Reconstructing High-Definition Infrared Spectroscopic Images Using Adaptive Sampling and Deep Learning
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Microscopic analysis of tissue is the current standard for making clinical diagnostic and prognostic decisions. Histology requires the use of chemical stains and dyes to provide contrast in bright-field imaging systems. Standard histological labels include hematoxylin and eosin (H&E), Masson's trichrome, and a wide range of immunohistochemical stains targeting proteins. Histological image analysis relies on the quantification of various labor-intensive methods, including cell counting, cell localization, and the measurement of tissue microstructures. Improving the performance of clinical histology requires overcoming two significant barriers: (1) automated tissue segmentation and (2) quantification of molecular composition. While various machine-learning approaches attempt to improve image segmentation, these methods are confounded by deviations between image quality and labeling protocols. One potential solution to both problems is spectroscopic imaging, which provides a quantitative image of the tissue sample, greater molecular detail, and a more robust foundation for segmentation. This dissertation proposes and evaluates a framework for performing label-free histological analysis through three major contributions. First, I develop deep learning architectures that dramatically improve the accuracy of histological segmentation. I then leverage similar architectures to synthesize label-free infrared images to corresponding high-resolution bright-field alternatives for histological interpretation. Finally, I develop an adaptive sampling technique with the potential to provide fast sub-cellular imaging using an emerging photothermal infrared imaging technology.