Automatic classification of human chromosome shapes using convolutional neural network models

Abstract

Metaphase chromosome classifications based on the positional relationship between sister chromatids are used to evaluate the function of the cohesin complex, which tethers sister chromatids until cell division. Currently, classification is manually performed by researchers, which is time consuming and biased. This study aims to automate the analysis using multiple convolutional neural network (CNN)-trained models. By improving our prototype model with a 73.1% concordance rate, one of the proposed new models achieved a maximum concordance rate of 93.33% after applying a fine-tuning method and ensemble learning method. The results suggest that CNN-based models can automatically classify chromosome shapes.