Real-time, multiplexed rendering of lymphoma diagnostic panels from unstained tissue sections using virtual staining.

Abstract

115 Background: Lymphoproliferative disorders are complex, heterogeneous diseases with an ever expanding array of subtypes and variants requiring stepwise assay approach to arrive at a precise diagnosis. While extensive characterization of lesional cells and the microenvironment (TME) via immunophenotyping and genotyping remains the gold standard, a full workup may be unachievable due to limited input tissue, turnaround time pressure, or unavailability of particular markers in-house. Even when available, the interpretation of large numbers of markers by pathologists is limited by the ability to mentally spatially align markers by recall-superimposition of single-plex markers, limiting the discovery of patterns relevant for novel treatments strategies. Here we present a deep-learning based virtual staining method which transforms a single unstained tissue section into a panel of perfectly aligned virtual immunohistochemical (IHC) stained images. Methods: 4um unstained sections were cut from 30 formalin-fixed paraffin-embedded blocks. Using a standard commercial slide scanner (Axio Scan.Z1, Zeiss), autofluorescence images were captured from unstained sections and paired with brightfield images captured from the same sections after chemical staining with H&E or various IHC stains of interest. The virtual staining was performed by a deep neural network trained as a conditional GAN using accurately co-registered patches of paired images . The training dataset comprised seven types of lymphoma plus benign lymphoid tissues. Results: We successfully created models for virtual CD3, CD8, CD20 and CD30 IHC stains as well as H&E. Analysis by board certified pathologists confirmed consistent and accurate results between virtual and IHC stained slides in a range of lymphomas and reactive lymphoid tissue. We quantitatively evaluated the co-expression of CD3/CD8 and the spatial distribution of all four markers in eight holdout slides. We demonstrated that our virtual multiplex IHC staining achieves comparable performance for WSI scoring and cellular localization with actual staining across lymphoma types. Conclusions: We present the feasibility of spatially aligned, rapid virtual H&E and IHC staining of four different markers from single, label-free autofluorescence images using deep-learning techniques. To the best of our knowledge, this is the first report of a panel of virtual IHC stains, perfectly aligned and with overlaid H&E structural context, for analysis of neoplastic and reactive lymphoid tissues. Our technique provides an opportunity for a rapid, comprehensive histopathology and immunophenotyping workflow. Multiplex virtual staining greatly expands diagnostic and discovery opportunities: high accuracy TME profiling, seamless downstream computational image analysis, and screening of large biomarker pools in retrospective analysis of clinical trials.