Separating the building from the checking — an AI foundation model for intraoperative frozen-section pathology, validated internally, externally and prospectively — a clinical trial (ClinicalTrials.gov)
A multicenter observational study putting frozen-section pathology tasks across multiple organ systems onto a single AI foundation model and testing how it performs. Retrospective development, internal validation and external validation are followed by prospective enrollment of patients actually undergoing intraoperative frozen section. Target enrollment is 33,000 and the primary outcome is the area under the ROC curve.
Trial overview (primary data)
- StatusActive, not recruiting
- ConditionsCancer, Intraoperative Pathology, Artificial Intelligence (AI)
- SponsorSun Yat-Sen Memorial Hospital of Sun Yat-Sen University
- Target enrollment33,000 participants
- Period2026-04-27 〜 2026-12-01
Key points
- A multicenter observational study developing and validating an AI foundation model for frozen-section pathology, with a target enrollment of 33,000.
- Retrospective development, internal validation and external validation are kept distinct from prospective multicenter validation. Few records spell out all three checks.
- The primary outcome is the area under the ROC curve, with accuracy, sensitivity, specificity and predictive values as secondary, on prespecified tasks across organ systems.
- The study is observational and, per the record, mandates no changes to routine clinical care for research purposes.
- What is evaluated is diagnostic performance, not operating-room waiting time or whether the decision made there changes.
1A diagnosis the operation waits for
During surgery, tissue is taken, frozen on the spot, cut thin and judged benign or malignant within tens of minutes. That is intraoperative frozen-section pathology. It exists so the extent of resection can be settled while the patient is still open, and in the operating room the surgery pauses for the answer. Section quality is poorer than in permanent specimens, and the judgment is correspondingly harder.
This study puts frozen-section tasks across multiple organ systems onto a single foundation model and sets out to measure how it performs.
2Separating the building from the checking
- 1Retrospective developmentBuild the model on past frozen-section data
- 2Internal validationCheck its behavior on data of the same origin
- 3External validationCheck it again on data from other institutions
- 4Prospective validationEnroll patients actually undergoing intraoperative frozen section, across multiple centers
The order is the substance of the design. Measured on the data it was built from, a model will of course score well, and that only confirms it was built. Only against another institution's data does it become clear whether it survives local habits of staining and equipment. Enroll prospectively on top of that, and the bias of having selected past cases falls away too. Records that spell out all three checks separately are not common.
3Scale, and how rare this is
A target of 33,000 places the study in the top 2% of the records here. That follows from the design: one model carrying tasks across organ systems needs enough cases within each organ. The primary outcome is the area under the ROC curve, with accuracy, sensitivity, specificity and positive and negative predictive value as secondary.
It is observational, and the record states that no changes to routine clinical care are mandated for research.
4The waiting is not what is measured
What the record commits to, though, is diagnostic performance, not how long the operating room waits or whether the decision made there changes. If AI for frozen section turns out to help, it should show up not only as correctness but as shorter waits and as support in the cases where the call is genuinely hard. The rung where performance is measured and the rung where results on the floor are measured are different rungs.
This trial is taking the former carefully, and taking it apart from the building of the model.
Why it matters
Writing development, internal validation, external validation and prospective validation as separate stages is itself a scale for how far a performance claim can be trusted. Few of the clinical trials held here mention external validation at all, so confirming that a model travels between institutions remains the exception. The structure transfers well beyond medicine, to any setting where model performance is claimed outside the team that built it.
FAQ
What is intraoperative frozen-section pathology?
Why does external validation matter?
Does the study change the care patients receive?
Sources (primary)
Source: ClinicalTrials.gov (U.S. NIH/NLM, public domain). This site does not provide medical advice. Verify the latest and exact details with the official source. This site is not endorsed or certified by the NIH/NLM.
- ClinicalTrials.gov (study record, original)
- NCT ID: NCT07708207