Improving Classification of Curved Chromosomes in Karyotyping using CNN-based Deformation

Abstract

Chromosomal image analysis is an important method to diagnose chromosomal disorders. However, the image can be curved after cultivation, resulting in difficulty in chromosome recognition and analyzing the bands. While manual work of straightening the chromosomes requires an intensive labor, the computer-aided method can increase the performance as well as preserve the image details. In this paper, we investigate a method of straightening the curved chromosomes using Spatial Transformer Network (SPN) and to what extend the method affects the chromosome classification using a CNN-based method. The experiments were carried on a dataset of 28,106 chromosome images. The results show that SPN achieved compatible performance to manual method on the curved chromosomes with straight ratio of higher than 90%, yielding improvements of the classification accuracy to that of the original curved images from 3% to 5% on average. The source code and processed data are shared to support further studies.