Comparing Machine Learning and Binary Thresholding Methods for Quantification of Callose Deposits in the Citrus Phloem

Abstract

Callose is a polysaccharide that can be fluorescently stained to study many developmental and immune functions in plants. High-throughput methods to accurately gather quantitative measurements of callose from confocal images are useful for many applications in plant biology. Previous callose quantification methods relied upon binary local thresholding, which had the disadvantage of not being able to differentiate callose in conditions with low contrast from background material. Here, a measurement approach that utilizes the Ilastik supervised machine learning imagery data collection software is described. The Ilastik software method provided superior efficiency for acquiring counts of callose deposits. We also determined the accuracy of these methods as compared to manual counts. We demonstrate that the automated software methods are both good predictors of manual counts, but that the Ilastik counts are significantly closer. Researchers can use this information to guide their choice of method to quantify callose in their work.