Questions
What people ask first
Six answers in full, including the one that matters most — how to read the number you get back.
What do I actually get?
A verdict in seconds — whether the models flagged your scan and how strongly — the region on the image they responded to, and, if you want it, a written report that explains the result in plain language and answers your follow-up questions. Everything is based on the image alone, so bring the result to a doctor who also has your symptoms and history.
How accurate is it?
Measured against RSNA Pneumonia Detection Challenge, regions drawn by radiologists — 26 335 chest X-rays, 5 850 of them annotated positive — the ensemble reaches an AUC of 0.81, catches 90% of pneumonia cases and leaves 57% of clear scans unflagged. Members on the same run: PadChest 0.78, Stanford 0.77, MIMIC 0.70. Run against NIH ChestX-ray14, labels parsed from report text instead, the same ensemble on the same images scores 0.72. The gap is the answer key, not the models: NIH's pneumonia labels were extracted from report text by a parser, while RSNA's were drawn by radiologists for this exact task. We publish both.
The number is not a percentage — so what is it?
It is the single thing most worth understanding about the result. Each model is rescaled so that 0.5 sits exactly at the point where that model starts flagging a scan; the score you see is the average of those, compared against a line at 0.45. It measures how strongly the models reacted against their own flagging point — think of it as reaction strength rather than a probability of disease.
What is the highlighted region?
The output of a separate model, trained by us on regions drawn by radiologists for the RSNA challenge. Its strongest point lands inside such a region 86% of the time, against 11% for a point placed at random. The edges of the highlight are where its response falls off, not a measured border of anything. On 18% of scans that do carry a finding it marks nothing — and the result says so rather than showing an empty picture.
Can a clinic use this?
For evaluation, yes — put your own scans through it and compare against the reports you already have. Not in a diagnostic pathway: there is no CE mark, no FDA clearance and no clinical validation behind it, and we would rather say that in the first minute than let anyone find it out in the last. An API and a workflow integration are what we are building toward; neither exists yet. If you are weighing it up, write to us and say what you would need measured before it would be worth a radiologist's hour.
What does it cost?
The screening — the verdict, the score, the marked region — is free and needs no account. The written report and the follow-up chat are opened with an access key during the pilot: no money changes hands, but the number of reports is capped, because each one is a paid call to a language model and we are two students. Keys go out through the channels we post on and on request; write to us and we will send one. Nothing on this site can charge you, and no payment details are collected anywhere.
What image formats are supported?
PNG, JPEG and DICOM, up to 10 MB. DICOM files are converted to an image in your browser before anything is uploaded — the window level stored in the study is applied, so the scan looks the way it was read. Compressed DICOM (JPEG2000, JPEG-LS, RLE) is not supported; export those as PNG.
Is my data stored?
Your image and the maps drawn from it are deleted about ten minutes after your result is ready — that is also how long the result stays open, so save anything you want to keep before then. If you ask for a written report, the report and the questions you asked are kept for seven days so you can reopen them, and are then deleted automatically. We set no cookies and use no analytics.
Still wondering something?
We answer these ourselves — there is no support desk in between.