Dr. Victor Salas

What did a reading shift look like before you had AI in the workflow?

A worklist is a queue, and for most of my career it filled in the order studies arrived. A stroke CT could sit behind a stack of routine follow-ups simply because it landed later in the morning. I spent real energy triaging by hand, skimming order histories and reasons for exam to guess which case needed my eyes first. On a heavy shift across several sites, that guesswork had a cost: the genuinely time-sensitive study, an intracranial bleed or a pneumothorax, might wait while I worked through less urgent films. None of it was unsafe by design, but it leaned on memory and attention at the exact moments those are thinnest. I wanted the queue itself to recognize urgency, so I could spend my judgment on the read instead of on sorting.

What was the first AI tool you brought into your routine, and how did it go?

The first was a triage algorithm for suspected intracranial hemorrhage. It runs on the images as they arrive and pushes a flagged study toward the top of my list. The promise was simple: see the likely emergencies sooner. In practice, the early weeks were a calibration exercise. It caught real bleeds quickly, which mattered, but it also flagged motion artifact and old, stable findings often enough that I had to learn what its confidence did and did not mean. I never let a flag shorten my read, and I never let the absence of one reassure me. Treated that way, it settled into something useful: a prompt to look now rather than later. The lesson stuck. The tool reorders my attention. It does not make the diagnosis, and I stopped expecting it to.

Concretely, how does AI fit your reading routine now?

Today it works in two places. First, prioritization: several triage models watch the incoming stream and reorder my worklist so suspected critical findings, large-vessel occlusion, hemorrhage, pulmonary embolism, rise to the top with a visible tag. I still open every study and read it in full, with prior exams and the clinical history alongside, because the algorithms see a single time point and none of the chart. Second, on certain studies a detection tool marks regions for a second look, a nodule on a chest CT or a subtle fracture. I read the case first, form my impression, then check its marks against what I found. Sometimes it points at something I would have caught anyway. Occasionally it nudges me back to a corner of the image worth a longer look. Either way, the report reflects my reading.

Where do you not trust it, and how do you stay accountable?

Where the stakes are highest, which in this work is everywhere. I sign every read, so the accountability is mine regardless of what any tool suggested. The failure mode I guard against most is automation bias: the quiet temptation to under-read a study the algorithm called normal, or to over-weight a flag because the software seems confident. A model tuned for high sensitivity will over-call to avoid a miss, so false positives are a feature, not a fault, and I treat them that way. I also keep patient information inside the systems meant to hold it, and I know which tools are cleared for what. A triage flag is a prioritization aid, not a diagnosis. I would never let one stand in for my own judgment or shorten the attention a patient's images deserve.

Was there a specific moment it earned its place?

One overnight stretch stays with me, and it is representative of why I kept the tools. A head CT came in during a genuinely busy hour, and the triage model flagged it and lifted it to the top of my list. I opened it next instead of forty minutes later, saw an early intracranial hemorrhage, and the patient moved toward treatment sooner than the old queue order would have allowed. I want to be careful here: I would like to think I would have reached that study in time on my own, and often I would. What the tool changed was the margin. On the shifts where volume is highest and attention is most stretched, shaving that wait is exactly where the value sits. It did not read the scan for me. It made sure the scan that needed me most was not waiting behind the ones that could.

What would you tell a radiologist who is skeptical?

Start skeptical; that instinct is correct. But be specific about what you are skeptical of. These tools are not a second radiologist, and treating them as one is where people get hurt. Treat them as a prioritization layer and an extra set of eyes that you always overrule. Learn what each one is actually cleared to do, because triage, detection, and diagnosis are different claims with different evidence behind them. Ask how performance is monitored over time, since a model can drift as scanners and patient populations change. In our group, new tools go through review before they touch a live worklist, and we track how they behave afterward. Adopted that way, with clear eyes about the false positives and the single-time-point blindness, the technology returns attention to the reads that need it.

How has it changed your work, and the experience for patients?

The clearest change is where my attention goes. The urgent studies find me faster, so turnaround on the cases that cannot wait has tightened, and that shows up downstream as patients reaching care sooner. For the routine volume, the second-read tools give me a little more confidence that a subtle finding will not slip past at the end of a long list. For patients, most of this is invisible, and it should be. They do not see the worklist reshuffle or the flag on a chest CT. What they experience is a network that answers time-sensitive questions a bit quicker and reads with one more layer of checking behind it. My name is still on every report and my eyes are still on every image. What changed is that the system around me now spends my limited attention more deliberately than a first-in, first-out queue ever could.