Skip to content
progresfulPublic

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

OncoScan

A working prototype of photo-first cancer triage. A patient sends photos of a worrying skin or mouth spot, a specialist reviews them, and the patient gets a result with a next step. Built at Junction 2026 for AstraZeneca's "Equal Care Across Finland" challenge.

Live demo: https://oncoscan.pgf.lv · Video (45 s): https://youtu.be/jBnBLPOL8io

Specialist screen: a patient's case with a real dermoscopy image and the signal check

Every screenshot on this page is a capture of the running product. The lesion images are real dermoscopy photos from the public HAM10000 dataset; nothing is mocked up or drawn.

The journey

1 · The patient adds a photo 2 · The case clock starts
Patient flow: photo step with a real image, checked for sharpness Patient flow: case page with a running 24-hour clock
Each photo is checked for darkness and blur before it is sent. The case goes to the front of the specialist's queue.
3 · The specialist reviews and signs 4 · The patient gets a next step
Specialist screen: verdict note sheet Patient flow: result page
One key for the verdict, one for the note. Three possible results. It never says "you have cancer".

The signal check

Specialist screen: demo case with a confirmed melanoma and the signal check open

For each case the specialist sees a plain verdict and two lists: signals that may indicate cancer, and signs against it. Items marked "(patient's report)" come from the patient's answers. After signing a demo case, the dataset's confirmed diagnosis is shown.

The landing page

Landing page Landing page: step-by-step section

Run it

node server.mjs

Then open:

A case sent from the patient page is saved on the server (photos and details under data/cases/, which is not in git). It appears first in the specialist queue on any device that can reach the server, and the signed verdict returns to the patient. Needs Node 18 or newer and no packages.

AI photo description (optional)

sh tools/set-key.sh

This asks for a Google Gemini API key and saves it to .env. Restart the server afterwards. The specialist's signal check then asks Gemini (gemini-3.5-flash-lite by default) to describe each photo. The photo, area and body site are sent to Google; the patient's description is not.

What we measured

Test Result
Full journey, automated in headless Chrome (tools/journey.mjs) 5 of 5 runs passed, twice
Photo check on a sharp, a dark and a blurred test image 3 of 3 labelled correctly
Gemini's score on 42 labelled HAM10000 images, photo only (tools/eval.mjs) Not usable for triage (below)
Gemini model Cancers flagged (of 18) Benign cleared (of 24)
gemini-3.5-flash-lite 13 5
gemini-3.8-flash 11 11

A general vision model could not tell cancer from benign lesions on the photo alone. So Gemini only describes the photo here. Its score is hidden, and the verdict comes from the patient's report and a sharpness check. Per-image results are in review/cases/ham10000/gemini-eval-*.json.

What is real and what is simulated

Real Simulated or not built
5-step patient submission (1 or 2 photos), saved on the server Sign-in: any email works, nothing is sent
Photo downscaling and a darkness and blur check Email delivery: shown as a preview
Case clock and status County cover list and the €29 price: examples
Specialist queue, swipe or keyboard verdicts, notes, undo Payment: no charge is made
Verdict returning to the patient as one of three results A model that can triage from the photo
Photo description by Gemini, when a key is set Licence check, calibration, payouts, second review
5 skin demo cases with real images and confirmed labels The 3 mouth demo cases: fictional, no image
The 24-hour turnaround: a clock, not a service
Landing-page "Case flow" panel: a labelled simulation

Limitations

  • No logins yet: anyone who can open the specialist page sees the queue and the photos. A patient's result is protected by a random token kept in their browser.
  • Cases are plain files on the server's disk, with no encryption or backup.
  • For the 5 skin demo cases the image, body area, age band and diagnosis come from the dataset. The patient's words and the hand-written example analysis are ours.
  • Without a Gemini key, the demo cases show that hand-written example analysis.
  • The sharpness check is crude: it also flags some smooth, low-contrast lesions.
  • Not tested with clinicians or patients, on phone-camera photos of lesions, or on a physical phone.
  • Not a medical device. Nothing here gives a diagnosis.

Images and data

The skin images are a 42-image sample of HAM10000 (Tschandl, Rosendahl and Kittler, Scientific Data 5, 180161, 2018), downloaded from Kaggle. See review/cases/ham10000/SOURCE.md for the source and licence notes: as far as we know the dataset is CC BY-NC 4.0, non-commercial use with attribution. Landing-page photos were supplied by the team.

Built during the hackathon vs pre-existing

  • Built during Junction 2026: everything in this repository.
  • Pre-existing: the concept draws on the team's earlier product, Oriscan. No Oriscan code, data or models are here.
  • External: Hanken Grotesk and Newsreader (Google Fonts), Solar icons via Iconify, GSAP and ScrollTrigger from a CDN, and the Google Gemini API when a key is set.

Repository layout

Path What it is
landing/, patient/, review/ The three screens
server.mjs Static server plus the Gemini call
tools/ journey.mjs (end-to-end test), eval.mjs (model test), set-key.sh
presentation/ Submission text, deck, video and the evidence ledger
docs/screenshots/ The images on this page

Made by Progresful (Emils Plavenieks, Jurgis Kacens, Krisjanis Mezulis) 2026

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages