Making sure follow-up scans really happen

Imaging scans often turn up something nobody was looking for, called an incidental finding, like a small spot on the lung seen on a scan done for another reason. About 6–12% of radiology reports recommend further action like a follow-up scan, which may be due in weeks, months or even years. Nobody owns making sure it happens, and 61% of them are missed in practice. In one study, 44% of the patients whose follow-up was never acknowledged by their doctor were at risk of serious harm, such as a suspected cancer.

Our first product went live 6 months after we started. Then came 2 pilots, and then the first 4 contracts, all from cold outreach, securing $212K in annual recurring revenue at 1 hospital and 3 imaging centers.

From scan to follow-up. The hospital’s own systems send each report as an HL7 v2 message through an interface engine. Inflo’s NLP finds each follow-up the report calls for, and the tracker follows it until the next scan is done.

Integrating with hospitals and imaging centers

Before software can read a report, it has to receive it. The scanner sends its images to the PACS, the image archive, as DICOM files. The radiologist reads them in the PACS viewer and dictates and signs the report in a reporting system, which passes it to the RIS, the radiology information system that holds orders, the schedule and reports. The images themselves stay in the PACS, though many vendors sell the two together as one RIS/PACS. In hospitals on Epic, the RIS is part of the EHR. Reports leave as HL7 v2 messages through an interface engine (e.g. Mirth Connect or Rhapsody), typically over a secure site-to-site VPN, and every site is set up a little differently. Our software integrated with what each site already had, so nobody had to change systems.

Reading reports before large language models

This was built before generative large language models, using natural language processing (NLP): smaller models each trained for one job, plus fixed rules. Reports are free text full of shorthand. Reading one meant finding each finding and its size while skipping any the report rules out (“no sign of”), translating the wording into standard medical terms (RadLex and SNOMED CT), then applying the published guidelines from the American College of Radiology and the Fleischner Society to decide which follow-up the finding needs and when. It covered ten kinds of incidental finding, among them lung and thyroid nodules, kidney, liver and adrenal masses, pancreatic cysts and abdominal aortic aneurysms. We built our own labelled set from a partner’s messy real-world reports to train and test it.

Closing the loop with care coordination

Each follow-up then had to be tracked and managed until the loop was closed, meaning the next scan actually happened for that patient. The application layer kept a worklist of patients who needed one, watched scheduling to see whether each was booked, matched it to a later exam for the same patient, and reached out to the doctor and the patient by text, email, letter or the doctor’s own systems when it wasn’t.

Our first product: an intelligent worklist

In 2019 I wrote an article arguing that radiology AI’s hard part would be fitting into hospital workflow. Our first product took that on for radiology groups: it decided which radiologist reads which scan next, working inside the PACS and RIS radiologists already used. Radiologists read from a shared queue of scans called a worklist. The most urgent scans, such as a suspected stroke or an emergency-room case, must come first, each against a turnaround deadline. A scan can only go to a radiologist licensed in the patient’s state and credentialed at that hospital, ideally one who specializes in that part of the body, and the work has to stay balanced across a team spread over many sites. Each site describes the same scan in its own free text, so routing needed language processing too. Piloted at a national radiology company present in 28 states.