Six steps, and what each one is allowed to say

1. 01

   Validation

   Is this a chest X-ray?

   The upload is checked before anything else runs, so every result the site produces is a result about an actual chest radiograph. Anything else stops here with a plain message.

   Front and side views both pass this check, as do adult and child chests. The models are built for adult frontal scans — which is exactly what the upload guide asks for before you choose a file.

2. 02

   Lung segmentation

   What counts as lung?

   A PSPNet outlines the lung fields at 0.97 Dice, from a method published in IEEE Transactions on Medical Imaging. The outline is drawn as an optional overlay on your result and never alters the image beneath it.

   The outline is trained to find air-filled lung, and consolidated lung looks like soft tissue to it. So the outline saying “this is lung” can be trusted; it saying “this is not” cannot. About one scan in thirty-five is rejected here because no lungs could be found at all.

3. 03

   The ensemble

   Do the models flag it?

   Three DenseNet121 classifiers — trained independently on Stanford CheXpert, MIMIC-CXR and Spain's PadChest — each read the scan. Each flags a scan at its own point, one of them as low as 0.044, so every score is first rescaled to put that point at 0.5. The rescaled scores are averaged and compared against a fixed line at 0.45.

   This is why the number you see is not a probability of disease. It is how strongly the models reacted, measured against the point at which they start flagging.

4. 04

   Finding location

   Where is it?

   A U-Net we trained ourselves on 26 684 images, 9 555 radiologist-drawn boxes from the RSNA challenge marks where the finding is. It casts no vote on whether — step 03 answers that. Its strongest point lands inside a radiologist-drawn region 86% of the time, against 10.6% for a point placed at random: 8.1 times chance.

   On 18% of scans that do carry a finding it marks nothing. That is an outcome, not a failure — the result says so plainly instead of showing an empty picture, and it does not soften the verdict.

5. 05

   Your result

   What you get for free

   The verdict, the score against the flagging line, how the three models split, and the localizer's map over your scan. No account, no payment, nothing held back about what the models found.

   The scan and the maps drawn from it are deleted about ten minutes later, and the result page dies with them. Save what you want to keep before then.

6. 06

   The written report

   What it means for you

   Claude Sonnet 5 reads every measurement on the page — the ensemble score, each model's own reading, how far above its own flagging point it landed, and where the localizer responded — and writes it up as a document addressed to you. It takes about half a minute. You can then ask it up to six questions about your own result, and download the whole exchange as a PDF to hand to a doctor.

   It works from the measurements rather than the picture — it never sees the image — so everything it writes can be traced back to a number on your result page.

Where each number came from, and on which images, is set out on the research page.