Computational Ki-67 Quantification: Where Image Analysis Helps and Where It Doesn't

For decades, the Ki-67 proliferation index has occupied an awkward middle ground in diagnostic pathology: a marker everyone agrees matters, scored by a method almost no one fully trusts. Manual counting — pathologists estimating the percentage of tumor nuclei staining positive, often by eye across a few high-power fields — is fast, cheap, and notoriously variable between observers. Computational image analysis promised to replace that estimate with something closer to an actual count. The question the field is now working through is how much of that promise has been kept, and where.

The Development

In January 2023, the FDA cleared Roche's uPath Ki-67 (30-9) image analysis algorithm for quantifying the Ki-67 proliferation index in breast cancer on whole-slide images [1]. That clearance matters beyond a single product. It anchored Ki-67 scoring in the Software-as-IVD regulatory paradigm — meaning the algorithm is treated as a diagnostic device, held to defined performance standards, and cleared for a specific, narrow intended use rather than sold as a general research tool. It was the first major cleared entry addressing the inter-observer variability problem head-on, and it set a template that subsequent tools have had to follow.

That template now governs a growing field. By one industry tracking count, the FDA had authorized 51 AI/ML-based digital pathology Software-as-IVD tools through early 2026 [2]. The practical consequence is standardization pressure: any computational Ki-67 tool aiming for clinical use — whether for breast, lung, lymphoma, or prostate — has to clear the same regulatory bar, submit validation data, and declare a bounded intended use. That's a meaningful shift from the earlier era of unvalidated scripts running quietly in the background of a research lab.

The Diagnostic Problem

Ki-67 is a genuinely cross-cancer marker, which is part of what makes standardizing it so difficult. In breast cancer, the proliferation index contributes to prognostic assessment and, in some contexts, to decisions about eligibility for particular systemic therapy classes. In lung neuroendocrine tumors, Ki-67 helps separate lower- and higher-grade categories. In lymphoma, a high proliferation fraction carries prognostic weight and helps distinguish aggressive from indolent disease. In prostate, its role is more contested but still actively studied.

The trouble is that each of these settings asks the marker to do a slightly different job, with different sampling conventions. Breast scoring emphasizes invasive tumor regions; lymphoma often depends on hot-spot sampling — finding and quantifying the most proliferative focus — a task where the choice of where to look drives the number as much as how you count. A manual estimate compounds two errors at once: sampling variability and counting variability. Automated analysis attacks the second directly. On a whole-slide image, an algorithm can classify every tumor nucleus rather than estimate across a handful of fields, which improves throughput and, more importantly, intra-observer consistency — the same slide scored twice gives the same answer.

The Evidence

The consistency gains are real, and they're the clearest thing the evidence supports. A 2025 peer-reviewed synthesis on digital pathology and AI quantification in oncology documents where AI-assisted Ki-67 scoring improves on manual reading: throughput and reproducibility of counting on a fixed image [5]. That's precisely the value proposition behind the uPath clearance [1].

But the same review is candid about where automation doesn't rescue the score. The failure modes it names are concrete: cross-platform drift, where the same algorithm produces different results on images from different scanners, and the absence of ground-truth consensus for tasks like lymphoma hot-spot selection and prostate grading contexts, where experts don't fully agree on what the right answer even is [5]. An algorithm can be perfectly reproducible and still be reproducibly wrong if the region it scores was chosen by an inconsistent rule.

This is where a technical assessment from the MGH informatics group — circulated as a technical report rather than a peer-reviewed publication — makes a pointed argument worth weighing carefully [3]. It identifies tissue preparation and scanner calibration as dominant sources of quantification error sitting upstream of any classifier. Fixation time, staining protocol, and scanning parameters all shift the pixel-level input before the AI ever sees it. The broader whole-slide imaging literature has long recognized pre-analytic variance as a first-order problem, and the implication here is uncomfortable: a cleared, well-validated algorithm cannot compensate for variance introduced before the image exists. Two labs running the identical software on identically-stained tissue can diverge if their fixation and scanning differ.

Where It Stands

For a practicing pathologist, the honest reading is a division of labor. In breast cancer, the cleared uPath tool offers a validated, regulated workflow for Ki-67 quantification, and its intended use is defined precisely enough to be relied upon within those bounds [1]. Outside breast — in lung neuroendocrine, lymphoma, and prostate contexts — computational Ki-67 quantification remains largely investigational, constrained by both the unresolved sampling conventions and the standardization gap the International Ki-67 Working Group has spent years trying to close.

That gap is now a matter of formal policy attention. A 2024 Friends of Cancer Research white paper called for harmonized validation frameworks for computational pathology biomarkers in both drug development and clinical care, explicitly naming the distance between image-analysis capability and cross-site standardization [4]. The argument is that no single cleared product closes this gap — reproducibility across sites requires multi-stakeholder agreement on staining, scanning, and scoring conventions that no algorithm can supply on its own.

The open question that most directly shapes a pathologist's next decision is what pre-analytic standardization would actually have to look like before Ki-67 numbers could be trusted across institutions. Before a value scored in one lab can be compared to one scored in another — for a multi-site trial, a second opinion, or a shifting treatment threshold — the field would need harmonized fixation windows, standardized staining protocols, and calibrated scanning defined tightly enough that the image entering the algorithm is genuinely comparable [3][4]. Until that upstream harmonization exists, a computational Ki-67 index is best read as reproducible within a lab and validated for its cleared use, not as a portable number. Knowing which of those two things you're holding is, for now, the pathologist's judgment to make.

---

References

1. Roche Diagnostics. uPath Ki-67 (30-9) image analysis, Breast — Product Page. Roche Diagnostics US (diagnostics.roche.com), 2023. https://diagnostics.roche.com/us/en/products/digital/upath-ki-67-30-9-image-analysis-breast-pid00000380.html

2. Shrestha Y. AI/ML in Digital Pathology and the Software-as-an-IVD Paradigm: 2026 Snapshot. Innolitics (innolitics.com/articles), 2026. https://innolitics.com/articles/ai-ml-in-digital-pathology-and-the-software-as-an-ivd-paradigm-snapshot

3. He S. Computational Pathology: From Whole-Slide Imaging to AI-Assisted Diagnosis. MGH/Harvard Medical School Medinformatics Lab Technical Report (medinformatics.mgh.harvard.edu), 2026. https://medinformatics.mgh.harvard.edu/resources/computational-pathology.html

4. Friends of Cancer Research. Supporting the Application of Computational Pathology in Oncology. Friends of Cancer Research White Paper (friendsofcancerresearch.org), 2024. https://friendsofcancerresearch.org/publication/supporting-the-application-of-computational-pathology-in-oncology

5. PMC Article PMC12861029. Applications and challenges of utilizing digital pathology and AI in oncology. PubMed Central (PMC12861029), 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12861029