受激拉曼散射和CycleGAN结合的胰腺穿刺组织虚拟染色仿真

Stimulated Raman Scattering and CycleGAN-Based Virtual Staining Simulation of Pancreatic Puncture Tissue

摘要: 传统的胰腺病理诊断依赖于福尔马林固定石蜡包埋(FFPE)染色技术，尽管其具有较高的诊断价值，但过程繁琐且对组织样本有一定的损伤。为了克服这一局限，本研究提出了一种基于受激拉曼散射(SRS)显微镜和CycleGAN模型的虚拟染色方法，用于超声内镜下细针穿刺活检(EUS-FNA)组织的福尔马林固定石蜡包埋染色仿真。通过利用SRS成像技术获取胰腺组织的高分辨率细胞图像，并应用CycleGAN模型进行弱监督学习。实验结果表明，虚拟染色图像在组织结构和病理特征上与传统FFPE染色图像高度相似，且保持了较高的细节信息，验证了该方法在胰腺组织快速诊断中的潜力。与传统染色方法相比，该技术具有无创、高效、节省时间和成本的优势。此研究为基于深度学习的虚拟染色方法在胰腺病理诊断中的应用提供了新的思路，也对其它生物组织的类似研究具有重要价值。

Abstract: Traditional pancreatic pathological diagnosis relies on formalin-fixed paraffin-embedded (FFPE) staining techniques, which, despite their high diagnostic value, are cumbersome and cause some damage to tissue samples. To overcome these limitations, this study proposes a virtual staining method based on stimulated Raman scattering (SRS) microscopy and CycleGAN models for simulating FFPE staining of endoscopic ultrasound-guided fine needle aspiration (EUS-FNA) tissue samples. By utilizing SRS imaging to acquire high-resolution cellular images of pancreatic tissue, and applying CycleGAN models for Weakly supervised learning, we successfully generate virtual staining images that closely resemble traditional FFPE-stained images. Experimental results show that the virtual staining images are highly similar to the FFPE images in terms of tissue structure and pathological features, while preserving fine detail. This validates the potential of this method for rapid pancreatic tissue diagnosis. Compared to traditional staining methods, this technology offers the advantages of being non-invasive, efficient, and cost- and time-saving. This study provides a new approach for the application of deep learning-based virtual staining in pancreatic pathological diagnosis, and holds significant value for similar research on other biological tissues.