Digital Pathology in Oncology Biomarker Validation: Where It Works and Where It Fails

Digital Pathology in Oncology Biomarker Validation: Where It Works and Where It Fails

Digital pathology is no longer a pilot technology. Whole slide imaging systems have been cleared for primary diagnosis in the United States since 2017, professional bodies on both sides of the Atlantic have published validation guidance, and NHS laboratories are reporting live cases on networked digital platforms. For oncology biomarker teams, the question has moved from “can we read slides on a screen?” to “when does a digital read make a biomarker result more reliable, and when does it quietly make it worse?”

This article takes the infrastructure and validation view. It does not assess the evidence for AI models as discovery tools; that is covered in our separate review of AI biomarker discovery evidence. Instead it looks at whole slide imaging, the guidance that governs its validation, digital IHC scoring for PD-L1, HER2-low and Ki-67, the failure modes that matter in practice, and what pharma sponsors should specify when they run central digital reads in oncology trials.

Key takeaways

• Whole slide imaging for primary diagnosis is established: the FDA authorised the first system (Philips IntelliSite Pathology Solution) through the De Novo route in April 2017, and later systems have followed via 510(k).
• Validation is local and use-specific. The CAP guideline (updated 2021) recommends at least 60 cases per application with a two-week washout, while the Royal College of Pathologists emphasises validation in real-world practice.
• Image analysis helps most where manual scoring is a counting problem (Ki-67, percentage-positive scores) and least where the call depends on faint intensity at a threshold (HER2 0 versus 1+).
• Most failures happen upstream of the algorithm: fixation, staining, scanner differences, colour handling, tumour heterogeneity and region-of-interest selection.
• Computational companion diagnostics are arriving: Roche’s TROP2 device received the first FDA Breakthrough Device Designation for a computational pathology CDx in April 2025 and is being used prospectively in a phase 3 trial.
• In the UK, NHS digital pathology is progressing network by network rather than through a single national switch-on.

Whole slide imaging: from novelty to primary diagnosis

The regulatory history is short but important, because it defines what a digital read is legally allowed to do.

In April 2016 the FDA finalised its guidance on technical performance assessment of digital pathology whole slide imaging devices. It set out system-level tests for colour reproducibility, spatial resolution, focus quality, whole slide tissue quality and stitching error, plus component testing of the light source, slide feeder, sensor and software. A year later, in April 2017, the FDA granted the De Novo request for the Philips IntelliSite Pathology Solution, the first whole slide imaging system authorised for primary diagnosis of surgical pathology slides from formalin-fixed, paraffin-embedded (FFPE) tissue. The supporting study covered roughly 2,000 cases read by 16 pathologists, and the decision placed whole slide imaging in Class II, creating a predicate for later 510(k) clearances. Frozen sections, cytology and non-FFPE haematopathology were outside the intended use.

Other vendors have followed. Roche, for example, announced 510(k) clearance for Roche Digital Pathology Dx (VENTANA DP 200) in June 2024. The practical point for biomarker teams is that clearance covers a defined system (scanner, viewer, display) for a defined use. A primary-diagnosis clearance does not mean the same images are validated for quantitative biomarker scoring.

Validation guidance: CAP and the Royal College of Pathologists

CAP (United States)

The CAP guideline on validating whole slide imaging for diagnostic purposes, updated with ASCP and the Association for Pathology Informatics and published in Archives of Pathology & Laboratory Medicine (2022;146:440–450), contains three strong recommendations and nine good practice statements. The core requirements are a validation set of at least 60 cases for one application, an additional 20 cases for each further application (immunohistochemistry or special stains, for example), intraobserver concordance between glass and digital, and a washout of at least two weeks between reads. Laboratories should be concerned if overall concordance falls below 95%, and every discordance should be reconciled. The panel found that going beyond 60 cases did not improve mean concordance.

Royal College of Pathologists (UK)

The RCPath best practice recommendations for implementing digital pathology (first published January 2018) take a different emphasis: validation of each pathologist in real-world practice, building confidence case by case with glass slides available, rather than a fixed retrospective set. The approach drew on the Leeds training and validation protocol. NHS England’s guidance on digital histopathology in the NHS cancer screening programmes tells providers to apply the RCPath recommendations for implementation, validation, ongoing monitoring and quality assurance.

Neither framework is designed to validate quantitative biomarker scoring. Both establish that a pathologist reaches the same diagnosis on screen as on glass. A trial that uses digital images for a PD-L1 tumour proportion score or a HER2 category needs a separate analytical validation for that endpoint, with its own concordance criteria at the clinical cut-off.

