Dual-Modal Autofluorescence Microscopy for CycleGAN-Based H&E-like Image Generation in Liver Tissue: A Preliminary...

Background/Objective: Hematoxylin and eosin (H&E) staining remains the routine reference in histopathology because it provides essential structural information for tissue evaluation. However, conventional H&E staining involves multiple reagent-based and operator-dependent procedures that may introduce variability in staining appearance. This study aimed to develop and preliminarily evaluate a technical proof-of-concept framework for generating H&E-like images from dual-modal autofluorescence microscopy data using a CycleGAN-based virtual staining approach. Methods: Human liver tissue sections were imaged using dual-modal autofluorescence, including DAPI-based nuclear fluorescence and endogenous tissue autofluorescence. A CycleGAN-based virtual staining framework with structural and perceptual constraints was developed to generate H&E-like images, followed by stain normalization to improve color consistency. The generated images were preliminarily evaluated by means of visual comparison, image-feature analysis, feature-space visualization, and nuclear counting. Results: The proposed framework generated H&E-like images with encouraging overall visual resemblance to conventional H&E images. Preliminary image-feature analysis suggested partial similarity between virtual and real H&E images, while feature-space visualization indicated that detectable differences remained. Nuclear counting on 168 images showed broadly consistent nuclear distribution between dual-modal autofluorescence and virtual H&E images, with minor discrepancies mainly related to thresholding artifacts. Some fine nuclear and chromatin-level details remained insufficiently reproduced in the current virtual H&E images. Conclusions: This study presents preliminary feasibility evidence for a technical proof-of-concept framework that translates dual-modal autofluorescence images into H&E-like images. In this small liver-tissue dataset, the generated images demonstrated encouraging overall H&E-like appearance and approximate nuclear localization. Further studies with larger and more diverse datasets, external validation, expert pathological assessment, diagnostic concordance analysis, and systematic workflow evaluation are warranted to assess robustness, diagnostic relevance, and potential utility in digital pathology workflows.