Before Histolyx existed, it was a repeated conversation. Not one conversation, not a pivot moment — just the same frustration surfacing over and over, in different rooms, with different radiologists, in enough different ways that eventually I stopped thinking of it as individual complaints and started thinking of it as a structural problem.
I had been working on medical image analysis research in the context of the San Diego biotech and academic medical corridor. The density of imaging centers here — the independent practices, the UCSD-affiliated outpatient sites, the community hospital radiology departments — gave me access to a cross-section of how radiology actually works at the operational level, not the academic-paper level.
The Conversation That Kept Repeating
The version that sticks with me most clearly was with a radiologist at an outpatient imaging center in the Kearny Mesa area — a facility doing around 200 to 250 studies a day with three staff radiologists. We were talking about workflow, not AI. I was asking genuinely dumb questions about how they managed their worklist, what cases they hated opening, where in the day their attention was sharpest.
His answer was immediate: the cases he dreaded most were not the hardest ones. They were the routine screenings at the end of a heavy shift. Not because the cases were complex, but because the stack was long, the cases looked normal at a glance, and his brain knew from experience that the normal-looking cases were exactly where he was most likely to miss something subtle. "The dangerous cases are the ones that don't look dangerous," he said. That is not a novel observation — it is documented in the radiology cognition literature. What struck me was how precisely he described the attention degradation: he knew it was happening and he had no tool to compensate for it.
I had heard versions of that same observation from researchers at Scripps, from radiologists at independent practices in Hillcrest and Clairemont, from a chest radiology fellow doing a rotation at a community hospital. The common thread was not workload — everyone in medicine has workload. It was the specific problem of sustained high-volume screening: too many cases where the base rate of positive findings is low, so attention drifts, and the one case that needed careful eyes gets 60% of the attention it deserved.
Why We Did Not Build an Autonomous System
The obvious AI response to that problem is an autonomous system that reads studies and issues verdicts. That is also the wrong response, and getting this distinction right was the single most important design decision we made early on.
An autonomous diagnostic system changes who is responsible for a finding. A pre-reading system that highlights regions of interest for radiologist confirmation does not. The radiologist is still reading the study — they now have a navigational aid that surfaces what the model found suspicious, so they confirm, measure, and report. Their judgment is still the clinical output. That is not just a regulatory framing; it is what radiologists actually wanted when we asked them.
The radiologists we talked to were not asking to be replaced or even substantially bypassed. They were asking for help not missing things at the end of a shift. That is a much narrower problem than "automated radiology," and it has a much more tractable technical solution: a pre-reading layer that marks findings before the radiologist opens the study, so their cognitive load at study-open is sorting and confirming rather than hunting from scratch.
We are not saying autonomous radiology AI has no future. We are saying it is not the right tool for the problem we set out to solve, and building the wrong tool well is still building the wrong tool.
The San Diego Ecosystem Advantage
We are based in La Jolla, at 4510 Executive Drive, in the dense biotech and medtech cluster between Torrey Pines Road and the 5. Being here is not incidental. San Diego has one of the highest concentrations of biotech and medical device companies per square mile of any US metro, with an academic medical infrastructure anchored by UCSD's health system and Scripps. For a small company building clinical software, that means access to the right conversations — not as paying customers, but as domain validators.
I want to be precise about what that means, because it is easy to overstate. We do not have a formal partnership with UCSD. We did not come out of a UCSD lab. What we have is geographic proximity to a dense network of radiologists and researchers who will take a meeting, look at an early build, and tell you plainly when something does not reflect how clinical workflow actually operates. That feedback loop matters enormously when you are building for a domain with high stakes for getting things wrong.
The San Diego radiology community also skews toward digital adoption. A meaningful fraction of the independent and hospital-affiliated practices here have been doing digital workflow for long enough that they are past the "we just switched to PACS" phase and into "we want to optimize what we already have." That is exactly the customer profile where a pre-reading layer adds value: established digital workflow, known pain points, openness to tooling that does not require replacing their existing system.
What We Got Wrong Initially and Fixed
The first version of Histolyx's output was too verbose. We were surfacing a large number of regions per study — partly because our recall-tuned thresholds were aggressive, partly because we thought more information meant more value. Radiologists told us quickly that was wrong. More annotations per study means more scanning time to review the annotations, which partially negates the efficiency benefit. What they wanted was fewer, more confident flags — the regions the model was genuinely unsure about, ranked so the most likely findings were first.
We rebuilt the output ranking layer based on that feedback. The model now surfaces a configurable number of regions per study, ranked by confidence, and defaults to a threshold that errs toward fewer high-confidence flags rather than more low-confidence ones. Labs can adjust that threshold, and we track annotation acceptance rates to help individual sites find the setting that works for their radiologist team.
That adjustment sounds small in retrospect. At the time it required rethinking what the product was for: not comprehensive AI coverage of every possible abnormality, but a focused pre-reader that changes how the radiologist starts a case. That framing is now in everything we build.
Where We Are Now
Histolyx is a small team. We have not raised a large round. We are not claiming to have solved radiology workflow. What we have built is a pre-reading tool for chest CT, mammography, and digital pathology slides that reflects what radiologists in this city told us they actually needed — not what the industry press said they should want, and not what a product roadmap built for investor pitch decks would have produced.
We are still learning. The conversations that started this company are still happening — with radiologists, with imaging center directors, with IT administrators who have to make DICOMweb integration actually work. Every one of those conversations changes something small in what we build. That is the only honest way to describe an early-stage product in a domain this consequential.