ABSTRACT
We present the rationale for the development of mathematical features used for classification of images stained for selected tight junction proteins. The project examined localization of zonula occludens-1, claudin-1 and F-actin in a model epithelium, Madin-Darby canine kidney II cells. Cytochalasin D exposure was used to perturb junctional localization by actin cytoskeleton disruption. Mathematical features were extracted from images to reliably reveal characteristic information of the pattern of protein localization. Features, such as neighbourhood standard deviation, gradient of pixel intensity measurement and conditional probability, provided meaningful information to classify complex image sets. The newly developed mathematical features were used as input to train a neural network that provided a robust method of individual image classification. The ability for researchers to make determinations concerning image classification while minimizing human bias is an important advancement for the field of tight junction cellular biology.
