A Multi-Stages Chromosome Segmentation and Mixed Classification Method for Chromosome Automatic Karyotyping

The chromosome karyotyping task is vital and indispensable but tedious work for birth defect diagnosis and biomedical research. In this work, we tackle chromosome automatic karyotyping using a multi-stages chromosome segmentation and mixed classification method. Firstly, we apply a global binary threshold-based method to segment the metaphase chromosome microscope grayscale image into several image slices, consisting of chromosome instances and chromosome clusters. Afterward, we propose a mixed chromosome classification method for identifying a given image is a chromosome cluster or corresponding instance label. After that, we use a deep learning-based approach to segment chromosome cluster images into chromosome instances and apply the mixed chromosome classification model to recognize their corresponding labels. Finally, we synthesize a chromosome karyotype from all corresponding instances and labels. In the mixed classification stage, the proposed method yields 99.53 ± 0.23% classification accuracy on the clinical dataset. In segmentation stages, the proposed method achieves 90.81% comprehensive segmentation accuracy and 85.00% instance segmentation accuracy with 90.63% \(AP_{50}\) precision. The experimental results show that our proposed method is promising for solving chromosome segmentation and classification task of the clinical chromosome automatic karyotyping.