When
Title:
Abstract:
Label-free multiphoton microscopy (MPM) provides endogenous biochemical and structural contrast without exogenous staining, offering a promising platform for rapid and quantitative tissue assessment. This dissertation develops deep learning methods to translate label-free MPM images into diagnostically and biologically meaningful predictions, using pancreatic neuroendocrine tumors as a primary disease model. Three complementary studies are presented. First, transfer learning enables accurate binary classification of tumor and normal pancreatic tissue, with held-out test AUCs above 0.95. Second, MPM is integrated with spatial transcriptomics to establish an optical phenotyping framework capable of predicting six molecularly defined tissue states from autofluorescence and second-harmonic generation images. Third, an unpaired animal-to-human image translation approach introduces biologically informed training data to address the limited availability of human specimens and improves classification on held-out human images. Together, these studies demonstrate how molecularly grounded labels, hierarchical image features, and cross-species domain translation can advance label-free microscopy from descriptive imaging toward predictive, data-efficient computational pathology.
Committee:
Dr. Travis Sawyer (Chair)
Dr. DK Kang
Dr. Ron Liang