组织病理学检查是癌症诊断分析的重要标准之一。而虚拟染色是近年来处理组织病理学切片的一种方式。 从病理学染色切片的制备方式来分类，常见的病理学组织切片主要包括新鲜冷冻切片（Fresh Frozen Slide, FF）和福尔马林固定石蜡包埋切片（Formalin Fixed Paraffin Embedded, FFPE）。FF切片质量低但易于制备；而FFPE切片精度较高但制备需要24小时。一种解决方案是基于FF切片进行虚拟染色，生成FFPE切片。 从病理学染色切片的制备原料来分类，常规的病理检查采用的是苏木精-伊红（Hematoxylin and Eosin, H&E）染色剂制备的切片，但其只能观察到组织或细胞形态学上的特征，更深层次的信息只能通过特殊染色类型、免疫组化染色或荧光图像切片获取。将某一块组织易于获得的切片（H&E）转化为其难以获得的染色类型切片（如Masson Trichrome, MT、Periodic Acid Schiff, PAS、Immunohistochemistry, IHC）或免疫荧光图像（Immunoffuorescence, IF）是解决这一问题的方法。 从计算病理学虚拟染色的方法来看，现有方法多是基于生成对抗网络来进行虚拟染色，2021年来，生成对抗网络在自然图像生成领域遭遇到了扩散模型的挑战。不过，扩散模型在非配对、无引导的病理学图像虚拟染色任务中应用较少。 本文的工作内容如下：1. 本文提出了ST-MKSC，一种FF-FFPE虚拟染色网络，它包含一个基于关键一致信息约束的多频域层次约束（MDHC）网络和释放约束损失（RC损失），利用FF切片虚拟染色生成出高质量的FFPE切片。同时还提出了一种多视场的自监督GAN，其中设计了三个辅助任务来优化虚拟染色过程，可以将FF切片转换为高质量FFPE切片，并尝试探究了扩散模型虚拟染色的可能性。 2.本文提出了基于注意力的变焦生成对抗性网络（AV-GAN）用于虚拟染色。将H&E染色切片转换为高质量的特殊染色切片（MT、PAS）。3. 本文提出了一种小波生成框架，该框架将输入图像解耦为频率上的多个分量，并提出了一种校正输出的不正确拟合阻尼机制和一种残差补偿机制，可以将H&E切片虚拟染色为高质量IHC切片或IF图像。 综上，本文提出的方法在多个虚拟染色任务中取得了最佳效果，让医生能够快速便捷地获取多种平时不易获取的染色切片图像，为诊疗提供了极大的便利。
Histopathological examination is one of the important tools for cancer diagnosis and analysis, and virtual staining is a recent way to process tissue pathological slides. According to the preparation method of pathological staining slides, common pathological tissue slides mainly include fresh frozen slides (FF slides) and formalin fixed paraffin embedded slides (FFPE slides). FF slides have low quality but are easy to prepare; FFPE slides have higher accuracy but require 24 hours for preparation. Classifying based on the raw materials used for pathological staining slides, conventional pathological examinations use slides prepared with hematoxylin eosin (H&E) staining agents, but they can only present morphological features of tissues or cells, and deeper information can only be obtained through slides of special staining types, immunohistochemical (IHC) staining slides, or immunofluorescence (IF) images.This essay proposes two methods for virtual staining based on FF slides and generates FFPE slides to address the aforementioned issue. Two methods are proposed to convert easily obtainable slides (H&E slides) of a certain tissue into other staining slides (such as special staining types or IHC slides) or IF images that are difficult to obtain. The above methods significantly improve the accuracy of existing virtual staining tasks. Specifically, the main contributions of this essay are as follows:1. An FF-FFPE slide virtual staining model based on multiple key structural constraints is proposed. It is an FF-FFPE slide virtual staining network that includes a multi-frequency domain hierarchical constraint network based on key consistent information constraints and a release constraint loss, which utilizes FF slides to generate high-quality FFPE slides through virtual staining.2. A multi-field self-supervised generative adversarial network is proposed, in which three auxiliary tasks are designed to optimize the virtual staining process. This model can convert FF slides into high-quality FFPE slides.3. A wavelet generation framework is proposed, which decouples the input image into multiple components at frequency, and includes a damping mechanism for correcting incorrect fitting of the output and a residual compensation mechanism, which can translate H&E slides into high-quality IHC slides or IF images.4. An attention based varifocal generative adversarial network (AV-GAN) for virtual staining is proposed, which converts H&E stained slides into high-quality special stained slides such as Masson Trichrome (MT) staining and Periodic Acid-Schiff (PAS) staining.The methods proposed in this essay achieve the best results in multiple virtual staining tasks, allowing doctors to quickly and conveniently obtain various staining slide images that are not usually easy to obtain, providing great convenience for diagnosis and treatment.
