Enhancing Translation of H&E to IHC with Robust Tumor-Infiltrating Lymphocytes Quantification Using Deep Generative Models: A Feasibility Study.

Breast cancer is the leading cause of female oncological mortality, making treatment response prediction a priority. While tumor-infiltrating lymphocytes (TILs) are recognized as prognostic biomarkers, visual quantification on hematoxylin and eosin (H&E) slides suffers from interobserver variability. Virtual H&E-to-IHC stain translation represents a potential alternative to conventional, labor-intensive IHC. In this feasibility study, we developed a deep generative framework to synthesize virtual IHC patches from H&E patches and an automated pipeline for TILs quantification. To facilitate H&E-IHC image alignment, H&E slides were destained and restained for IHC, followed by a coarse-to-fine registration. Our framework incorporates two key components for better synthesis of TIL-specific chromogen: (1) a spatial attention mechanism guided by a cell segmentation map to focus on small TILs and (2) a chromogen loss function based on color separation for accurate diaminobenzidine (DAB) or alkaline phosphatase (AP) synthesis. Furthermore, we propose a TILs quantification pipeline designed to account for cells in overlapping regions. Integrating spatial attention and chromogen loss significantly improved performance, reducing mean absolute error (MAE) from 12.793 to 9.480 for DAB and from 21.113 to 17.755 for AP (p < 0.001 for both) per patch in the best-performing model. Reader studies demonstrated that the automated pipeline achieved moderate to high absolute count agreement with pathologists (intraclass correlation coefficient (ICC), 0.565-0.852) and strong relative density ranking (Spearman, 0.721-0.798) that exceeded inter-reader consistency. Furthermore, visual scoring confirmed that synthesized chromogens were comparable to ground truth IHC (p = 0.640).