Case study: HER2 IHC quantification with AI

Context

Accurate subtyping is pivotal in breast cancer management, guiding therapeutic decision making and impacting patient outcomes. HER2 assessment has long underpinned therapy selection, particularly for targeted agents including antibody-drug conjugates (ADCs).

However, emerging evidence suggests that traditional scoring may be too simplistic. Recent studies^1,2 have shown that even patients with extremely low HER2 levels—previously classified as HER2-negative—benefited from treatment with a targeted ADC, experiencing better outcomes than with standard chemotherapy. The result— a paradigm shift for HER2 scoring in breast cancer, highlighting the need to detect and quantify subtle HER2 expression that falls below conventional thresholds yet retains therapeutic significance.

Recently, AI image analysis has helped to better quantify HER2 IHC scoring and reduce inter-rater disagreement³, ultimately making sure patients can benefit from transformative therapies.

Methods

A first model was trained to detect and segment tumor cells from roughly 50,000 annotated cell nuclei. This model uses a lightweight IHC-specific FM (86M params) originally designed for diagnostic purposes and pretrained on 10,000 IHC whole slide images (WSI) covering 100+ markers.

A second model was then used to extract the staining intensity inside the membrane, later converted into “Negative”, “Faint”, “Moderate” and “Strong” classes. Cell-level predictions were compared to more than 7,000 annotations provided by 3 expert pathologists.

For comparison, the same analysis was performed using the Virchow2 state-of-art public FM (632M parameters), pretrained on approximately 300k IHC WSIs.

Results

Our IHC-specific foundation model reached on par performance with current state-of-the-art pathology FMs on the HER2 prediction task while being 8x smaller and trained on 30x less data.

On the validation cohort, our framework distinguished cancer cells vs. non-cancer cells with 89% balanced accuracy (BA), compared to 82% for the Virchow2-based architecture, and achieved a BA of 67% (see Table) in correctly classifying cell-level staining intensities (63% for Virchow2)⁴.

Impact

Our cell-level staining quantification framework achieves state-of-art accuracy while providing quantitative, ADC-relevant HER2 metrics in compliance with ASCO/CAP guidelines. It is readily adaptable to future scoring systems, with native support for spatial feature integration, ensuring robustness and scalability for evolving clinical and research needs.

These results are a first step into demonstrating supercharged IHC scoring and interpretation with quantitative, reliable and interpretable results, to streamline biomarker assessment.

References

_{1. Modi, S. et al. Trastuzumab Deruxtecan in Previously Treated HER2-Low Advanced Breast Cancer. N Engl J Med 2022;387:9-20. DOI: 10.1056/NEJMoa2203690}

_{2. Bardia, A. et al. Trastuzumab Deruxtecan after Endocrine Therapy in Metastatic Breast Cancer. N Engl J Med 2024;391:2110-2122 DOI: 10.1056/NEJMoa2407086}

_{3. Mulder, D. et al. Use of artificial intelligence–assistance software for HER2-low and HER2-ultralow IHC interpretation training to improve diagnostic accuracy of pathologists and expand patients' eligibility for HER2-targeted treatment**.** ASCO 2025 Rapid oral abstract session. 10.1200/JCO.2025.43.16_suppl.1014}

_{4. Accepted peer reviewed abstract, currently under embargo.}