AI use case

AI cancer detection in digital pathology slides

AI cancer detection in digital pathology is software that analyzes scanned biopsy slides, marks areas likely to contain cancer, suggests a grade and prompts a second look or reorders the worklist. A pathologist reviews the slides and signs every report.

Last verified 18 November 2026 · 2 public deployments

30%
Reported cycle time reduction
The University of Texas Medical Branch, organization claim.
USD 15,000 to USD 270,000
Indicative value per year
A pathology laboratory that reports 5,000 prostate biopsy cases a year. Worked example, see how it is calculated.

What problem does it solve?

A prostate biopsy can produce a dozen or more tissue cores on several slides, and the pathologist has to find small foci of cancer among mostly benign tissue, grade them and decide whether to order extra stains. Missed or undergraded foci change treatment, and many laboratories also order immunohistochemistry (IHC) stains to confirm borderline findings, which adds cost and days to the result. Pathologist capacity is short in many health systems while biopsy volumes keep rising.

Once a laboratory scans its slides, software can read the whole slide image. The deployments on this page use it for prostate biopsies, where approved products exist. In Articulate Pro, three NHS centres used it as a second read or a concurrent read, and the University of Texas Medical Branch (UTMB) built it into routine sign out. The software does not issue a diagnosis. It raises a flag, a pathologist decides, and the evidence so far shows fewer unnecessary stains and shorter turnaround, with a small number of diagnosis changes that need the pathologist's judgment.

How does it work?

  1. Scan the slides. The laboratory digitizes the glass slides on a whole slide scanner and sends the images to the AI product through the pathology image management system.
  2. Analyze each slide. The model marks areas suspicious for cancer, gives a probability per slide or per case and, depending on the product, suggests a grade group and flags features such as perineural invasion.
  3. Show the result to the pathologist. In a second read workflow the pathologist reads first, and the system prompts a review where its view differs. In a concurrent read workflow the pathologist sees the marks from the start.
  4. Decide and sign. The pathologist decides on additional cuts, stains or a second opinion and signs the report. The AI never reports a diagnosis by itself.
  5. Monitor. The laboratory tracks discrepancies between the AI and the pathologist, turnaround time and stain use, and checks performance after any change to a scanner or staining protocol.
Audience
Employee facing
Autonomy
Copilot
Adoption
Early adopters
Channels
Internal tools

What is it worth?

Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.

Value benchmarks for AI cancer detection in digital pathology slides
KPIMedianReported rangeData pointsClaimed by
Cycle time reductionToo few to pool
30%
11 organization

Value drivers: Employee productivity, Speed and cycle time, Risk and loss reduction.

Indicative value

A pathology laboratory that reports 5,000 prostate biopsy cases a year

USD 15,000 to USD 270,000

Annual IHC cost avoided per year

How this is calculated

Formula: cases * stainsPerCase * stainReduction * costPerStain. The low scenario uses every low input, the high scenario every high input.

InputLowHighBasis
Prostate biopsy cases per year cases, cases5,0005,000The reference laboratory.
IHC stains ordered per case before AI stainsPerCase, stains per case13Editorial assumption, replace with your own order data.
Share of IHC stains avoided stainReduction, fraction of stains0.150.3Below the figure on this page (UTMB reports 38% lower IHC utilization after implementation, in a before and after comparison of 139 cases).
Cost of one IHC stain, reagent and technician time costPerStain, USD per stain2060Editorial assumption, replace with your own laboratory cost.

What it leaves out: Gross stain cost only. It leaves out the license and integration cost of the AI product and the scanner, the cost of pathologist review time spent on AI flags, and the value of shorter turnaround and of cancers found earlier.

Which organizations already do this?

2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.

This library is not a list of Blits.ai customers. It documents public AI deployments from any vendor, each with its source.

The University of Texas Medical Branch

United States · Healthcare · 2024

ProductionGrade B

The Department of Pathology at the University of Texas Medical Branch in Galveston validated an AI assisted prostate biopsy decision support tool on routine cases within its fully digital practice and then put it into routine sign out. The validation met all predefined acceptance criteria. A before and after comparison of three month periods found shorter diagnostic turnaround and less use of immunohistochemistry, with the pathologist keeping responsibility for the diagnosis.

