Validating Two Pathological Foundation Models for Breast Biomarker Detection

Keywords: vision foundation model · DPO · pathology · LoRA · breast biomarker · PR status

Highlights. First to transfer Direct Preference Optimization from LLM alignment to computational pathology. Designed a pseudo-probability framework — cosine similarity, temperature scaling (γ=10), sigmoid — converting TITAN encoder geometry into DPO preference likelihoods; LoRA (r=16) trains only 2.7% of base parameters with a frozen reference model for implicit KL regularization. Designed a label-free preference-pair pipeline where a baseline classifier flags misclassified samples and UMAP+DBSCAN over correctly classified clusters yields “preferred” embeddings vs. misclassified “rejected” targets. Awarded Honorable Mention at the UChicago Applied Data Science Capstone Showcase.

Recommended citation: Atiya, S., Liang, J., et al. (2025). "Validating Two Pathological Foundation Models for Breast Biomarker Detection." In preparation.