Classification of Human Metaspread Images Using Convolutional Neural Networks

Abstract

Chromosomes are the genetic information carriers. Any modification to the structure or the number of chromosomes results in a medical condition termed as genetic defect. In order to uncover the genetic defects, the chromosomes are imaged during the cell division process. The images thus generated are termed as metaspread images and are used for identifying the genetic defects. It has been observed that the metaspread images generally suffer from intensity inhomogeneity and the chromosomes are also present in varied orientations, and as a result finding genetic defects from such images is a tedious process. Therefore, cytogeneticists manually select the images that can be used for the purpose of uncovering the genetic defects and the generation of the karyotype. In the proposed approach, a novel method is being presented using DenseNet architecture of the convolutional neural networks-based classifier, which classifies the human metaspread images into two distinct categories, namely, analyzable and non-analyzable based on the orientation of the chromosomes present in the metaspread images. This classification process will help to select the most prominent metaspread images for karyotype generation that has least amount of touching and overlapping chromosomes. The proposed method is novel in comparison to the earlier methods as it works on any type of image, be it G band images, MFISH images or the Q-banded images. The proposed method has been trained by using a ground truth of 156 750 metaspread images. The proposed classifier has been able to achieve an error rate of 1.46%.