An Improved Segmentation for Raw G-Band Chromosome Images

Karyotyping is one of the most common procedure in cytogenetics to identify and evaluate the presence of genetic defects or disorders. Developing an automatic analysis algorithm for this procedure was studied by numerous other researchers before. One of the fundamental parts of this process is the segmentation of chromosomes, which is followed by classification. Despite recent improvements in deep neural networks and image classification, high-quality segmentation is essential to achieve accurate classification results. Still, automatic segmentation and extrication of touching and overlapping chromosomes are current problems. In the light of the above, this paper is intended to present an automatic segmentation and separation of G-band human chromosomes. we have tried to present an algorithm to overcome all the difficulties such as accurate thresholding, separation of touching and overlapping chromosomes. Unlike previous studies, we are also focusing on issues that may increase the segmentation quality beyond the separation of overlapping chromosomes. In this work, we present an end-to-end segmentation of chromosome images. This process includes noise removal and rejection of unwanted objects, separating fore and background, a binary watershed approach to divide intuitively and easily detectable clusters, finding a geodesic path between chromosomes in remaining clusters and finally disentangling of overlapped and complex clusters. We also built a user interface and let a little human intervention to detect unnoticed clusters. We have tested the proposed method on 145 chromosome images that contain 6678 chromosomes and 6532 (97.8%) of them have been correctly extracted.