Built on radiology workflow reality.
Our approach to validation starts with radiologists, not benchmarks.
How we measured what we built.
We did not benchmark ImageAssist against datasets designed to be easy. We built our validation methodology around the real-world question radiologists and radiology operations directors actually care about: when the worklist is full and the urgent case is buried, does ImageAssist surface it in time for the read to matter?
The answer is not just a sensitivity number. It is a workflow number. We measured time-to-first-read on flagged cases across our pilot deployments, and we tracked radiologist agreement with each triage decision at those same sites. Attending radiologists reviewed the AI scoring on every flagged case and confirmed or overrode the prioritization. Cases where disagreement was consistent pointed to model performance gaps, which we addressed before expanding the pilot.
We report what we can stand behind. For findings where radiologist agreement is strong and time-to-read impact is measurable, we state a number. For areas where the evidence is thinner, we say so directly. We would rather tell you what we do not know than hand you a number that breaks under scrutiny at your institution.
What we measured on real worklists.
From our pilot deployments. Not curated benchmark datasets.
These figures come from internal measurement across our pilot deployments, not from a published clinical trial or peer-reviewed study. We are transparent about that distinction and happy to discuss the methodology with your clinical or IT team.
The radiologist decides. The AI prepares the queue.
ImageAssist is designed as decision support, not autonomous diagnosis. The radiologist's read is always the final step. The AI's job is to make sure the right study is in position one when the radiologist opens their worklist.
Every flagged case shows the radiologist the AI urgency score alongside the study. The radiologist can review, confirm, or override. There is no alert fatigue design: flagged cases are moved up the queue, not pushed as pop-up interruptions. The radiologist's reading flow is preserved.
Override data is captured and fed back into model performance tracking. When radiologists systematically disagree with a flag category, we treat that as a model performance signal, not a user error.
Questions about our approach?
Our clinical team will walk you through the validation methodology in full, and answer your questions directly.