Synthesis: Digital pathology and AI-assisted histology

Evidence grades (GRADE-adapted): A high — multiple well-conducted RCTs or systematic reviews converge. B moderate — single pivotal RCT or consistent observational evidence. C limited — single observational study, mechanistic, or expert consensus. D preclinical / hypothesis-generating.

Why digital pathology + AI matters now

Pathology has traditionally been a microscope-based field with substantial human-judgment-dependent components. Inter-pathologist agreement for some tasks (HER2 IHC scoring, TIL scoring, mitotic counting) is poor enough that the same physical tumor can produce different clinical decisions depending on which pathologist scored it. As biomarker-driven precision oncology has expanded, the cost of pathology imprecision has grown.

Two converging trends have made AI-assisted pathology newly viable:

• Whole-slide imaging (WSI). Slide scanners (Aperio, Leica, Roche, others) now routinely produce 20x and 40x digitized slides at clinical-grade quality, enabling digital workflows. Major academic medical centers have completed transitions to digital primary diagnosis.
• Deep learning. Convolutional neural networks and now vision transformers can extract features from images at levels approaching or exceeding human performance on selected tasks. Training datasets of consensus-annotated pathology images have grown to enable reliable model development.

HER2 IHC scoring with AI assistance

The HER2-low concept (see IHC synthesis) has made the IHC 0 vs 1+ scoring boundary clinically consequential because it determines T-DXd eligibility. Inter-pathologist agreement at this boundary is notoriously poor (kappa ~0.26 in Fernandez 2022)^[1]B. AI-assisted scoring is the most plausible near-term remediation:

• Roche Digital Pathology (Roche) and several other vendor-developed AI tools provide automated HER2 IHC scoring with reported accuracies of 85–90% vs expert consensus.
• Multiple academic-developed models have demonstrated improvement over human inter-observer agreement particularly at the 0 vs 1+ boundary.
• FDA-cleared standalone HER2-IHC scoring AI does not yet exist as of mid-2026; pathologist-assist tools that augment but don't replace pathologist judgment are deployed at multiple academic centers.

Companion-diagnostic-grade FDA clearance for AI-assisted HER2 scoring requires substantial validation across slide preparation methods, scanner platforms, and patient populations. Multiple vendors are pursuing this; first clearances are anticipated within 2–3 years.

TIL scoring with AI assistance

TIL scoring per the International TILs Working Group standard (see TME and TILs synthesis) has known inter-pathologist variability. AI-based approaches:

• Computational TIL scoring from H&E whole-slide images, trained on consensus-annotated datasets. Multiple academic and commercial models have shown high concordance with expert reference scoring.
• Quantitative TIL features beyond percentage — spatial organization (intratumoral vs stromal density), proximity to tumor cells, tertiary lymphoid structure identification — that humans don't routinely capture but AI can extract from the same images.
• The International Immuno-Oncology Biomarker Working Group has reported computational TIL scoring assessment frameworks^[2]B.

AI TIL scoring tools are research-deployed at multiple academic centers and are being incorporated into clinical trial central-review workflows. FDA-cleared clinical-deployment tools are not yet available.

Ki-67 proliferation index with AI

Ki-67 staining produces a percentage of nuclei staining positive in a tumor section. The percentage is prognostic but inter-pathologist agreement is variable. AI-assisted Ki-67 quantification is more advanced than HER2 or TIL scoring; multiple FDA-cleared tools provide automated Ki-67 quantification with high reproducibility.

In TNBC, Ki-67 is typically high (often > 30%) and doesn't currently drive treatment decisions in the way it can in HR+ early-stage disease. AI Ki-67 scoring is therefore less clinically impactful for TNBC than for HR+ disease.

Computational biomarker discovery from H&E

A separate AI pathology direction extracts biomarker information directly from H&E-stained images, without specialized IHC staining. The hypothesis: morphologic features that pathologists cannot consciously articulate but that correlate with treatment response or biomarker status may be extractable by deep learning models. Studies have shown:

• HER2 status prediction from H&E. Deep learning models can predict HER2 IHC status from H&E sections with moderate accuracy (AUC 0.7–0.8), without the patient needing the IHC test at all. Could provide screening or triage^[3]C.
• pCR prediction from pre-treatment H&E. Models that predict neoadjuvant response from pre-treatment biopsy H&E images have shown AUC 0.65–0.80, comparable to or better than some single-biomarker predictors.
• Molecular subtype prediction from H&E. Deep learning predicts intrinsic subtype and Lehmann/Pietenpol subtype from H&E with moderate accuracy. Whether this is clinically useful depends on whether it can replace molecular profiling (cheaper, faster) or only complement it.
• BRCA status prediction from H&E. Some studies have shown moderate-accuracy BRCA-deficiency prediction from H&E features, motivated by the distinctive morphology of BRCA-deficient tumors.

