Liu, Jonathan T. C.Serafin, Robert2025-08-012025-08-012025Serafin_washington_0250E_28197.pdfhttps://hdl.handle.net/1773/53699Thesis (Ph.D.)--University of Washington, 2025Slide-based histopathology is the current gold-standard for cancer diagnosis and treatment planning. Conventional histological methods, developed over a century ago, require tissue specimens to be chemically fixed, embedded in wax, and physically sectioned onto glass slides. These slides are stained with chromogenic dyes, most commonly hematoxylin and eosin (H&E), and imaged via brightfield microscopy. Histopathology has many disadvantages, including destructive sectioning of valuable tissue specimens, limited 2D sampling, and in some cases high rates of inter-observer disagreement. In recent years, multiple 3D imaging techniques have emerged to address the limitations of histopathology. One such modality is open-top light-sheet microscopy (OTLS) which enables rapid, non-destructive imaging of clinical specimens (i.e. surgical excisions or biopsies) stained with fluorescent analogs of H&E. However, OTLS datasets are often hundreds of gigabytes or more in size, posing an exceptional informatics challenge, often making the analysis of OTLS datasets without computational assistance tedious and impractical. This dissertation summarizes three computational workflows developed to address some of these challenges: (1) Physics-based virtual H&E staining of 3D pathology datasets. Pathologists are trained to interpret tissue morphologies using a combination of chromogenic staining and brightfield microscopy. However, modern fluorescent microscopy systems, like OTLS, acquire grayscale data. Grayscale images lack the familiar color cues pathologists rely on for diagnosis. This color-space mismatch presents a potential barrier to clinical interpretation of non-destructive 3D pathology modalities. Thus to facilitate pathologist interpretation of OTLS images I developed FalseColor-Python, a physics based virtual staining platform that converts OTLS datasets into the color-spaces of standard H&E histology. (2) Prognostic analysis of 3D nuclear morphologies for prostate cancer risk assessment. 3D pathology datasets present a valuable opportunity to discover novel prognostic biomarkers by providing comprehensive sampling of diagnostically relevant microstructures. A previous study using OTLS demonstrated that glandular morphologies measured in 3D outperformed their 2D counterparts at stratifying prostate cancer (PCa) patients into high- and low- risk categories. The results motivated the development of additional computational pipelines built to analyze other 3D tissue structures to determine their clinical significance. 2D nuclear morphologies are known prognostic biomarkers in multiple cancer species but remained under-explored in 3D. I developed a 3D nuclear segmentation pipeline designed for 3D pathology datasets. This workflow enabled the extraction and quantification of both 2D and 3D nuclear features from a cohort of 46 PCa patients with known clinical outcomes. The results provide additional evidence that 3D pathology provides superior diagnostic information by showing that, much like our analysis of glands, 3D nuclear morphologies were better at stratifying patients into low- or high-risk categories than their 2D counterparts. (3) Automated detection of prostate cancer via generative immunolabeling. Prior 3D pathology studies done by our group demonstrated the potential prognostic value of analyzing PCa microstructures in 3D . These studies used 3D segmentationtechniques to quantify tissue structures contained in the cancer-enriched regions of core-needle prostate biopsies. Identification of these cancer-enriched regions required manual review of OTLS data by pathologists and at the time there was no method to differentiate between cancerous and benign prostate glands which are often spatially intermixed. To address these limitations, I developed an automated cancer detection platform SIGHT: Synthetic Immunolabeling for Generative Heatmaps of Tumor. SIGHT uses 3D image translation models, trained in a fully supervised fashion, to convert H&E-analog OTLS data into multiplexed 3D immunofluorescence datasets that facilitate tumor detection. SIGHT is trained to synthetically label two cytokeratin markers differentially expressed in cancer and benign prostate glands. Using these synthetic labels, I developed an algorithm that generates 3D heatmaps of cancer-enriched regions in prostate tissues. Validation of SIGHT against ground-truth annotations from a panel of genitourinary pathologists demonstrated that this method could detect PCa at a level comparable to expert human annotators. Finally, to demonstrate the prognostic value of SIGHT, I analyzed the 3D features of prostate glands extracted from 75 3D pathology datasets of 3 mm prostate punches. Glandular features were calculated across the entirety of the tissue and the cancer-enriched regions detected by SIGHT. Using known time-to-recurrence outcomes of the 75 patients the results show that the features contained within the cancer enriched regions of each dataset significantly improved downstream risk stratification.application/pdfen-USnoneComputer visionLight sheet microscopyProstate CancerPathologyBiomedical engineeringMechanical engineeringThesis