Deep-learning assisted photoacoustic histology for real-time intraoperative pathological diagnosis

项目摘要 尽管癌症治疗取得了进展，但手术仍然是基石，80%以上的癌症 在癌症演变的某个阶段，患者会进行外科手术。在肿瘤外科中，术中 病理学提供手术指导和肿瘤边缘的识别，允许确认完整的 在肿瘤外科医生关闭手术伤口之前切除肿瘤，帮助患者避免第二个肿瘤 切除手术。大多数边缘阴性切除的局部肿瘤显示出改善的患者预后和 较低的肿瘤复发几率。然而，术中冰冻切片技术存在一系列问题。 局限性：组织丢失，冷冻伪影导致的质量下降，脂肪标本切割不佳，以及 无法诊断骨质损伤。 在我们的初步结果中，我们发展了三维轮廓扫描紫外光声显微镜(UV- PAM)获取厚骨标本的类似组织学的图像，这解决了长期存在的 术中骨组织学检查。骨标本的快速光声组织学图像与 苏木精伊红(H&E)染色的常规组织学图像，使病理学家能够识别 癌症特征很容易遵循现有的模式识别参数。尽管这些结果表明 术中光声组织学取骨标本的可行性，该系统成像速度相对较慢 速度从根本上受到UV激光重复频率低的限制，并且仅适用于骨骼标本。 该研究计划旨在开发一个高通量的光声组织学平台，供病理学家和 外科医生在手术中进行远程诊断，成像速度至少比任何 发布了反射式UV-PAM系统。 具体目标1：建立用于无载玻片超快组织学成像的结构照明UV-PAM 标本。目标1.1。我们将开发一种使用多焦点照明的超快反射模式UV-PAM 单元件换能器。目标1.2。我们将设计和制造用于结构照明UV-PAM的DO 具有不规则表面的无载玻片样品的扩展焦深，以允许高通量成像 临床环境中的无载玻片标本。 具体目标2：实现用于光声组织学虚拟染色的神经网络和实时 术中诊断。目标2.1。我们将实施神经网络和无监督深度学习 用于对各种组织类型的光声图像进行虚拟染色的技术。虚拟染色PAM的应用 术中诊断的图像将由临床实践中的病理学家进行评估。目标2.2。我们会 开发和训练深度学习神经网络，用于对不同组织中的肿瘤类型和分期进行分类 光声组织学为术中实时诊断搭建了计算机辅助平台。

Project Summary Despite the advances in cancer treatment, surgery remains the cornerstone, and more than 80% of cancer patients have a surgical procedure at some point in their cancer evolution. In oncology surgery, intraoperative pathology provides surgical guidance and identification of tumor margins, allowing confirmation of complete tumor resection before oncology surgeons close the surgical wound and helping patients avoid a second tumor resection surgery. Most localized tumors with negative margin resection show improved patient outcomes and a lower chance of tumor recurrence. However, the intraoperative frozen section technique suffers from a series of limitations: tissue loss, compromised quality due to freezing artifacts, suboptimal cutting of fatty specimens, and inability to diagnose bony lesions. In our preliminary results, we have developed the 3D contour scanning ultraviolet photoacoustic microscopy (UV- PAM) to acquire histology-like images of thick bone specimens, which addresses the long-standing challenge of intraoperative bone histology. The rapid photoacoustic histology images of bone specimens well match the conventional histology images stained by hematoxylin and eosin (H&E), allowing pathologists to identify the cancerous features following existing pattern recognition parameters readily. Although these results showed the feasibility of intraoperative photoacoustic histology of bone specimens, the system has a relatively slow imaging speed fundamentally limited by the low laser repetition rate of UV lasers and applies only to only bone specimens. This research proposal aims to develop a high-throughput photoacoustic histology platform for pathologists and surgeons to diagnose intraoperatively and remotely with an imaging speed at least 100 times faster than any published reflection-mode UV-PAM systems. Specific Aim 1: Develop a structured illumination UV-PAM for ultrafast histology imaging of slide-free specimens. Aim 1.1. We will develop an ultrafast reflection mode UV-PAM using multifocal illumination with a single element transducer. Aim 1.2. We will design and fabricate DOEs for structured illumination UV-PAM with an extended depth of focus for slide-free specimens with irregular surfaces to allow high-throughput imaging of slide-free specimens in clinical settings. Specific Aim 2: Implement neural networks for virtual staining of photoacoustic histology and real-time intraoperative diagnosis. Aim 2.1. We will implement neural networks and unsupervised deep learning techniques to virtually stain photoacoustic images in various tissue types. The utilization of virtual stained PAM images for intraoperative diagnostic will be evaluated by pathologists in clinical practices. Aim 2.2. We will develop and train a deep learning neural network to classify tumor types and stages in different tissues using photoacoustic histology to build a computer-aided platform for real-time intraoperative diagnosis.