(F-229) Label-free machine learning-based viability analysis of 3D tissue cultures

Introduction::
Quantitative analysis of cell response and viability is a fundamental measurement across biology and pharmacology. Traditional techniques use labels and often destroy samples, preventing longitudinal analysis. For example, one widely used industry-standard viability assay is CellTiterGlo (CTG). After cell lysis, the CTG assay reacts with released ATP, providing an indirect and fast measurement of cell viability. However, we and other researchers noticed that the quantitative CTG readouts from 3D cell cultures are not universally in agreement with expert assessments. Given the critical role of cell viability results, developing an alternative approach that is also compatible with longitudinal investigations is urgently needed.




Advances in computational power coupled with machine learning-based algorithms have greatly accelerated the identification of features in biological images, such as cell shape, size, and arrangement. Machine learning-based image analysis has the potential to provide information about biological function and structure, simultaneously. In this work, we developed a machine learning-based analysis approach for label-free 3D tissue cultures. Our method performs comparably well to human expert analysis, out-performs CTG analysis and reports information regarding the health of the spheroids across multiple days without requiring cell or spheroid lysis.




Materials and Methods::
The Segmentation Algorithm to Assess the ViabilitY of 3D spheroid cultures, or SAAVY, combines machine learning-driven region identification and a viability assignment algorithm. SAAVY provides a significant amount of information for every image on a per-spheroid level and as the image average such as, spheroid size, circularity, viability, area, and total count.




As proof-of-concept, we challenged the ability of this CNN-based algorithm against blinded human expert-assigned viabilities to 2D images of spheroids. We further compared performance against the industry-standard viability assay (CTG). Unless indicated, pancreatic ductal adenocarcinoma (PDAC) cell spheroids that are not labelled are analyzed. We also created a series of samples with a range of artificially noisy backgrounds by incorporating nanoparticles and challenge SAAVYs ability to accurately assess viability and differentiate spheroids from background noise.




Results, Conclusions, and Discussions::
SAAVY performs similarly to human experts at identifying and classifying spheroids as alive or dead, but with significantly faster decision-making. Our algorithm offers improved viability estimation compared to the CellTiter Glo (CTG) assay, which often conflicts with expert evaluations, and is tolerance to a medium degree of background noise. SAAVY’s time-agnostic viability measurement revealed differences in cellular response in a therapeutic-response longevity study. Although further expansion of the training dataset is required to estimate the viability of other cell types more accurately, SAAVY provides a promising starting point for real-time, label-free viability analysis of 3D tissue culture.




Acknowledgements (Optional): :

References (Optional): :