Journal of Advances in Mathematics and Computer Science

Biomedical laboratories often use different cell types in the same assay or the same cell type in different assays. One cell type can become contaminated by another, or cells can be mis-identified, giving poor results. Addressing these issues by DNA analyses can be time-consuming, labor intensive or costly to implement. Here we uniquely employ Legendre moments (LM), Zernike moments (ZM), circularity and a genetic algorithm (GA) to advance a computer-based vision system, and we task it to identify four cell types used in virology: HeLa, Vero, BHK and PC3. By employing a k-nearest neighbor (kNN), multilayer perceptron (MLP), Convolutional Neural Networks (CNN) classifiers and a GA-selected 9-vector candidate comprising 4 ZMs, 4 LMs, and circularity, we provide adaptive system for deep machine learning. Our approach provides avenue to measure the performances of two of the conventional and popular classifiers (kNN and MLP) with a relatively recent classifier (CNN). We provide detailed mathematical treatments of the image signatures for accessibility and reproducibility in computer vision. Our methods are unique in biomedical applications. The performance of the kNN for k = 1, 2, and 3 using 10-fold cross-validationyielded accuracies of (83.59%, 82.03%, 81.25%) and (84.38%, 82.82%, 82.03%) for 8-class and 4-class training sets, respectively, drawn from the same data while those of the MLP and CNN were 86% and 87.25% respectively. These results establish the feasibility of reliable automated cell identification, with diverse applications in biological and biomedical research.