Chromosome Extraction Based on U-Net and YOLOv3

Abstract

Karyotype analysis based on chromosome banding and microscopic imaging is an important means for the diagnosis of genetic symptoms. Chromosome extraction is one of the key steps in karyotype analysis, but it faces some complex situations i.e. chromosome overlaps and adhesions, which are still a challenge for traditional algorithms. Here, we proposed a method for chromosome extraction based on deep learning. In this method, U-Net was used to segment the original micrographs to remove background noise such as nuclei and other interferences. Then YOLOv3 was used to detect and extract each chromosome. Further, U-Net was used again to extract the single chromosomes precisely. The results show that this method can remove effectively the interferences outside the chromosomes, and accurately extract the overlapping and adhesive chromosomes. The accuracy of extracting chromosomes from the raw G-band chromosome images reaches 99.3%. This method is of great significance for the development of automatic karyotype analysis technology.

Keywords

• Chromosome extraction
• U-Net
• YOLOv3