The FDA-cleared landscape of emergency imaging AI, and the gap Atlas targets

A crowded but narrow field

The US FDA has authorized many hundreds of AI and machine-learning-enabled medical devices, and radiology is by far the largest category [1]. The American College of Radiology's Data Science Institute maintains a public directory of these cleared products [2], and one pattern stands out: most are narrow. A typical cleared device does one thing, flags one finding, and notifies a human.

Where the clearances cluster

The densest clusters are in neuro and chest, where minutes change outcomes. Triage-and-notification tools for intracranial hemorrhage and large-vessel-occlusion stroke are offered by companies such as Viz.ai [3], RapidAI [4], and Aidoc [5], and pulmonary embolism and cervical-spine tools are common too. These products earned their place by attaching to a time-critical pathway and showing they shorten it.

What triage and notification means, and does not

Many of these clearances are for computer-aided triage: the device prioritizes a worklist or pushes an alert, but it does not produce a standalone diagnosis or measurement. That is a sensible regulatory lane. It also means the output is usually a single yes/no flag for one condition, not a localized, multi-finding read of the whole study.

Where Atlas sits, honestly

Atlas is research-stage and not FDA cleared. But its shape differs from the cleared norm in three ways: it covers six abdominal emergencies at once rather than one, it localizes each finding to a bounding box and a named abdominal region rather than emitting a single flag, and it drafts report-ready text. Broad multi-disease classifiers exist in the literature [6], but classification without localization is a different product. The claim is not that Atlas beats cleared tools; it is that the abdominal, multi-pathology, localization-first niche is comparatively empty.