Computational Synthesis of Histological Stains: A Step Toward Virtual Enhanced Digital Pathology

Histological staining plays a crucial role in anatomic pathology for the analysis of biological tissues and the formulation of diagnostic reports. Traditional methods like hematoxylin and eosin (H&E) primarily offer morphological information but lack insight into functional details, such as the expression of biomarkers indicative of cellular activity. To overcome this limitation, we propose a computational approach to synthesize virtual immunohistochemical (IHC) stains from H&E input, transferring imaging features across staining domains. Our approach comprises two stages: (i) a multi-stage registration framework ensuring precise alignment of cellular and subcellular structures between the source H&E and target IHC stains, and (ii) a deep learning-based generative model which incorporates functional attributes from the target IHC stain by learning cell-to-cell mappings from paired training data. We evaluated our approach of virtual restaining H&E slides to simulate IHC staining for phospho-histone H3, on inguinal lymph node and bladder tissues. Blind pathologist assessments and quantitative metrics validated the diagnostic quality of the synthetic slides. Notably, mitotic counts derived from synthetic images exhibited a strong correlation with physical staining. Moreover, global and stain-specific metrics confirmed the high quality of the synthetic IHC images generated by our approach. This methodology represents an important advance in automated functional restaining, achieved through robust registration and a model trained on precisely paired H&E and IHC data to transfer functions cell-by-cell. Our approach forms the basis for multiparameter histology analysis and comprehensive cohort staining using only digitized H&E slides.