Assessment of AI-based computational H&E staining versus chemical H&E staining for primary diagnosis in lymphomas: a brief interim report.

Abstract

Microscopic review of tissue sections is of foundational importance in pathology, yet the traditional chemistry-based histology laboratory methods are labour intensive, tissue destructive, poorly scalable to the evolving needs of precision medicine and cause delays in patient diagnosis and treatment. Recent AI-based techniques offer promise in upending histology workflow; one such method developed by PictorLabs can generate near-instantaneous diagnostic images via a machine learning algorithm. Here, we demonstrate the utility of virtual staining in a blinded, wash-out controlled study of 16 cases of lymph node excisional biopsies, including a spectrum of diagnoses from reactive to lymphoma and compare the diagnostic performance of virtual and chemical H&Es across a range of stain quality, image quality, morphometric assessment and diagnostic interpretation parameters as well as proposed follow-up immunostains. Our results show non-inferior performance of virtual H&E stains across all parameters, including an improved stain quality pass rate (92% vs 79% for virtual vs chemical stains, respectively) and an equivalent rate of binary diagnostic concordance (90% vs 92%). More detailed adjudicated reviews of differential diagnoses and proposed IHC panels showed no major discordances. Virtual H&Es appear fit for purpose and non-inferior to chemical H&Es in diagnostic assessment of clinical lymph node samples, in a limited pilot study.