  • Cycle time reduction: 30%, Diagnostic turnaround time, three months after implementation against three months before
    "Following clinical implementation, diagnostic TAT decreased by 30% and IHC utilization decreased by 38%."
    Claimed by: organization

Oxford University Hospitals NHS Foundation Trust

United Kingdom · Healthcare · 2022

PilotGrade B

Three NHS specialist centres in England, led from the University of Oxford and Oxford University Hospitals, reported prostate biopsies in routine care with a commercially available AI system (Paige Prostate) alongside the histopathologist. Of 1,613 cases, 1,049 were reported with AI assistance, either as a second read after the initial diagnosis or as a concurrent read from the start. The study measured changes to diagnosis, effect on patient management and turnaround time, and the pathologist signed every report.

No outcome disclosed.

How do you implement it?

A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.

Data you need

  • Consecutive prostate biopsy cases scanned in the laboratory's own workflow, with the final pathologist diagnoses as ground truth
  • Records of IHC orders, second opinions and turnaround time before the AI to compare against
  • Scanner and stain settings documented, since performance can change when they change

Systems to integrate

  • Whole slide scanner and pathology image management system
  • Laboratory information system for orders, results and sign out
  • Pathologist viewing workstation, so the AI marks display during the read

Complexity: High

The laboratory needs a fully digital workflow first: whole slide scanners, an image management system and storage for large images. The AI product is a regulated diagnostic device bought from a vendor, so the build is the connection to the image system and laboratory information system, a local validation on the laboratory's own slides and stains, and a written sign out protocol.

  1. 1

    Digitize the workflow

    Scan all cases of the target type before any AI is added, so turnaround and stain use can be measured as a baseline.

  2. 2

    Validate on your own slides

    Run the product on consecutive past cases with known diagnoses and compare sensitivity and specificity with the vendor's claims. Repeat on a sample from each scanner.

  3. 3

    Choose second read or concurrent read

    Start with the AI as a second read after the pathologist's first view, then move to concurrent reading where the evidence supports it. Write the protocol for what happens when the AI and the pathologist disagree.

  4. 4

    Run a prospective period

    Report a defined set of cases with the AI, record every change to the diagnosis or grade, and compare turnaround and stain use with unassisted cases.

  5. 5

    Monitor performance

    Review discrepancies each quarter and after any change to the scanner, the staining protocol or the patient mix.

Guardrails

  • A pathologist reviews and signs every case, and the AI never issues a report
  • The AI runs only for the case types and scanners it was validated on
  • Local validation is repeated after a scanner, stain or software change
  • The laboratory tracks disagreement cases and reviews them with a second pathologist

KPIs to instrument

  • Turnaround time from scan completion to sign out, with and without AI
  • IHC stains per case and second opinions requested
  • Share of cases where the AI prompted a change to the diagnosis or grade, and the outcome on review
  • Disagreement between AI and pathologist by case type and scanner

Human in the loop

Pathologists read every slide and own the diagnosis and the decision to order more stains. The laboratory director owns the validation, the sign out protocol and the monitoring of discrepancies.

Common failure modes

Domain shift
A model trained on one scanner, stain or population performs differently on another. A 2026 practice review of the two FDA cleared prostate products stresses domain shift and the risk of unequal performance in under represented patients, and says clearance must be followed by local validation and ongoing monitoring.
Pathologists over trust the flag
A reader who accepts a negative AI result stops looking for small foci. Use the second read order for new users and audit a sample of cases the AI marked benign.
Alert fatigue
A model with high sensitivity and lower specificity flags benign tissue and adds review work. Track the false positive flags per case and adjust the workflow, not only the model.

What are the risks and rules?

EU AI Act

High risk

Article 6(1) with Annex I: an AI system that is an in vitro diagnostic device under Regulation (EU) 2017/746, or a safety component of one, is high risk when the device needs a third party conformity assessment. Software that analyzes slides to support a cancer diagnosis is normally classified in a class that needs a notified body, so it carries the AI Act duties for risk management, data governance, logging and human oversight on top of the device route. It is not an Annex III use. Laboratories that build and use a device only inside their own health institution can fall under the in house exemption of Article 5(5) of that regulation.

