OCT as a Platform for Non-Invasive Virtual H&E Biopsy

该项目的主要目标是开发成像仪器和算法技术， 进行非侵入性、实时、体内、3D、虚拟H&E活检。全世界每四个人中就有一个 最终会受到癌症的影响。手术切除是大多数实体癌的主要治疗方法。外科医生 其任务是切除足够的组织以避免留下残留的癌细胞， 同时又不会切除太多的组织，因为这会损害器官功能。这对大脑尤其重要。 肿瘤是儿童中最常见的实体瘤类型，也是儿童癌症死亡率的主要原因。的 检测大多数实体癌和确认肿瘤边缘的金标准是苏木素和伊红（H&E） 染色的组织切片，这需要侵入性活检程序。不幸的是，目前的非侵入性体内 成像模态不能产生具有可比有用性的图像。我们提出了一种新的成像方式 一种被称为“虚拟H&E活检”的技术，可以在真实的时间内生成类似H& E的活体组织图像。非侵入性地向上 至组织中1 mm。这种成像方式将能够提供肿瘤边缘的实时诊断， 通过扫描大的组织区域以寻找残留的癌细胞来提高侵袭性。这些信息将指导治疗 例如脑癌和皮肤癌。除了临床益处外，这项技术还可以 用于研究肿瘤发展和肿瘤对治疗的反应，通过提供体内H& E样 随时间变化的健康和肿瘤组织微观结构的图像。 为了生成虚拟H&E图像，我们将优化我们基于 光学相干断层扫描（OCT）和图像翻译的生成对抗神经网络 （甘）。使我们能够训练GAN生成虚拟H&E图像的关键突破是一种称为 光学条形码，我们使用它来获得OCT图像和相应的真实的H&E图像的数据集 以单细胞精度对齐。我们已经用离体人体皮肤组织演示了这种虚拟H&E系统 样品在建议的项目中，我们将首先训练GAN生成健康小鼠的虚拟H&E图像 脑组织和胶质母细胞瘤小鼠离体脑组织（Aim 1）。第二，我们将利用迁移学习进行再培训 GAN生成体内OCT扫描的小鼠脑组织的虚拟H&E图像（Aim 2a），并跟踪 第一次H&E图像如何随着小鼠胶质母细胞瘤肿瘤的发展而变化（目的2b）。最后，我们将评估 GAN是否可以通过在小鼠大脑训练的GAN上应用迁移学习来跨物种重新训练 并使用它来生成离体低级别人类胶质瘤的虚拟H&E活检（Aim 3）。据我们所 这将是迁移学习首次在生物医学图像中跨物种应用。 这种迁移学习可以加速虚拟活检研究，因为小鼠样本更容易获得。 获取和处理，从而在获取人类数据集用于训练 虚拟活检GAN将难以或不可能实现（例如，视网膜）。

The broad objective of this project is to develop imaging instrumentation and algorithmic technology to perform non-invasive, real-time, in-vivo, 3D, virtual H&E biopsies. One in four people worldwide will ultimately be affected by cancer. Surgical removal is the main treatment for most solid cancers. The surgeon is tasked with the delicate balancing act of excising enough tissue to avoid leaving behind residual cancer cells while not removing too much tissue, which can harm organ function. This is particularly important for brain tumors, the most common type of solid tumor in children and the leading cause of pediatric cancer mortality. The gold standard for detecting most solid cancers and confirming tumor margins is hematoxylin and eosin (H&E) stained tissue sections, which require an invasive biopsy procedure. Unfortunately, current non-invasive in-vivo imaging modalities do not produce images of comparable usefulness. We propose a novel imaging modality called a "virtual H&E biopsy'' that would generate H&E-like images of living tissue in real time. non-invasively up to 1 mm into the tissue. This imaging modality would be able to provide real-time diagnosis of tumor margins and invasiveness by scanning a large tissue area for residual cancer cells. Such information would guide treatment decisions for diseases such as brain and skin cancer. Beyond its clinical benefits, this technology can also be used for research into tumor development and tumor responses to treatment by providing in-vivo H&E-like images of healthy and tumorous tissue microstructures changing over time. To generate virtual H&E images, we will optimize a new imaging instrument we have developed based on optical coherence tomography (OCT) and image translation by a generative adversarial neural network (GAN). The key breakthrough enabling us to train a GAN to generate virtual H&E images is a technique called optical barcoding, which we used to obtain a dataset of OCT images and corresponding real H&E images aligned to single-cell precision. We have demonstrated this virtual H&E system with ex-vivo human skin tissue samples. For the proposed project, we will first train a GAN to generate virtual H&E images of healthy mouse brain tissue and glioblastoma mouse brain tissue ex-vivo (Aim 1 ). Second, we will use transfer learning to retrain the GAN to generate virtual H&E images of mouse brain tissue of in-vivo OCT scan (Aim 2a), and track for the first time how H&E images change as a mouse glioblastoma tumor develops (Aim 2b}. Finally, we will assess whether the GAN can be retrained across species by applying transfer learning on the mouse-brain trained GAN and use it to generate a virtual H&E biopsy of ex-vivo low-grade human glioma (Aim 3). To the best of our knowledge, this will be the first time transfer learning has been applied across species for biomedical images. Such transfer learning can accelerate virtual biopsy research since mouse samples are significantly easier to obtain and handle, thereby opening up applications in locations where acquiring a human dataset for training a virtual biopsy GAN would be difficult or impossible to achieve (e.g., the retina).