Digital IHC scoring: PD-L1, HER2-low and Ki-67

PD-L1

PD-L1 is the clearest example of a biomarker where the assay, the scoring system and the reader all contribute variability. The Blueprint phase 2 study (Tsao et al., J Thorac Oncol 2018) stained 81 real-world lung cancer samples with five assays. It found 22C3, 28-8 and SP263 broadly comparable for tumour cell scoring, while SP142 stained fewer tumour cells, and it reported good agreement between scores made on glass and on digital scans. For tumour proportion score, image analysis is well suited: it is a percentage of positive tumour cells, and cell detection plus classification can reduce the drift that comes from estimating by eye. Immune cell scores and combined positive scores are harder, because the denominator and the definition of a relevant immune cell are less tractable for an algorithm. Our clinical review of PD-L1 as a predictive biomarker covers the scoring systems in more detail.

HER2-low and HER2-ultralow

HER2 shows where manual scoring fails and where digital scoring has not yet fixed it. In a 2022 JAMA Oncology study (Fernandez et al.) that drew on CAP proficiency surveys and a cohort of 170 breast biopsies, 18 pathologists reached at least 90% agreement on 0 versus 1+ in only 26% of the relevant cases, compared with 58% for 2+ versus 3+. That matters because the PATHWAY HER2 (4B5) assay has been the FDA-approved companion diagnostic for HER2-low status since 2022, and in January 2025 its label was expanded to HER2-ultralow for trastuzumab deruxtecan in HR-positive metastatic breast cancer. The clinically meaningful thresholds now sit at the faint end of the staining range.

Image analysis can measure membrane staining continuously and consistently, but it inherits every source of intensity variation: antibody clone, platform, fixation, section thickness and scanner colour response. At the 0 / ultralow / 1+ boundary, a small calibration shift moves patients between categories. For ADC programmes, where expression thresholds increasingly decide eligibility, this is the central validation problem; see our analysis of ADC and immunotherapy patient selection.

Ki-67

Ki-67 is the case where digital scoring most clearly helps. The International Ki67 in Breast Cancer Working Group (Nielsen et al., JNCI 2021) concluded that analytical validity can be achieved with careful attention to pre-analytics and calibrated, standardised scoring, and endorsed use of Ki-67 of 5% or less, or 30% or more, to estimate prognosis in stage I/II ER-positive, HER2-negative disease. Ki-67 is fundamentally a counting task, which is where automated nuclear detection outperforms human estimation, and independent work has shown high agreement between guideline-based algorithms and reference scores. The intermediate range remains clinically contested, which no amount of automation will resolve.

Where digital pathology fails

When digital biomarker scoring fails, the algorithm is rarely the root cause. The common failure points are upstream.

• Pre-analytics. Cold ischaemia time, fixation duration and block age change antigenicity. An algorithm will quantify degraded staining precisely and confidently. The IKWG made pre-analytical control a precondition for valid Ki-67, and the same logic applies to every quantitative IHC endpoint.
• Staining variability. Different autostainers, antibody lots and retrieval conditions shift intensity distributions. This matters most for thresholds defined by intensity (HER2) and least for binary nuclear calls.
• Scanner differences. Scanners differ in optics, sensors, compression and colour processing. Models trained on one laboratory’s images often perform worse on another’s, largely because of scanner and staining differences (StainCUT, J Imaging 2022). A multi-site trial that allows any scanner has introduced a variable it has not measured.
• Colour normalisation. Stain normalisation and colour augmentation can improve generalisation, but they also alter the very intensity signal that HER2 or TROP2 scoring depends on. Normalisation should be a validated, locked step, not an ad hoc fix.
• Tumour heterogeneity. A biopsy samples a fraction of the tumour. Digital analysis can quantify heterogeneity within the slide, but it cannot correct for an unrepresentative sample, and primary-versus-metastasis discordance remains a clinical reality.
• Regions of interest. Who draws the tumour annotation, and to what rules, often decides the score. Hot-spot versus global Ki-67, exclusion of in situ components and handling of crush artefact all need written rules and reader training.

Multiplex immunofluorescence and spatial images

Multiplex immunofluorescence and spatial platforms add marker co-expression and cell-to-cell distances to the picture, which brightfield IHC cannot provide. The validation challenge grows accordingly: spectral unmixing, autofluorescence, panel-specific antibody performance, cell segmentation and phenotyping thresholds each add variability, and none is covered by primary-diagnosis clearances or the CAP and RCPath frameworks. Most of these readouts remain exploratory or used for hypothesis generation in trials. Where spatial data are used to select patients, the analytical validation burden looks like a complex companion diagnostic, not an image viewer. We compare the main platform families in spatial proteomics vs spatial transcriptomics.