Regulatory pathway

FDA approval pathways for AI pathology devices vary:

• 510(k) clearance for AI tools that are substantially equivalent to predicate devices — the route most commonly used for pathologist-assist tools that augment but don't replace human judgment.
• De Novo authorization for novel AI devices without predicate — the route for standalone AI scoring devices and for biomarker prediction tools that don't have predecessor devices.
• Pre-Market Approval (PMA) for highest-risk devices — uncommon for AI pathology to date but may apply to standalone companion diagnostic AI tools.

FDA has cleared multiple AI pathology tools to date, including Paige Prostate (the first AI tool for cancer detection in any FDA-cleared category), Roche Digital Pathology suite, and several others. Breast cancer-specific AI pathology clearances are accumulating; HER2 scoring is the most active development area.

Operational deployment challenges

Routine clinical deployment of AI pathology tools faces several challenges:

• Slide preparation variability. Tissue fixation, embedding, sectioning, and staining vary across labs. AI models trained on one lab's slides may underperform on another lab's slides. Robustness to preparation variation is the key validation requirement.
• Scanner variation. Different scanner manufacturers produce different image characteristics; models must generalize across scanner platforms.
• Whole-slide vs region-of-interest analysis. AI tools that require pathologist selection of regions of interest are limited; tools that operate on whole slides are more useful but face computational and storage challenges.
• Integration with laboratory information systems. AI outputs must flow into the standard pathology reporting workflow rather than living in separate dashboards.
• Pathologist workflow disruption. Even excellent AI tools that disrupt established workflows face adoption resistance. User-interface and workflow integration matter.

Evidence table

Open questions and active investigation

• FDA-cleared standalone AI HER2 scoring. The first FDA-cleared standalone AI scoring tool for HER2 (or for any TNBC-relevant IHC) will set a regulatory precedent and accelerate adoption. Timing depends on vendor validation studies; expected within 2–3 years.
• AI biomarker discovery vs replacement. If H&E-based AI can predict HER2 status or BRCA status as well as the gold-standard test, do we still need the test? Likely yes initially (regulatory and verification reasons) but the answer may evolve.
• Foundation models in pathology. Large pre-trained image foundation models (analogous to ChatGPT in language) are being developed specifically for pathology. Path Foundation, UNI, CONCH, and others are being released. Whether these change the field by enabling rapid task-specific adaptation is being demonstrated.
• Generalizability across populations. AI tools trained on Western cohorts may underperform on non-Western patients (different prep methods, different tumor mix, possibly different morphologic features). Inclusion of diverse training data is a fairness imperative.
• Pathologist workforce implications. AI augmentation could reduce pathologist time per case, allowing more cases per pathologist. The net effect on the workforce is unclear; some tasks (routine IHC scoring) may automate while others (complex diagnostic interpretation) won't.
• Multimodal AI integration. Combining H&E features, molecular markers, clinical features, and imaging features into composite predictors is an active machine-learning direction (see ML multimodal fusion synthesis when available).

For the HER2 IHC scoring problem that AI assistance aims to address, see the IHC synthesis. For TIL biology, see the TME and TILs synthesis. For complementary genomic biomarker information, see the genomic profiling synthesis.

References

Each citation links to the original publication via DOI. The same records are searchable in the evidence library by title or DOI.

1. Fernandez AI, Liu M, Bellizzi A, et al. Examination of Low ERBB2 Protein Expression in Breast Cancer Tissue. JAMA Oncol. 2022;8(4):607–610. doi:10.1001/jamaoncol.2021.7239. ↩
2. Amgad M, Stovgaard ES, Balslev E, et al. Report on computational assessment of tumor infiltrating lymphocytes from the International Immuno-Oncology Biomarker Working Group. NPJ Breast Cancer. 2020;6:16. doi:10.1038/s41523-020-0154-2. ↩
3. Coudray N, Ocampo PS, Sakellaropoulos T, et al. Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. Nat Med. 2018;24(10):1559–1567. doi:10.1038/s41591-018-0177-5. ↩

Last reviewed: 2026-06-04. Researcher-layer synthesis page. Evidence grades follow the GRADE-adapted rubric defined at the top of this page.