MIRROR forces an AI radiology report writer to work blind, so it can never describe a finding its own classifier did not flag.
The system chains three pieces: a multi-label classifier that scores findings, a Grad-CAM localizer that turns each positive score into a named anatomical region, and a language model that drafts the report from labels, probabilities, and regions alone, never the image itself. A shared registry holds the taxonomy, anatomy, and phrasing for chest X-ray, brain MRI, and head CT, so adding a new modality is mostly a data update. All three modalities are routed and tested, but only the chest X-ray classifier is trained so far, reaching a macro AUROC of 0.729 on the ChestMNIST benchmark and beating chance on all 14 labels.
That auditability is the real contribution: because the writer cannot see pixels, every sentence in a MIRROR report can be checked against the probability vector that produced it. But the underlying classifier is weaker than the headline number suggests. At the default 0.5 threshold, it predicts zero positive cases for 11 of the 14 labels, and its Brier score of 0.045 barely beats the 0.047 a model that ignores the image entirely would get by just guessing normal.
That is the quieter finding here: strong looking aggregate metrics can flatter a detector that has effectively learned to say nothing, and the paper's own example, a cardiothoracic ratio the system never actually measured, shows how confidently a well architected report can still be wrong on the numbers even when it cannot lie about the labels.