The TROP2 device points to where this is heading. Roche’s VENTANA TROP2 (EPR20043) RxDx Device combines an IHC assay, specified scanners and an algorithm that computes a normalised membrane ratio. It received FDA Breakthrough Device Designation in April 2025, and in January 2026 Daiichi Sankyo and AstraZeneca began the TROPION-Lung17 phase 3 trial with prospective central TROP2 NMR testing using the investigational device. The score is one a pathologist cannot estimate by eye, so the whole system, not just the reader, becomes the regulated diagnostic.

NHS digital pathology: where the UK stands

The UK has no single national switch-on date for digital pathology. Progress has come through regional networks and centres of excellence. In 2020 the government announced £50 million to expand three digital pathology and imaging AI centres of excellence, with digital upgrades planned for an additional 38 NHS trusts. The National Pathology Imaging Co-operative (NPIC), based at Leeds Teaching Hospitals, has become the most visible national network: all six trusts in the West Yorkshire network were live by April 2024, three North Central London trusts went live in 2025, and in July 2026 the network reached 15 connected NHS sites, with two children’s hospitals expected to join.

For sponsors running UK trials, this means digital capability varies by site. Some NHS laboratories report primary diagnosis digitally and could supply images directly; others still ship glass. Trial protocols should not assume uniform scanner models, file formats or image management systems across UK sites.

Practical guidance for pharma trials using central digital reads

• Fix the system, not just the algorithm. Specify assay, staining platform, scanner model, scan settings, file format and software version. Treat any change as a change control event requiring bridging data.
• Validate at the cut-off. Report concordance around the clinical threshold (TPS 1% and 50%, HER2 0 versus 1+), not just overall agreement, which is inflated by easy cases.
• Write region-of-interest rules. Define tumour annotation, exclusions and minimum cell counts, and train and test readers against them.
• Control pre-analytics at sites. Collect fixation time, block age and specimen type, and set acceptance criteria before scanning.
• Plan the pathologist’s role. Decide whether the algorithm assists, verifies or replaces the human score, and how disagreements are adjudicated.
• Think about the CDx path early. If a digital score selects patients, it is likely to become part of a companion diagnostic, with the regulatory obligations that implies in the US, under IVDR in the EU and with the MHRA in the UK.

Diagnostics, scanner, image management and CRO companies supporting this work can find biomarker and CDx decision-makers through our companion diagnostics conference in London.

Continue the conversation in London

These questions are on the programme at the biomarker and immuno-oncology conference in London on 3–4 March 2027, the 2nd annual ImmunoMark Summit UK Edition, with a target of 400+ attendees. Stream 2: Biomarkers & Companion Diagnostics includes the session “Digital Pathology & Spatial Transcriptomics in Oncology” on Day 1, followed on Day 2 by “Multiplex & Programmable Biomarker Platforms” and the “Panel: Global CDx Approval Pathways – EMA, FDA & MHRA Perspectives”. The format combines keynotes, panels, case studies, pre-arranged 1:1 meetings and a Day 1 drinks reception.

Register for ImmunoMark London 2027

If your company wants to meet this audience directly, whether you supply scanners, image analysis software, IHC and multiplex assays or central laboratory services, see the London 2027 sponsorship opportunities.

FAQs

What is digital pathology in oncology?

Digital pathology in oncology means scanning tissue slides into high-resolution whole slide images that pathologists read on screen and that software can analyse. It is used for primary diagnosis, remote and network reporting, central review in clinical trials and quantitative biomarker scoring such as PD-L1, HER2 and Ki-67.

Is whole slide imaging approved for primary diagnosis?

Yes. The FDA authorised the first whole slide imaging system for primary diagnosis, the Philips IntelliSite Pathology Solution, through the De Novo pathway in April 2017, and other systems have since been cleared through 510(k). Clearance covers a defined system and intended use, and it does not on its own validate quantitative biomarker scoring.

How should a laboratory validate whole slide imaging?

The CAP guideline recommends at least 60 cases for one application, 20 more for each additional application such as IHC, and a washout of at least two weeks between glass and digital reads, with concern if concordance falls below 95%. In the UK, the Royal College of Pathologists recommends validation in real-world practice and NHS screening guidance tells providers to follow it.

Does image analysis make IHC biomarker scoring more reproducible?

It does for counting tasks such as Ki-67 and PD-L1 tumour proportion score. It is less reliable for intensity-based calls near a threshold, such as HER2 0 versus 1+, because staining, fixation, scanner and colour differences all shift the measured signal. Reproducibility depends on locking the whole system, not just the algorithm.

What is the status of NHS digital pathology?

Adoption is progressing through regional networks rather than a single national rollout. The Leeds-based National Pathology Imaging Co-operative reached 15 connected NHS sites in July 2026, including the West Yorkshire network and several London hospitals, with more sites expected to join.