Liu, Jonathan T.C.Gao, Gan2026-04-202026-04-202026Gao_washington_0250E_29204.pdfhttps://hdl.handle.net/1773/55528Thesis (Ph.D.)--University of Washington, 2026Slide-based histopathology for biopsies or surgical specimens is the gold standard for guiding treatment decisions. Examples include intraoperative assessment of resected tissue margins based on frozen section analysis and diagnosis and patient prognosis using FFPE slides. However, slide-based microscopic pathology has many limitations, including severe under-sampling, ambiguous 2D visualization of 3D morphologies, slow turnaround times, and destruction of valuable specimens. To address these limitations, open-top light-sheet (OTLS) microscopy has been developed for non-destructive slide-free pathology of clinical specimens. This technology enables rapid imaging of intact tissues at high resolution in 2D or 3D, providing the same level of detail as traditional pathology. To narrow the gaps to clinical adoption, this dissertation focuses on developing computational processing approaches and optimizing OTLS imaging workflows for ex-vivo surgical guidance and non-destructive 3D pathology. Specifically, for the application of surgical guidance, I demonstrate 1) the ability to generate false-colored H&E mimicking images of freshly excised surgical margin surfaces stained for < 1 min with a single fluorophore, 2) rapid OTLS surface imaging at a rate of 15min/cm^2 followed by real-time post-processing of datasets within RAM at a rate of 5min/cm^2, and 3) rapid digital surface extraction to account for topological irregularities. I show that the image quality generated by our rapid surface-histology method in an intraoperative-acceptable timeframe approaches that of gold-standard archival histology. Furthermore, for the application of rendering final diagnostic determinations, the ability to effectively guide pathologists through spatially heterogeneous 3D pathology datasets is critical. To this end, I developed a context-aware deep learning triage approach that automatically identifies the highest-risk 2D slices within 3D volumetric tissue, enabling time-efficient pathologist evaluation. The AI-triaged 3D pathology workflow is implemented for two clinical scenarios, triage of higher-grade prostate cancer as well as screening of esophageal dysplasia/cancer for Barrett's esophagus patients, achieving area-under-curves (AUCs) >0.9 for identifying high-risk slices in both cases. Preliminary clinical validations with panels of Genitourinary and Gastrointestinal pathologists show that AI-triaged 3D pathology improves the identification of higher-grade malignancies compared to standard-of-care 2D slide-based pathology.application/pdfen-USnone3D pathologyLight-sheet microscopySurgical guidanceWeakly supervised learningMedical imagingBiomedical engineeringArtificial intelligenceMechanical engineeringThesis