Tweel, James2026-03-112026-03-112026-03-112026-03-06https://hdl.handle.net/10012/22968Histological staining remains the gold standard of diagnostic pathology, enabling visualization of tissue structure and cellular morphology. However, traditional staining workflows are time-consuming, destructive, and chemically intensive, limiting the number of stains that can be applied to valuable biopsy samples. These processes also introduce delays, variability in stain quality, and high resource demands. To address these limitations, this thesis presents a label-free histology framework that combines Photon Absorption Remote Sensing (PARS) microscopy with deep learning–based virtual staining to replicate commonly used histochemical stains without altering or consuming the tissue. The first component of this work focuses on the development of an automated whole slide PARS system designed for imaging thin, transmissible tissue sections. The system captures sub-micron resolution radiative and non-radiative absorption contrasts using 266 nm UV excitation, targeting endogenous chromophores such as DNA and extracellular matrix components to reveal nuclear and connective tissue structures. Whole slide imaging is achieved through automated focusing, tiling, and contrast leveling, producing gigapixel-scale images directly comparable to standard hematoxylin and eosin (H&E) slides. The second component introduces a deep learning virtual staining pipeline based on the unpaired CycleGAN architecture, with direct comparison to the paired Pix2Pix model. These models are trained on one-to-one whole slide images of PARS data and chemically stained H&E slides. The first masked clinical concordance study is conducted using breast needle core biopsies, where board-certified pathologists independently diagnose and assess the virtual and real H&E slides. The study demonstrates substantial diagnostic agreement, validating the clinical viability of the PARS-based virtual staining approach. The final component expands the PARS imaging system through the integration of a secondary long-wave UV excitation wavelength (355 nm), enabling sensitivity to additional biomolecular absorbers and thereby expanding the captured label-free contrasts. The additional label-free contrast contributes to improved emulation of histochemical stains beyond H&E, including Masson’s Trichrome, Periodic acid–Schiff, and Jones methenamine silver. To further improve performance, a more advanced registration-guided GAN model (RegGAN) is adopted, outperforming both Pix2Pix and CycleGAN. The resulting whole slide virtual images closely match their ground truth counterparts in qualitative appearance, quantitative metrics, and masked pathology review. Together, this work presents a non destructive histology pipeline capable of generating high-resolution, multi-stain images of commonly used stains without chemical labeling, representing a step toward integrating label-free microscopy and deep learning virtual staining into routine pathology workflows.enPhoton Absorption Remote Sensing (PARS) microscopyVirtual stainingLabel-free optical imagingWhole slide imagingDeep learningHistochemical stain emulationAbsorption microscopyMulti-wavelength excitationDoctoral Thesis