-identified slides under a data use agreement, set up with your IT and privacy teams.
Is it FDA cleared?
No. It is a research prototype for research use only.
The screening worklist
Patients whose slides were read but who were never tested for MSI, ranked by priority. Choose where to draw the line, then open any patient.
| # | Patient | Priority | Action | Truth |
|---|
How well it finds MSI-high patients
Held-out test set: 20% of patients, split off before training and used once at the end. Every score comes from a model that never saw that patient.
Screening curve
Testing in Histofinder's order vs at randomROC curve
Share of flagged patients who are MSI-high
Operating points
What happens if you send different shares of untested patients to testing.
These are held-out results inside one public dataset (TCGA); small test sets give wide confidence intervals. External validation of a related rule-out tool on 600 new cases reached an AUROC of 0.88 (Saillard et al., 2023); testing on new hospitals and scanners is Histofinder's next step.
Upload a whole-slide image
Whole-slide images are 0.5–2 GB and need a GPU, so they are analysed in the live analyzer rather than in this page.
Load the notebook's output folder
Pick the histofind_msi folder. This page reads demo_data/ and results/ in your browser only; nothing is uploaded.
Why Histofinder exists
The information that decides a cancer patient's treatment, and sometimes their family's future, is often already sitting on a glass slide. Histofinder exists so that no one misses it because a test was never ordered.
Tawsifur Rahman
Founder. PhD candidate in Biomedical Engineering at Johns Hopkins University. His research is on teaching AI to read whole-slide pathology images when labels are weak or scarce, which is the core problem behind predicting a tumour's molecular status from a routine slide. His work on computational pathology has appeared in Nature Digital Medicine and Modern Pathology.
Pathologists, labs and cancer centres
If your patients miss MMR/MSI testing, we would like to run Histofinder on a de-identified set of your slides and show you what it finds.
Research prototype. Not for diagnostic use. Data: The Cancer Genome Atlas (TCGA Research Network). All slide images are de-identified research slides from TCGA, used for research only.