AI for R&D

Computer Vision for Automated Pathology Slide Analysis

ANG Associates
Life Sciences & AI Consulting

Sep 2025 6 min read

Background

Digital pathology converts glass histopathology slides into gigapixel whole-slide images that can be viewed, shared, and analyzed on screen instead of under a microscope. That shift in format is what makes computer vision possible: a deep learning model can only "read" a slide once it exists as pixels. Over the past five years, convolutional neural networks and, increasingly, vision transformers have been trained on these whole-slide images to support pathologists in three recurring tasks: detecting and grading tumors, counting mitotic figures, and scoring biomarkers such as HER2 and Ki-67 in immunohistochemistry-stained tissue. The clinical motivation is straightforward. Pathology departments face rising case volumes and a shrinking specialist workforce, and manual tasks like mitotic counting are known to suffer from significant interobserver variability between pathologists looking at the same slide. Computer vision does not replace the pathologist's judgment; it is positioned, in essentially every cleared or CE-marked product on the market today, as a decision-support layer that flags regions of interest, quantifies what a human would otherwise estimate by eye, and provides a second read.

How the models are used in practice

The clearest commercial application is prostate cancer detection on core needle biopsies. Paige Prostate, authorized by the FDA in 2021 as the first AI product in digital pathology, and Ibex Medical Analytics' Galen/Prostate Detect platform both work by scanning biopsy whole-slide images and generating heatmaps that direct the pathologist's attention to tissue regions suspicious for cancer, catching foci that might otherwise be read as benign on a busy sign-out day. A related and growing category is quantitative biomarker scoring: a 2025 multireader study in JCO Precision Oncology evaluated a fully automated AI solution for HER2 immunohistochemistry scoring in breast cancer, aiming to standardize a test whose manual scoring has long been recognized as inconsistent between labs. Mitotic counting, used for tumor grading in breast cancer and many other malignancies, is another target: a controlled reader study published in Veterinary Pathology had 23 pathologists across 11 laboratories count mitotic figures in canine mast cell tumor slides with and without deep learning assistance, and found that computer assistance improved both detection performance and agreement between readers. Beyond single-task tools, whole-slide scanner and viewer platforms such as Roche's Digital Pathology Dx and Indica Labs' HALO AP Dx have been cleared for primary diagnosis itself, meaning a pathologist can render an official diagnosis from the digital image on screen rather than the original glass slide, a workflow milestone that computer vision tools are increasingly layered on top of.

Tumor detection and localization on biopsy and resection specimens, surfaced as heatmaps or overlays
Mitotic figure counting and hotspot selection to support tumor grading
Quantitative immunohistochemistry biomarker scoring, such as HER2 and hormone receptor status
Multimodal risk stratification that combines histopathology image features with clinical variables

Evidence and regulatory pathways

Regulatory clearance for these tools follows two main routes. In the United States, novel AI pathology software with no existing predicate device typically enters through the FDA's De Novo pathway, as Paige Prostate did in 2021; once a predicate exists, follow-on products and expansions can use the faster 510(k) pathway, which Ibex Medical Analytics used for its Prostate Detect clearance in February 2025 and Roche used for its Digital Pathology Dx system in June 2024. Some newer entrants combine image features with clinical data under a multimodal AI designation, as Artera did when it received 510(k) clearance in 2026 for a breast cancer risk stratification tool built on digitized histopathology plus patient variables. In Europe, these products are regulated as in vitro diagnostics, and Ibex's Galen Prostate became the first standalone AI-powered cancer diagnostics solution to obtain CE marking under the stricter IVDR framework rather than the legacy IVDD regime.

The evidence base behind these clearances is thinner and more uneven than the headlines suggest. A 2024 review in npj Digital Medicine examining 26 CE-marked or FDA-approved AI products for digital pathology found that only 38% had a peer-reviewed internal validation study and only 42% had a peer-reviewed external validation study publicly available, with independent, vendor-uninvolved evidence present for only about 17% of publications.

"Publicly available information for products can be variable, with few sources of independent evidence," the authors wrote, describing this as a transparency gap in the AI-driven digital pathology market.

For diagnostics developers and the laboratories that adopt these tools, that finding is a reminder that regulatory clearance is a floor, not a substitute for rigorous, site-specific validation before a model touches a real diagnostic workflow.

Where a firm like ANG Associates fits in

Deploying a computer vision pathology tool inside a regulated lab or pharma environment is not primarily a machine learning problem; it is a validation, integration, and change management problem. ANG Associates works with diagnostics and life sciences organizations at exactly that intersection: helping teams define an AI adoption strategy that matches the right regulatory pathway and evidence plan to the intended use of the model, building the GxP-aligned computer system validation package a lab's own quality system and inspectors will expect (installation, operational, and performance qualification, data lineage, and change control for model updates), and running the IT delivery work, whether through SAFe or another agile framework, needed to integrate a cleared algorithm into an existing laboratory information system and imaging pipeline without disrupting sign-out. This is deliberately unglamorous work. It does not produce a new algorithm, and it does not claim to close the evidence gaps documented in the literature. It is the discipline that lets an organization deploy a genuinely validated tool with confidence, on a timeline and audit trail that hold up under regulatory and clinical scrutiny.

Sources

digital pathologycomputer visionAI diagnosticsFDA De Novo clearanceGxP validation

Interested in this topic?

Let's discuss how we can apply these approaches to your organization.