Advancing biophysical cytometry with deep-learning (ABCD) for new-generation cell-based diagnostics

Abstract

In recent years, the field of imaging cytometry has experienced a significant transformation, propelled by innovative breakthroughs in optical microscopy and deep learning techniques. This transformation has revolutionized the role of imaging from a qualitative visual examination tool to quantitative analytic technology, particularly in the realm of label-free imaging. Thanks to its high-throughput and high-content characteristics, label-free imaging cytometry emerged as a powerful technique for conducting large-scale cell-based bioassays in diverse areas, ranging from drug discovery to basic biomedical and biological research. Nonetheless, long-standing challenges remain not only in bridging the gap between qualitative image assessments and robust quantitative measurements, but also in achieving unbiased robust cellular biophysical analysis. These persistent issues constraint the potential and the widespread adoptions of label-free imaging cytometry in biological research and clinical diagnosis.

This thesis aims to address these challenges through utilizing multiplexed Asymmetric Time-stretched Optical Microscopy (multi-ATOM), a high-speed microscopy technique that enables cellular imaging of high-throughput and high-resolution. It enables simultaneous readouts of single-cell brightfield images and quantitative phase images (QPI) at subcellular resolution, which critically empowers large-scale and high-content label-free biophysical phenotyping at single-cell precision. Using multi-ATOM, this work established the first open-access, large-scale label-free image data repository, which served as a foundational resource for comprehensive data-driven cellular biophysical profiling. This approach transforms label-free imaging cytometry into a robust quantitative analytics tool which was once inconceivable.

Furthermore, we present a novel generative deep learning algorithm termed Cyto-Morphology-Adversarial-Distillation (CytoMAD). CytoMAD mitigates the inherent biases in human-mediated cellular phenotyping through automating the process with data-driven deep learning algorithms. Also equipped with the image contrast translation capabilities, CytoMAD further augments and thus enriches the morphological information of cells. Notably, CytoMAD is, to the best of our knowledge, the first deep-learning based model tailored for tackling the notorious batch effect in imaging cytometry. Its batch-aware characteristic, rooted in weakly supervised learning, distills the underlying biological information from the technical confounders/distortions, significantly improving data reproducibility and allowing robust integrative analysis.

The effectiveness of CytoMAD has been rigorously validated using multiple datasets, including laboratory-cultured cell lines, and clinical biopsy samples from non-small-cell lung cancer (NSCLC) patients. The application of label-free imaging cytometry in NSCLC exemplifies success in delineating tumor cells with biophysical phenotypes of cells and its potential in exploring epithelial-mesenchymal transition during cancer progression.

Overall, this thesis represents a pivotal step forward in advancing biophysical cytometry from rudimentary inspection to robust quantitative analytics powered by deep learning. The establishment of comprehensive cellular imaging datasets, in conjunction with the innovative applications of CytoMAD, underscores the transformative potential of label-free imaging cytometry in both scientific research and clinical practice. We anticipate that this work will not only have a profound impact on the utility of label-free imaging cytometry, but also catalyze new strategies for cancer prognosis, ultimately improving clinical outcomes.