Deep learning based label-free virtual staining and classification of human tissues using digital slide scanner.

Hematoxylin and eosin (H&E) staining has long been a cornerstone of histopathology, typically applied to formalin-fixed, paraffin-embedded (FFPE) tissue sections to ensure stable and consistent sample preparation. However, the chemical staining procedure is inherently irreversible, making stained tissue unsuitable for subsequent analysis. Moreover, it is time-consuming, requires chemical handling, and incurs significant costs. While label-free microscopy, combined with deep learning, has been explored as an alternative, its clinical adoption remains limited due to slow imaging speeds, single-slide processing constraints, and operator-dependent variability in system management. To address these challenges, we propose a high-throughput virtual histology framework that integrates deep learning-based virtual staining and classification with a high-throughput digital multi-slide scanner. This approach enables the rapid and automated processing of unstained bright-field (UBF) images acquired from formalin-fixed, paraffin-embedded (FFPE) tissue sections, including those from lymph nodes, brain, and liver samples. We employ instance-wise hard negative example generation for Contrastive Learning in the Unpaired Image-to-Image Translation (NEGCUT) model to generate virtual H&E (VHE) images that faithfully replicate conventional staining while preserving critical morphological features. Furthermore, we introduce a residual network (ResNet)-based classification model that jointly leverages both VHE and UBF images, achieving 95.9 % accuracy in cancer detection. In a blind evaluation, board-certified pathologists confirmed that VHE images were diagnostic and indistinguishable from conventionally stained slides. By leveraging widely available digital slide scanners, our framework offers a scalable and cost-effective alternative to conventional staining, enabling high-throughput histopathology with reduced time, labor, and chemical usage.