Rules that apply

Guidance

Controls to put in place

  • Documented local validation on the laboratory's own scanners, stains and case mix before any reporting use
  • Written sign out protocol that states the AI is a second reader and who decides on a disagreement
  • Audit trail that shows the AI output, the pathologist and the final diagnosis per case
  • Scheduled review of discrepancies and of performance after any change to scanner or stain

Frequently asked questions

Does AI replace the pathologist who reads a biopsy?

In the deployments on this page it does not. In the Articulate Pro study at three NHS centres the AI assisted the histopathologist, and UTMB states that the pathologist keeps responsibility for the final diagnosis. The software flags and grades, and the pathologist signs the report.

How often does AI change a prostate biopsy diagnosis?

In the Articulate Pro study, staged AI assistance as a second read changed the initial diagnosis or Grade Group in 21 of 386 patients (5.4%), and 5 of them (1.3%) could affect clinical management. The study reports the changes and does not call every one an error corrected.

What does AI do to turnaround time in pathology?

UTMB reports a 30% shorter diagnostic turnaround and 38% less IHC use after implementation, in a small before and after comparison of 139 cases. Articulate Pro reports a mean reduction of 30.1 hours with concurrent reading at one of three sites. The sites differ, so measure your own baseline.

Is pathology AI high risk under the EU AI Act?

Usually yes. Software that analyzes slides to support a cancer diagnosis is an in vitro diagnostic device that needs a notified body, which puts it in the high risk tier through Article 6(1) and Annex I. The vendor carries the conformity work, and the laboratory carries the duties of a deployer.

How to cite this page

Blits.ai AI Use Case Library, "AI cancer detection in digital pathology slides", last verified 18 November 2026, https://www.blits.ai/ai-use-cases/digital-pathology-cancer-detection. Licensed under CC BY 4.0. Method: how we verify use cases.

Changelog
  • 18 November 2026: First published

Related use cases

Healthcare

AI support for reading screening mammograms in breast cancer screening programs

AI support for screening mammography scores every examination in a population based breast cancer screening program and either triages it to single or double reading or acts as decision support for the radiologists who read it. A radiologist still makes every recall decision.

Deployments
3 public, best grade B
Reported productivity gain
44.2%
Mammography screening programme in southwest Sweden (Malmö, Lund, Landskrona, Trelleborg), organization claim
Healthcare

AI prioritization of radiology and imaging worklists

AI radiology worklist triage analyzes a medical image right after a scan and flags time sensitive findings, so the most urgent cases are read and acted on first. Examples include a brain bleed, a stroke causing large vessel occlusion or a pulmonary embolism. It reorders the radiologist's worklist and notifies the care team, and a radiologist confirms every finding before it changes a patient's treatment.

Deployments
13 public, best grade B
Reported productivity gain
44.2%
Mammography screening programme in southwest Sweden (Malmö, Lund, Landskrona, Trelleborg), organization claim
Healthcare

Autonomous AI screening for diabetic retinopathy at the point of care

Autonomous AI screening for diabetic retinopathy reads retinal photographs taken in a primary care clinic or other point of care setting and returns a result, such as referable or not referable, without a specialist reading the image first. People with a positive or ungradable result are referred to an eye specialist.

Deployments
3 public, best grade B
Reported screening completion rate
100%
Johns Hopkins Medicine, organization claim
Healthcare

AI agent for patient appointment scheduling, reminders and no show reduction

An AI agent for patient scheduling and reminders books, moves and cancels patient appointments by phone and messaging, and reminds and confirms patients. It follows the provider's scheduling rules (referral, triage level, clinician and visit type, preparation), confirms and reminds patients in two way conversations, predicts who is likely to miss an appointment, and offers freed slots to patients on the waiting list. Unlike a general branch booking or reminder agent, it writes into the electronic health record and must respect clinical constraints, so anything clinical goes to staff.

Deployments
13 public, best grade B
Median no show reduction
30%
5 deployments