AI slashes unnecessary thyroid FNAs

A thyroid nodule is a lump in the thyroid, and ultrasound finds them so often that EndoText says up to 60% of people may have one on imaging. Most are benign, and the usual cancer risk is about 7% to 15%, which is why the hard part is deciding who actually needs a needle biopsy. (endotext.org) That biopsy is called fine-needle aspiration, which means a clinician uses a thin needle to pull out a tiny sample of cells from the lump. It is a standard test in thyroid care, but EndoText notes that even a well-done sample can come back nondiagnostic in about 5% to 10% of cases because the slide is too sparse or too bloody to read. (endotext.org) Doctors do not biopsy every thyroid nodule, because ultrasound patterns and size are supposed to act like a triage system before the needle ever comes out. The American College of Radiology built Thyroid Imaging Reporting and Data System, or Thyroid Imaging Reporting and Data System, to standardize which nodules should be watched and which should be sampled. (acr.org) Those rules already cut down on extra procedures, but they still send many benign nodules to biopsy. A 2021 American Journal of Roentgenology review said the American College of Radiology system reduced unnecessary biopsies of benign nodules by about 19.9% to 46.5% compared with other risk systems, which shows both the progress and the remaining waste. (ajronline.org) The new wrinkle is using artificial intelligence as a “goalkeeper” in front of the biopsy decision, like adding one more reviewer before a shot reaches the net. In a 2025 clinical study highlighted in Endocrine Practice, investigators tested artificial intelligence in real-world thyroid nodule management to identify nodules that radiologists had flagged for biopsy but were likely benign. (sciencedirect.com) That study reported the number getting attention online: the unnecessary biopsy rate fell from 68.5% to 9.1% when the artificial intelligence filter was applied. In plain English, most of the biopsies avoided by the model were biopsies that would have landed on benign nodules anyway. (sciencedirect.com) A Mayo Clinic commentary published in 2026 framed that setup as diagnostic de-escalation, which is a clinical way of saying “do less when less is enough.” The piece describes the model as a tool for health care resource stewardship in thyroid clinics, where high imaging volume can push many patients toward procedures that do not change treatment. (sagepub.com) The catch is that a lower biopsy count is only a win if the right nodules are the ones being deferred. The American College of Radiology white paper says even highly suspicious nodules smaller than 1 centimeter are often not biopsied right away under existing rules, so any artificial intelligence layer has to fit into a system that already accepts some watchful waiting. (jacr.org) There is also a second bottleneck after the needle: what happens to the sample once it is taken. EndoText says thyroid biopsy results can be benign, malignant, indeterminate, or nondiagnostic, and indeterminate samples often trigger molecular testing to sort out who needs surgery and who does not. (endotext.org) So the real scorecard is bigger than one headline percentage. If artificial intelligence sharply reduces biopsies, clinics will want to know the missed-cancer rate, the follow-up schedule for deferred nodules, the adequacy rate of the biopsies that still get done, and whether fewer samples also means fewer chances to run downstream molecular tests on borderline cases. (sciencedirect.com)