PO.BCS01.07 · 生物信息与计算
3D全息断层成像实现厚层肾脏内科组织的虚拟多重染色用于全面肾脏病理学
3D holotomography-enabled virtual multiplexed staining of thick medical kidney tissue for comprehensive renal pathology
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
肾脏病理学依赖多种特殊染色——包括H&E、PAS、AFOG和PAM——来评估肾小球、肾小管和间质的改变。这些染色需要连续切片，易产生变异，并涉及破坏性处理，限制了一致性并妨碍了真正的3D解读。全息断层成像(HT)能够对厚层组织进行无标记的3D折射率(RI)成像，生成式转换模型可以直接从RI输入合成染色等效对比。我们开发了一个虚拟多重染色框架，从单一无标记输入生成四种主要肾脏染色，包括在常规染色变得不可靠的厚切片中。用3D HT对肾脏内科样本(薄切片和厚达50 μm的厚切片)进行成像以获得RI体积。在配准和质控后，在配对的RI-染色瓦片(H&E、PAS、AFOG、PAM)上训练了一个条件生成模型。在留出区域上计算相似性指标——SSIM、PSNR、LPIPS、PCC。该模型应用于(i)无标记RI和(ii)化学染色切片以产生跨染色预测，从而实现直接比较。在0.5-mm肾脏区域上进行了全切片虚拟染色。该框架在各染色中再现了诊断特征，结构相似性通常>0.85。虚拟染色恢复了关键要素——包括基底膜、系膜基质、肾小管上皮和间质胶原——与化学参照高度匹配。大区域预测保留了连贯的结构。虚拟PAS和AFOG突出了基底膜、富含糖原的区域、胶原沉积和管型，支持对肾小球损伤和间质纤维化的解读。在厚层组织(10-50 μm)中保持了稳健的性能，而化学染色往往在这些组织中显示出不均匀的渗透；虚拟染色保持了均匀的对比并勾勒出重叠结构。多重染色实现了在同一切片内对基底膜增厚、系膜扩张和炎症的一致可视化，消除了连续切片伪影。全息断层成像与生成式转换相结合，能够实现可靠的无标记虚拟多重染色，从单一未染色或单染色输入产生H&E、PAS、AFOG和PAM等效图像。该方法保留了肾脏显微解剖结构，并在常规染色可能失败的厚切片中保持稳健。通过消除连续切片和重复的化学处理，它提供了一个统一、非破坏性的工作流，改善了可重现性并实现了直接的跨染色比较。这使虚拟多重染色成为一种可扩展的、可用于病理学的技术，适用于肾脏疾病评估、转化研究和计算病理学流程。
查看英文原文 English abstract
Renal pathology relies on multiple special stains-including H&E, PAS, AFOG, and PAM-to assess glomerular, tubular, and interstitial changes. These stains require serial sections, are prone to variability, and involve destructive processing, limiting consistency and preventing true 3D interpretation. Holotomography (HT) enables label-free 3D refractive index (RI) imaging of thick tissues, and generative translation models can synthesize stain-equivalent contrast directly from RI inputs. We developed a virtual multiplexed staining framework that generates four major renal stains from a single label-free input, including for thick sections where conventional staining becomes unreliable.Medical kidney samples (thin and thick sections up to 50 μm) were imaged with 3D HT to obtain RI volumes. A conditional generative model was trained on paired RI-stain patches (H&E, PAS, AFOG, PAM) after registration and QC. Similarity metrics-SSIM, PSNR, LPIPS, PCC-were computed on held-out regions. The model was applied to (i) label-free RI and (ii) chemically stained slides to produce cross-stain predictions, enabling direct comparison. Whole-slide virtual staining was performed on 0.5-mm kidney regions.The framework reproduced diagnostic features across stains, with structural similarity typically >0.85. Virtual stains restored key elements-including basement membranes, mesangial matrix, tubular epithelium, and interstitial collagen-closely matching chemical references. Large-region predictions preserved coherent architecture. Virtual PAS and AFOG highlighted basement membranes, glycogen-rich areas, collagen deposition, and casts, supporting interpretation of glomerular injury and interstitial fibrosis. Robust performance was maintained in thick tissues (10-50 μm), where chemical stains often show uneven penetration; virtual stains preserved uniform contrast and delineated overlapping structures. Multiplexing enabled consistent visualization of basement membrane thickening, mesangial expansion, and inflammation within the same section, removing serial-section artifacts.Holotomography with generative translation enables reliable, label-free virtual multiplexed staining, producing H&E, PAS, AFOG, and PAM-equivalent images from a single unstained or singly stained input. The method preserves renal microanatomy and remains robust in thick sections where conventional staining can fail. By eliminating serial sectioning and repeated chemical processing, it provides a unified, non-destructive workflow that improves reproducibility and enables direct cross-stain comparison. This positions virtual multiplexed staining as a scalable, pathology-ready technology for renal disease assessment, translational research, and computational pathology pipelines.
利益披露 Disclosure
S. Jeong, None..
J. Park, None.