Document Type : Original Research

Authors

• Mahmoud Bagheri ^{1, 2}
• Alireza Ghanadan ³
• Mobin Saboohi ¹
• Maryam Daneshpazhooh ³
• Fatemeh Atyabi ⁴
• Marjaneh Hejazi ^{1, 2}

¹ Department of Medical Physics and Biomedical Engineering, School of Medicine, Tehran University of Medical Sciences, Tehran, Iran

² Research Center for Molecular and Cellular Imaging, Bio-Optical Imaging Group, Tehran University of Medical Sciences, Tehran, Iran

³ Department of Dermatology, Razi Hospital, Tehran University of Medical Sciences, Tehran, Iran

⁴ Department of Pharmaceutical Nanotechnology, Faculty of Pharmacy, Tehran University of Medical Sciences, Tehran, Iran

Abstract

Background: The use of Hematoxylin-and-Eosin (H&E) staining is widely accepted as the most reliable method for diagnosing pathological tissues. However, the conventional H&E staining process for tissue sections is time-consuming and requires significant labor. In contrast, Confocal Microscopy (CM) enables quick and high-resolution imaging with minimal tissue preparation by fluorescence detection. However, it seems harder to interpret images from CM than H&E-stained images.
Objective: This study aimed to modify an unsupervised deep-learning model to generate H&E-like images from CM images.
Material and Methods: This analytical study evaluated the efficacy of CM and virtual H&E staining for skin tumor sections related to Basal Cell Carcinoma (BCC). The acridine orange staining, combined with virtual staining techniques, was used to simulate H&E dyes; accordingly, an unsupervised CycleGAN framework, trained to virtually stain CM images was implemented. The training process incorporated adversarial and cycle consistency losses to ensure a precise mapping between CM and H&E images without compromising image content. The quality of the generated images was assessed by comparing them to the original images.
Results: The CM images, specifically focusing on subtyping BCC and evaluating skin tissue characteristics, were qualitatively assessed. The H&E-like images generated from CM using the CycleGAN model exhibited both qualitative and quantitative similarities to real H&E images. 
Conclusion: The integration of CM with deep learning-based virtual staining provides advantages for diagnostic applications by streamlining laboratory staining procedures.

Keywords

• Basal Cell Carcinoma
• Microscopy
• Confocal
• Pathology
• Deep Learning