Abstract
Chromosome analysis and identification from metaphase images is a critical part of cytogenetics based medical diagnosis. It is mainly used for identifying constitutional, prenatal and acquired abnormalities in the diagnosis of genetic diseases and disorders. The process of identification of chromosomes from metaphase is a tedious one and requires trained personnel and several hours to perform. Challenge exists especially in handling touching, overlapping and clustered chromosomes in metaphase images, which if not segmented properly would result in wrong classification. We propose a method to automate the process of detection and segmentation of chromosomes from a given metaphase image, and in using them to classify through a Deep CNN architecture to know the chromosome type. We have used two methods to handle the separation of overlapping chromosomes found in metaphases - one method involving watershed algorithm followed by autoencoders and the other a method purely based on watershed algorithm. These methods involve a combination of automation and very minimal manual effort to perform the segmentation, which produces the output. The manual effort ensures that human intuition is taken into consideration, especially in handling touching, overlapping and cluster chromosomes. Upon segmentation, individual chromosome images are then classified into their respective classes with 95.75\% accuracy using a Deep CNN model. Further, we impart a distribution strategy to classify these chromosomes from the given output (which typically could consist of 46 individual images in a normal scenario for human beings) into its individual classes with an accuracy of 98\%. Our study helps conclude that pure manual effort involved in chromosome segmentation can be automated to a very good level through image processing techniques to produce reliable and satisfying results.
Abstract (translated)
URL
https://arxiv.org/abs/2209.05414
