Digital Quantification of Cluster Features for Salivary Tissue Engineering

Abstract

Objectives: Nearly all patients with head and neck cancer are prescribed radiation therapy, which although effective, also induces progressive and permanent salivary gland damage. The resulting hyposalivation reduces quality of life through impacts on taste, speech, and overall oral health. Tissue engineering is a potential route towards de novo regeneration. Human salivary stem/progenitor cells (hS/PCs) encapsulated within HA-PEGDA hydrogels survive and proliferate in vitro, but do not branch spontaneously as observed in the in vivo morphology. One approach to promote branching is to use multiphoton laser ablation to create a 3-D space within the hydrogel to guide cell growth over time. Digital imaging tools will be used to quantitatively evaluate and optimize the clustering of cells.

Experimental Methods: To identify the best clustering cells from 6 patient samples over time, Ilastik, a supervised machine learning software, quantified various features of the clusters using a random forest classifier and user annotation. Preprocessing of microscope images occurred to ensure compatibility with Ilastik. Pixel classification was conducted to differentiate between cells and the wells in which they were contained. Positive and negative control images, identified subjectively, were annotated to train the model. Afterwards, batch processing occurred to analyze the remaining images. Subsequently, object classification was used to identify clusters, where various features were quantified including object area, number of branches, total counts, kurtosis, and variance.

Results: For each feature, averages within each well were calculated. Then averages and standard error mean were found within the 6 replicates in each plate. Poor clustering patient samples were identified due to significant differences compared to other samples over time in various features analyzed.

Conclusion: Comparison of features revealed patient samples that were not appropriate for future use due to poor clustering. However, further image analysis is required to objectively identify which patient samples were more suitable.

This study was supported by the UTSD Student Research Program, and NIH/NIDCR grant R03DE028988.