Raw G-Band Chromosome Image Segmentation Using U-Net Based Neural Network

Emrecan Altinsoy, Can Yilmaz, Juan Wen, Lingqian Wu, Jie Yang, Yuemin Zhu

Abstract: Chromosome analysis plays an important role in investigating one’s genetic disorders and abnormalities. Many works are done on automating this operation for decades. Segmentation of chromosomes is the first step of this process, and it is essential for the next step which is classification. However, it is not an easy task due to a very noisy background, the presence of other cells and the variation of chromosome structures. In this paper, we propose a raw G-band chromosome image segmentation method using U-net based convolutional neural network. To this end, we constructed a raw G-band chromosome dataset which consists of 40 images. In order to prevent over-fitting, we implemented augmentations on the training and the validation set images. The trained model achieved 96.97% dice score. The experimental results showed that, the convolutional neural network can provide satisfying results, especially with highly noisy images.