A TROP2 lung cancer study presented at PathVisions 2026 found near-perfect agreement between pathologists using an AI biomarker. The variation that remained came from staining and scanning, and that changes what a lab has to control. This article is based on an educational session at the Digital Pathology Association's PathVisions 2026 meeting (October 2026), presented by Roche's medical affairs team with practicing pathologists, and on Roche's public announcements. Study figures are as presented at the session; check the published study before citing them.
A biomarker the eye cannot read
TROP2 is a cell-surface protein found in most non-small cell lung cancers and the target of antibody-drug conjugates such as datopotamab deruxtecan. When pathologists scored TROP2 staining the usual way, with an H-score, the score did not predict who responded to the drug. Read differently, the same stain did. The normalized membrane ratio (NMR) compares staining intensity at the cell membrane with intensity in the cytoplasm, pixel by pixel, for every tumor cell on the slide, as a proxy for how much of the target sits where the drug can bind and be internalized. In retrospective analyses of the TROPION-Lung01 and Lung02 trials, patients whose tumors were NMR-positive had higher response rates and longer progression-free survival.
No pathologist can calculate that ratio for every tumor cell on a slide by eye, or check the number by looking at the slide. One panelist called it a subvisual finding. In May 2025 Roche's VENTANA TROP2 RxDx Device became the first computational pathology companion diagnostic to receive FDA Breakthrough Device Designation; it is not yet approved, and the research version reports scores, not treatment eligibility.
What the study tested
Twelve laboratories in eight countries took part, with three pathologists each: academic centers, community hospitals and reference labs. The case set of 36 lung cancers was deliberately weighted toward borderline cases near the cutoff. It ran in two arms:
- Digital read: slides stained and scanned centrally, then scored by each site with the algorithm. Overall agreement was 100%.
- Full device: unstained slides sent to each lab, stained on the lab's own stainer, scanned on its own scanner and read locally. Overall agreement was 94% (96% positive, 92.5% negative), and 99.8% once borderline cases were set aside.
For comparison, panelists pointed out that manual scoring of biomarkers like PD-L1 shows far more disagreement between pathologists.
The variation moved
The headline is not that the algorithm is accurate. It is where the remaining error came from. With interpretation close to perfect, what varied between sites was the pre-analytical and system side: staining protocols, scanners, section preparation. The presenters described the test itself as a system with five parts that are validated together: the antibody and detection kit, the staining platform, the staining protocol, the whole slide scanner and image management system, and the algorithm. Change one and you have potentially changed the test.
The NMR window that separates positive from negative is narrow, so small shifts in stain intensity or scan quality can move a borderline patient across it. That is why the next round of studies looks at fixation, section thickness, block age, cytology and cryobiopsy specimens, stainers and scanners, and whether the pathologist's subspecialty matters.
The pathologist's job changes, and stays
The pathologist still signs the result. In the study workflow, quality control came in two stages: before the algorithm runs (is every piece of tissue on the slide in the image, are there at least 100 viable tumor cells, are there folds, bubbles or out-of-focus regions) and after (did the overlay find the tumor, did it count necrosis or normal tissue). A failure at either stage means a predefined action: restain, recut, rescan or reject. Artifacts that a person would read around can silently corrupt an algorithm's input.
Validation changes too. Once a test like this is cleared, a lab verifies the manufacturer's claims on its own patients with all five components matching. The College of American Pathologists is preparing guidance on AI implementation, and panelists expect proficiency testing and ongoing monitoring for drift in models, patient populations and lab processes to become routine.
What labs should do now
- Inventory the chain. Record every stainer, protocol, scanner, image management system and software version that touches a slide, with an owner for each.
- Put it under change control. A scanner firmware update, a new antibody lot or an IMS upgrade is a change to any computational test that depends on it; test it before it reaches a live case.
- Build the quality gates in. Tissue completeness, tumor content and image quality checks before analysis, with logged outcomes, not just a pathologist's memory.
- Watch for drift. Track scores and rescan rates over time so a slow shift in staining or scanning shows up before a patient is misclassified.
- Size the infrastructure for it. Scanners, storage and the network between them become part of a clinical test, and need the monitoring a clinical system deserves.
Where we come in
Most of this work is not pathology; it is systems discipline. CarolinaEMR, part of PazNetworks, designs and supports the infrastructure and integrations under digital pathology: scanner networks, storage, LIS and image management interfaces, and the configuration records and monitoring that make change control possible. See our digital pathology delivery playbook and our pages on PathAI, PathPresenter and Epic Beaker, or talk to us about your lab.
Published October 16, 2026. Study figures are as presented at the session and have not been independently verified by CarolinaEMR. VENTANA, BenchMark and navify are trademarks of Roche; Datroway is a trademark of AstraZeneca and Daiichi Sankyo. CarolinaEMR is not affiliated with these companies.
