Commercial AI closes in on onsite cytopathology during EUS tissue acquisition
Access to rapid onsite evaluation remains one of the most practical bottlenecks in endoscopic ultrasound-guided tissue acquisition. In this prospective single-center study, commercially available artificial intelligence systems performed unexpectedly well when asked to answer two of the most operationally important questions in the procedure room: is the specimen adequate, and does it look neoplastic?
The answer was yes for adequacy and for neoplastic lesions. The models matched cytopathologists on those tasks, a finding with immediate relevance for endoscopy units that cannot always staff an onsite cytopathologist during EUS-guided tissue acquisition. Benign disease was a different story.
Specimen adequacy is the first gatekeeper in EUS-guided sampling
EUS-guided tissue acquisition has become a core method for diagnosing gastrointestinal malignancies, but the diagnostic chain is only as strong as the specimen that reaches the slide. When tissue is scant, crushed, or poorly representative, the entire procedure can fail before pathology has a fair chance to render a diagnosis. Rapid onsite evaluation helps prevent that failure by confirming in real time whether additional passes are needed.
The clinical logic is straightforward. If a specimen is adequate, the endosonographer can stop. If it is not, the operator can redirect the needle path, change the target, or collect additional material before the patient leaves the procedure room. That is where onsite cytopathology has value: it shortens the feedback loop between sampling and diagnosis, and it can reduce unnecessary repeat procedures.
This study addressed a real workforce problem. Onsite cytopathology is not universally available because of staffing shortages, logistics, and cost. The investigators evaluated whether commercial AI could fill that gap without degrading procedural quality. Their design included consecutive patients undergoing EUS-TA of solid mass lesions, with Diff-Quik-stained rapid onsite evaluation smears and dedicated biopsy specimens for histopathology. Predominantly 22-gauge fine biopsy needles were used, which reflects everyday clinical practice in many endoscopy suites.
AI performed well where procedural decision-making is most immediate
The strongest result came in specimen adequacy assessment. Commercial AI models matched cytopathologists in determining whether a slide was adequate, which is the question that most directly drives the next move in the room. If a model can reliably identify whether the sample contains enough diagnostic material, it can function as a real-time quality check rather than a passive archive tool.
The same was true for neoplastic lesions. AI correctly classified neoplastic disease, suggesting that commercial models can recognize malignant or pre-malignant cytomorphologic patterns well enough to support procedural triage. In practical terms, this could help endoscopy teams confirm that sampling has captured the lesion they intended to sample, especially when access to expert onsite review is limited.
One sentence captures the core operational tension here: AI may be ready for the procedure room before it is ready to stand alone in the pathology lab.
The distinction matters because the study did not show equivalent performance for benign disease. Benign specimens are often more heterogeneous, more sparsely cellular, and more dependent on subtle contextual interpretation. Reactive change, inflammation, fibrosis, and sampling artifact can blur the visual boundary between nonmalignant tissue and a truly representative specimen. A model that can identify obvious neoplasia may still struggle when the clinical question turns on the absence of malignancy rather than its presence.
Why benign disease remains harder for algorithms
From a biologic and imaging standpoint, benign tissue is less standardized than cancer. Neoplastic lesions often create denser cellularity, more obvious architectural distortion, and more repetitive cytologic patterns that machine learning systems can classify with relative confidence. Benign samples, by contrast, may be thin, mixed, and variably inflamed. They can contain blood, fibrosis, gastrointestinal contaminants, or scattered normal elements that make algorithmic certainty harder to achieve.
That limitation is clinically important because the most consequential errors in gastrointestinal sampling often occur in the gray zone. A false reassurance that a bland or scant specimen is truly representative can delay diagnosis. A false call of adequacy can also prompt premature cessation of sampling, especially in lesions where malignant cells are patchy rather than diffuse.
The study’s message is therefore not that AI replaces human expertise wholesale. It suggests a narrower and more useful role: commercial AI may serve as a live safeguard during EUS-TA, flagging whether the sample is usable and whether the collected material contains neoplastic features. In that capacity, the technology could reduce procedural error, improve efficiency, and support less experienced operators.
A quality-assurance role fits the current evidence better than autonomy
The most defensible near-term use of commercial AI is as an adjunct, not a substitute. In a center that already uses onsite cytopathology, AI could reinforce quality assurance by confirming representative sampling and giving an immediate second read. In centers without onsite cytopathology, it may partially close the access gap, especially where the alternative is no rapid review at all.
This approach also fits the study’s real-world framing. The investigators tested commercially available systems in a prospective setting rather than a purely experimental one, which makes the findings more relevant to day-to-day endoscopy practice. Still, single-center performance is not the same as broad generalizability. Commercial tools can behave differently across institutions, staining workflows, lesion types, and operators, and the article itself notes that such algorithms have not yet been validated across diverse settings for routine clinical use.
For now, the key question is not whether AI can outthink cytopathologists. It is whether AI can reliably prevent wasted passes, identify inadequate sampling before the patient leaves the room, and preserve diagnostic yield where onsite expertise is unavailable. This study suggests that answer may be yes for adequacy and neoplastic lesions, but not yet for benign disease.
That leaves a clear boundary line for deployment: use AI to confirm representative sampling and support the procedure, not to eliminate human review where the cytologic interpretation is least certain. Wider validation across settings and lesion types will determine how far that boundary can move.