LO
Pathologist — tasks, one by one
The unit of analysis is the task, not the job title. Each one below carries its direction, whether the judgement rests on evidence or on platform inference, the reasoning, and what it does not establish.
Every task on this page#
Detecting and grading cancer
Being augmented✓ Evidence-backedFinding cancer in biopsies and assigning its grade, such as the Gleason grade for prostate cancer.
This is where AI is furthest along, and it arrives as a second reader. The FDA cleared Paige Prostate as an adjunct whose output should not be used as the primary diagnosis, and later cleared a second-read tool that flags cases first diagnosed as benign for another look. In a prospective study of 1,613 cases across three NHS centres, AI second reads changed the diagnosis or grade for 21 of 386 patients (5.4%) and reduced cases needing extra stains at all sites. In the PANDA challenge, algorithms reached pathologist-level agreement on Gleason grading in validation sets from two continents.
Almost all of it is prostate biopsy; the clearances are add-on only, and the prospective study and one deployment were written or funded by the developers.
Searching slides for small findings
Being augmented✓ Evidence-backedLooking through many slides for small things, such as metastases in lymph nodes.
The search is the part machines take most naturally. In a reader study, pathologists assisted by an algorithm detected micrometastases in lymph nodes more often (91% against 83%) and took about half the time per image (61 against 116 seconds). The UK's Royal College of Pathologists names searching lymph nodes for cancer as routine, time-consuming work AI could take on.
A 70-slide reader study by the algorithm's developers and a professional body's view; neither measures workloads in practice.
Scoring biomarkers
Being augmented≈ Platform inferenceQuantifying stains such as PD-L1 that decide whether a patient gets a treatment.
Counting is where pathologists disagree most, and AI narrows the spread. In a ring study of 21 pathologists scoring PD-L1 in breast cancer, agreement rose from an intraclass correlation of 0.618 by eye to 0.931 with AI assistance, and 80% of the AI results were accepted.
One biomarker in one cancer, with a model built by the authors; it does not show routine use.
Frozen sections during surgery
Still human-led✓ Evidence-backedReading a rapidly frozen sample while the surgeon waits, and calling the result into the theatre.
Frozen sections are harder material, and the first digital-slide clearance left them out: the FDA's 2017 decision for Philips' system says it is not intended for use with frozen section. A study of frozen donor-kidney biopsies found that frozen sections cause variable scoring and inappropriate discard, and that its model's scores tracked transplant outcomes — but it is retrospective and about organ allocation, not tumours during surgery.
No record here measures AI on intraoperative tumour frozen sections; the only frozen-section study is retrospective and about donor kidneys.
Signing out the diagnosis
Still human-led✓ Evidence-backedIntegrating everything into the final report and answering for it.
Every tool cleared so far keeps the diagnosis with the pathologist. The FDA says Paige Prostate's output should not be used as the primary diagnosis and must be used with a complete standard-of-care evaluation; the second-read tool sends flagged cases for further review by a pathologist. The Royal College of Pathologists says pathologists will still be needed to interpret and produce an overall pathological assessment.
Regulatory scope and a professional position; they describe the rules now, not what future clearances will allow.