3D microscopic imaging with computational image analysis to improve cancer assessment

Abstract

Non-destructive, slide-free, and comprehensive pathology of clinical specimens can improve clinical workflow efficiency and diagnostic performance. Recent advances in optical-sectioning microscopy and optical clearing techniques enable volumetric imaging of thick tissue specimens with subcellular resolution. In addition, computational analysis of 3D microscopic image datasets can objectively assist human observers with more efficient and accurate pathological analysis. In this dissertation, we explore two emerging optical-sectioning microscopy techniques, microscopy with ultraviolet surface excitation (MUSE) and open-top light-sheet (OTLS) microscopy, along with various computational methods to enhance, synthesize or analyze the 3D pathological images of tissue specimen, to improve clinical assessment of breast cancer and prostate cancer. For accurate intraoperative lumpectomy margin assessment, we present a fully automated MUSE system that incorporates 3D deconvolution along with a fluorescent analogue of histology stain, facilitating comprehensive pathology of fresh breast specimen surfaces. For rapid diagnosis of prostate cancer, we developed a compact workflow (1Hr2Dx), consisting of fluorescence labeling, tissue clearing, and 3D OTLS microscopy, to diagnose a set of 12 prostate needle cores within an hour of receipt, which can provide patients with a preliminary on-site diagnosis after a biopsy procedure, thereby alleviating anxiety and potentially expediting treatments. To facilitate more reliable cancer risk evaluation, we developed a workflow for non-destructive 3D pathology and computational analysis of whole prostate biopsies that are labeled with a fluorescent analog of H&E staining. The analysis is based upon the development of an annotation-free deep-learning-based volumetric segmentation strategy, namely image-translation-assisted segmentation in 3D (ITAS3D). Based on the glandular features extracted from our 3D gland segmentations, we show that 3D glandular features are superior to 2D features for risk stratification based on patient outcomes.