Analyzable Metaphase Image Selection Using Deep Learning

Identification of good metaphase is a prerequisite and an essential step in human karyotyping which is used for analysis of genetic abnormalities in human beings. Most of the research work is focused on automation of karyotyping process but despite being crucial step, very scant attention is given to metaphase image selection. The conventional method, still used by cytogeneticists is a manual visual search of good metaphase spread from microscopic slides. This system is highly dependent on individual observations and suffers from drawbacks such as complexity, tediousness, subjective, time consuming and needs a trained expertise. In this study, we propose a novel approach for the selection of analyzable metaphase images using three publicly available datasets. We evaluated seven critical features namely chromosome spread quality, morphology, focus and clarity, staining quality, metaphase stage, background, and field of view for classification of metaphase images using deep learning models, including MobileNetV2, ResNet50, DenseNet121, VGG16, and EfficientNetBO. Among these, EfficientNetBO outperformed the others, establishing a new baseline for this application. To the best of our knowledge, this is the first work to employ publicly available datasets for the automated selection of analyzable metaphase images. Moreover, our proposed method can run efficiently on any platform with a Python interpreter installed, achieving high classification performance and flexibility.