From Flagging Lesions to Drafting Reports: Aidoc’s Generative AI for Chest X-Rays Moves Toward FDA Review

Most medical imaging AI systems only highlight suspicious areas, leaving radiologists to synthesize the entire image, interpret related abnormalities, and write the report. Aidoc’s First Read extends into this core part of the work: it first analyzes a chest X-ray, then generates preliminary report text for a physician to revise and sign. This makes it more than another alert tool and an early test case for whether generative AI can withstand regulatory scrutiny as a diagnostic medical device.

The U.S. Food and Drug Administration (FDA) has granted First Read Breakthrough Device designation under number Q260882. However, the designation covers the detection and description of four life-threatening imaging findings, not all the capabilities the company plans. Aidoc told the media that the product is expected to cover more than 100 predefined findings, but it has not released the full list.

According to the company, the system first uses detection models to identify imaging features and then organizes the results into a structured report. This differs from existing AI that marks the location of a single lesion or reprioritizes a worklist: the vision-language model must integrate information from the entire X-ray, determine which findings should be included in the report, and generate text suitable for clinical use. Every draft must still be reviewed and signed by a radiologist; ultimate responsibility has not been transferred to the model.

First Read uses the same underlying architecture as an Aidoc abdominal computed tomography triage tool that has already received FDA clearance, but a shared architecture does not mean that the new use has been validated. Publicly available information has not provided the sample size of First Read’s validation study, the composition of participating medical institutions, disease prevalence, performance across different equipment and populations, or finding-by-finding sensitivity, specificity, or report error rates. Its more than 100 interpretation capabilities therefore cannot currently be assessed independently.

This also highlights the review challenges posed by end-to-end generative diagnostic AI. Regulators must not only confirm whether the four major findings can be identified reliably, but also contend with large numbers of negative and rare cases, images containing multiple abnormalities, and omissions, inaccurate descriptions, and inappropriate certainty that may arise during text generation. When the output shifts from a single alert to a complete report, the unit of validation expands from one lesion to multiple findings and their combinations.

Although human review is an important safeguard, it does not automatically eliminate risk. If draft quality is inconsistent, physicians may spend more time checking each sentence; if the text is mostly correct, automation bias may instead cause them to overlook a small number of critical errors. Real-world evaluation should therefore also include how physicians revise reports, whether errors are intercepted in time, and workload across different clinical settings—not merely the model’s accuracy on a static dataset.

Breakthrough Device designation can give developers more intensive FDA feedback and the opportunity for priority review, but it does not mean that a product has been authorized for marketing. First Read currently remains for research use, and Aidoc must still complete validation studies and obtain marketing authorization. The real regulatory test will be how to turn a designation covering only four life-threatening findings into a complete body of evidence sufficient to support broad chest X-ray reporting capabilities.