Deep learning microscope for slide-free and digital histology

点击翻译按钮获取中文摘要

Project summary/abstract: Anatomic histopathology plays a central role in disease diagnosis and in surgical procedure guidance to ensure delivery of quality healthcare and treatment. At the time of surgery, for example, tumor margins are ideally assessed with fast frozen section pathology to help ensure complete tumor resection while sparing normal tissue. Unfortunately, the time- and labor-intensive slide preparation process requires expensive equipment and specialized personnel, so it is not widely available in many settings including the rural US; even in settings with the clinical infrastructure to perform frozen section, only a small fraction of the margin is manually examined. In resource-limited global settings, a dire shortage of pathologists makes it more challenging to provide routine diagnostic pathology. Therefore, there is a critical need for affordable tools to support quality histopathology programs throughout the world. The goal of this proposal is to use recent advances in optical fabrication and artificial intelligence to develop a new and affordable tool, the deep learning extended depth-of-field (DeepDOF) platform, to rapidly examine fresh tissue resections without extensive slide preparation, while providing computer-aided image analysis at the point of care. We will demonstrate and validate its use for tumor margin assessment in patients with oral squamous cell carcinoma, the sixth most common malignancy worldwide. In Aim 1, we will develop key modules of the DeepDOF platform for rapid, subcellular imaging of freshly resected tissue samples. A deep learning network will be developed to design and optimize the DeepDOF microscope to image highly irregular tissue surfaces (up to 200 µm) at subcellular resolution without mechanical refocusing; we will combine it with fast vital dyes and deep ultraviolet illumination to achieve high contrast imaging. In Aim 2, we will carry out a clinical evaluation of DeepDOF to determine its ability to assess oral tumor margin status immediately following surgery. The clinical workflow of DeepDOF for intraoperative oral tumor margin assessment will be optimized, and its performance will be evaluated by comparing to gold standard histopathology. In Aim 3, we will develop a machine learning framework to identify positive margins in and assist annotation of large-area, cellular-resolution DeepDOF maps of oral surgical specimens. Using clinical data acquired in Aims 1 and 2, we will train an algorithm to complete segmentation tasks for identifying key diagnostic features such as nuclear enlargement and abnormal clustering; the results will be further used to annotate and quantify positive margins at the point of care. Taken together, we will develop a first microscopy platform with AI-driven optics and algorithms for rapid and slide-free histology of intact tissue samples, and we will provide important clinical evidence to show the DeepDOF platform can improve patient care during oral cancer surgeries. Equipped with a computer-aided image analysis, the broader impact of the DeepDOF platform extends to global settings including low- and middle-income countries that lack access to high quality histopathology services.