What Radiology and Pathology Have Learned About Living With AI
Imaging and pathology adopted clinical AI earlier than most of medicine. Their experience offers practical lessons on worklists, validation, and the limits of algorithmic triage for every specialty.
For many physicians, clinical AI still sounds like a future event. For radiologists and, increasingly, pathologists, it is already part of the shift. Algorithms flag suspected findings, reorder worklists, measure structures, and pre screen digital slides. The FDA has cleared a large number of AI enabled devices, and imaging has accounted for most of them. That head start makes these specialties a useful preview for everyone else.
AI changes the order of work before it changes the work
Some of the most common deployments do not interpret anything for the physician. They triage. A study with a suspected urgent finding moves to the top of the list so a human reads it sooner. That sounds modest, but it reshapes the day. Cases the algorithm flags get fast attention. Cases it does not flag wait in the ordinary queue, which means a missed finding by the algorithm can quietly become a delayed read.
The lesson for other specialties is that a prioritization tool is not neutral. Any AI that decides what a physician sees first also decides what they see later, and the group should understand both sides of that tradeoff.
Local validation is not optional
Algorithms trained on one population, scanner type, or staining protocol can perform differently elsewhere. Departments that use these tools well tend to run a period of local evaluation before relying on them, comparing algorithm output against their own physicians' reads on their own cases. They also keep watching after go live, because equipment upgrades and patient mix changes can shift performance.
- Ask vendors what data the model was trained and tested on and how that compares with your population.
- Track discrepancies between algorithm output and final physician interpretation.
- Assign someone to review performance after any equipment or protocol change.
An algorithm that works well somewhere else has not yet shown that it works well here.
Keep the human read meaningful
Automation bias is a real risk. When an algorithm marks a study as normal, a busy reader may look less carefully. When it marks a finding, a reader may anchor on it and miss something nearby. Experienced departments counter this by training physicians on the tool's known failure patterns, being explicit about which outputs are advisory, and encouraging readers to form an impression before looking at algorithm markings when practical.
Pathology is facing similar questions as digital slide scanning expands. Slide level triage and quantitative measurement can help, but they require new workflows for quality control, image storage, and sign out.
If your specialty is evaluating its first clinical AI tool, talk to a radiologist or pathologist colleague about how their department chose, validated, and monitors theirs. The process they built is likely more valuable than any vendor demonstration you will see.
Dr. Reyes practices emergency medicine and writes about clinical decision making under pressure and the technology entering the ED.
This article is for professional education and does not replace clinical judgment. Treatment decisions should be based on the individual patient and current guidelines.
