Non-invasive automated wound analysis via deep learning neural networks

项目概要： 每年有数百万美国人发展为慢性伤口，这需要先进的伤口护理， 估计每年花费500亿美元。然而，我们对慢性伤口的理解以及如何治疗它们， 由于缺乏已建立的方法来客观地表征和测量伤口特征而受到限制。 在临床和研究实验室中，伤口的详细评估通常通过组织学分析进行 组织活检。这些信息可以提供深入了解细胞迁移到伤口，细胞增殖 感染和纤维化然而，组织学的收集、创造和分析 切片本身具有侵入性、耗时性和定性性。这个提案的目的是为了塑造一个形象 分析管道，可以提供伤口的自动化定量分析，并为非创伤分析奠定基础。 可以提供与标准组织病理学相同信息的侵入性实时“光学活检”。我们的中央 假设使用深度学习卷积神经网络的人工智能方法可以 结合体内多光子显微术和现有的定量图像分析方法来实现这一点 与传统的活检组织学染色和专家分析相同的准确性。目标1： 将训练和验证能够分割和量化标准伤口组织学的神经网络， 接受三个独立的伤口愈合研究实验室的培训在目标2中，我们将调整该网络以执行 皮肤伤口的体内无标记多光子显微镜图像的分割和量化， 伤口组织和代谢功能的快速读数。最后，在目标3中，我们将开发和验证一个 能够从我们的无染色非侵入性体内MPM图像生成虚拟组织学图像的网络， 这些网络可以与目标1和目标2中开发的网络相结合， 伤口微观结构和代谢。在短期内，这一建议将制定一系列强有力的分析， 可应用于现有的H& E染色或未染色的皮肤组织切片的工具， 疗愈研究者从长远来看，无标记多光子显微镜和机器的结合 基于学习的图像分析将实现完全非侵入性的伤口组织学， 在护理点实时指导清创和伤口护理。

Project Summary: Each year millions of Americans develop chronic wounds, which require advanced wound care that has been estimated to cost $50 Billion annually. However, our understanding of chronic wounds and how to treat them has been limited by a lack of established methods to objectively characterize and measure wound features. Detailed assessments of wounds in the clinic and research laboratory often occur through histological analysis of tissue biopsies. This information can provide insight into cellular migration into the wound, cellular proliferation at the edge of the wound, infection, and fibrosis. However, the collection, creation, and analysis of histology sections is inherently invasive, time-consuming, and qualitative. The goal of this proposal is to develop an image analysis pipeline that can provide automated quantitative analysis of wounds and lay the groundwork for a non- invasive real-time “optical biopsy” that can provide information identical to standard histopathology. Our central hypothesis is that artificial intelligence approaches using deep learning convolutional neural networks can be coupled with in vivo multiphoton microscopy and existing quantitative image analysis methods to achieve this goal with the same accuracy as traditional biopsies with histological staining and expert analysis. In Aim 1, we will training and validate neural networks capable of segmenting and quantifying standard wound histology based on training from three independent wound healing research labs. In Aim 2, we will adapt this network to perform segmentation and quantification of in vivo label-free multiphoton microscopy images of skin wounds to provide rapid readouts of wound organization and metabolic function. Finally in Aim 3, we will develop and validate a network capable of generating virtual histology images from our stain-free non-invasive in vivo MPM images, which can be coupled with the networks developed in Aim 1 and 2 to provide a comprehensive assessment of wound microstructure and metabolism. In the near-term, this proposal will develop a series of robust analysis tools that can be applied to existing H&E-stained or unstained skin tissue sections commonly studied by wound healing researchers. In the long-term, the combination of label-free multiphoton microscopy and machine learning-based image analysis will enable completely non-invasive wound histology that can be performed in real-time at the point of care to guide debridement and wound care.