The pipeline

Four models, each with one job

Every scan runs through a chain, and each link answers exactly one question. Three classifiers say whether. A localizer trained on regions drawn by radiologists says where. A segmentation model finds the lung fields. A validator confirms the image is a chest X-ray to begin with. Keeping them separate is what lets each one be judged on its own evidence.

Whether

Pneumonia ensemble

3 models, 3 datasets, 3 institutions

Three DenseNet121 models read the scan and vote — one trained at Stanford, one at MIT and Harvard, one at a university hospital in Alicante. Different continents, different patients, different teams, which is the whole point: models that fail in the same places would agree with each other and tell you nothing. Their scores are rescaled onto one axis before they are averaged, because each flags a scan at a different point.

Details Where

Finding localizer

Peak lands in a radiologist's box 86% of the time — 8.1× chance

The one model here we trained ourselves, on 9 555 regions drawn by radiologists for the RSNA challenge. It answers only where, which is what lets it be judged on its own evidence: its strongest point lands inside a marked region far more often than the classifiers' own maps ever managed. When it does not single out a region, the result tells you plainly rather than showing an empty picture.

Details What is lung

Lung segmentation

Traces both lungs at 0.97 Dice — published in IEEE TMI

A PSPNet out of IEEE Transactions on Medical Imaging. It outlines the lung fields so a finding can be described in anatomical terms rather than in pixels, and so we can tell when a highlighted region falls outside them. Worth knowing about its outline: it is trained to find air-filled lung, and consolidated lung looks like soft tissue to it — so it saying "this is lung" can be trusted, and it saying "this is not" cannot.

Details Is this a chest X-ray

X-ray validator

Checked on 25 images — a spot check, not a test set

The gate at the front. It confirms the upload is a chest radiograph before anything else runs, so every result the site produces is a result about an actual chest X-ray. Two things it does not do: it has no large-scale evaluation behind it, only a 25-image spot check; and it does not separate a front view from a side one, or an adult chest from a child's.

Details