Benchmarking Open-Source Pathology Foundation Models for Breast Cancer Biomarker Prediction from H&E Whole-Slide...

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology image analysis

Background:

• Immunohistochemistry (IHC) for breast cancer biomarker detection faces challenges like long turnaround times and lab variability.
• Open-source vision-language foundation models offer a potential solution for biomarker inference from H&E-stained whole-slide images (WSIs).

Purpose of the Study:

• To evaluate the performance of two open-source pathology foundation models, TITAN and CHIEF, in predicting ER, PR, and HER2 status from H&E-stained breast cancer WSIs.
• To assess the potential of AI models for biomarker detection in digital pathology.

Main Methods:

• Two models, TITAN (48.5M parameters) and CHIEF (1.2M parameters), were evaluated on H&E-stained breast cancer WSIs from TCGA-BRCA.
• Biomarker status (ER, PR, HER2) was predicted and compared against ground truth labels.
• Performance was assessed using AUROC, AUPRC, sensitivity, specificity, and PPV across 10 independent random partitions.

Main Results:

• Both TITAN and CHIEF demonstrated comparable and strong performance for ER and PR prediction (ER AUROC: 0.877-0.885, PR AUROC: 0.791-0.799).
• HER2 prediction showed moderate performance (AUROC: 0.71-0.74) but very low sensitivity at the default threshold.
• Positive predictive values (PPV) were high for ER (0.90) and moderate for PR (0.79-0.81) at the default operating point.

Conclusions:

• The evaluated AI models show significant potential for ER and PR status prediction in breast cancer digital pathology.
• HER2 prediction requires further optimization, including threshold calibration and multimodal data integration, before clinical application.
• Prospective external validation is crucial before widespread clinical deployment of these AI models.