Prediction of Centromere Location in Human Chromosome Using Convolutional Neural Networks

Abstract

Accurate determination of chromosome centromere location is of high importance in cytogenetics, particularly in karyotyping, chromosome classification and determination of exposure to genotoxic environmental effects. This study investigates the ability of CNN to accurately predict the human chromosome centromere location and the effect centering chromosomes in images, by predicted centromere location, has on classification accuracy. Dataset, used to train and test CNN models, contained 8283 annotated individual chromosome images. Prior to performing centromere detection, followed by chromosome classification, the individual chromosome images are preprocessed using sequence of filtering algorithms. The CNN model achieved an average error of 0.5586 and 0.4543 in predicting x and y coordinates of centromere location, respectively. The achieved classification accuracy of randomly oriented and centered chromosomes in images, is 71.10 and 96.73%, respectively. Achieved increase in chromosome classification accuracy of 25.63% highlights importance of chromosome centromere detection, importance of positional variation removal, and high performance of CNN in prediction of centromere location and chromosome classification.