A companion piece to Flight Path, episode one, with Mach7 CEO Teri Thomas.
Teri Thomas has spent a year listening to Mach7 customers, and one theme comes up more than almost any other. The appetite for AI in imaging is real, and it runs into a wall the moment a health system tries to use its own data. On episode one of Flight Path, she made the case that trustworthy imaging AI starts earlier than most people think. It starts with access.
The demand side is not in question. Radiology now accounts for more than three-quarters of the AI-enabled devices the FDA has cleared, the largest share of any specialty. Hospitals and imaging groups are surrounded by tools that promise to read faster, flag sooner, and take work off stretched teams.
When the answer is life or death, Teri asks, how do you know the output is something you can trust completely. Part of that answer is population fit. A model built on one patient population can lose accuracy when it meets a different one. External validation studies bear this out: algorithms with strong results on their original data have given up meaningful accuracy when tested on outside populations and other sites. The model is not broken. It is meeting patients it was never trained on.
That is why so many health systems now want to build and validate models on their own patients, or at least confirm that a vendor's model holds up on the people they actually serve. Teri sees this shift toward organizations creating their own foundation models as one of the most promising things happening in the field.
Some imaging vendors do not make it easy to reach your own data, Teri notes. Data that is locked to one system, in one format, behind one viewer, is data a health system cannot use for research, cannot use to train a model, and cannot check a vendor's claims against.
This is the problem Mach7 set out to solve. The company's stated goal is to make imaging data liquid: portable, usable, and owned by the customer. A vendor neutral archive keeps images accessible and standards-based rather than trapped, so a health system can pull its own data for clinical research, for building its own models, and, where it chooses, to license de-identified data to AI developers who need enough of it to clear a real trust threshold. That last option stays in the customer's hands, tied to consent and de-identification. The data belongs to the health system, not the vendor.
Teri is clear that AI in the reading room has to help without getting in the way. Surface twenty findings a radiologist judges to be noise and you have slowed the read and added exposure, not removed it. The goal is the opposite: AI that prioritizes the urgent study, sharpens the worklist, and brings the right prior images forward, with a clinician making the call. The human stays in the loop.
Where does this go? Teri's five-year read is that the software starts to bend to the user instead of the other way around. AI will learn the user more than the user learns the software, she says, describing an AI layer between the data and the clinician that tailors the experience to how each person works. That future rests on the same foundation as everything above: comprehensive data and an operating system trusted and flexible enough to build on.
Mach7's Enterprise Imaging family is built around that idea. It spans the Vendor Neutral Archive, the Communication Workflow Engine, the Enterprise Viewer (eUnity), and AI enablement, with Flamingo coming as the next offering. The through line is the one Teri keeps returning to: give health systems their data, keep clinicians in control, and AI gets to be useful and safe at the same time.
Hear the full conversation with Teri Thomas on Flight Path, episode one.
eUnity is a diagnostic viewer intended to display medical images and associated clinical reports to aid in diagnosis by trained healthcare professionals. Professional and prescription use only. In the United States only, eUnity supports digital pathology whole-slide imaging (WSI) for reference and referral use and is not intended for primary diagnosis.