Segmentation Of Raw G-band Chromosome Images Using Convolutional Neural Networks

Analysis of the chromosome images plays an important role to discover one’s genetic information and possible disorders.Nowadays,genetics is considered to be related to many disorders and abnormalities.Scientists work hard for decades to automate this vital operation.Segmentation has a very substantial place in the analysis of the human genome.However,without an automatic solution,it becomes a very time-consuming and error-prone procedure.Thus,a solution for this purpose is proposed here.Because the extraction of chromosomes is the first step in the analysis concept,it is that important as well,because the accuracy of segmentation affects all the following works.But,it’s not an easy task.First of all,the background of the chromosome images are very noisy and it contains some other cell particles,stain debris.The other problem is that chromosomes have a big variety in their shapes and structures.In addition,because they are string-like objects,they tend to overlap on each other or touch other chromosomes.To overcome the problems,firstly,background removal was studied.A U-Net based full convolutional neural network model and an adaptive local thresholding method were compared for this purpose.With the heavy usage of optimization methods such as batch normalization,data augmentation,and dropout,deep learning based method achieved very good results on the images what the adaptive thresholding had very low success.To separate touching chromosomes,two approaches,watershed segmentation,and geodesic path algorithms were deployed simultaneously.Watershed transforms improved the algorithm in terms of time consumption,and the geodesic path method provides an improvement for smoother and more accurate separation.Finally,overlapping chromosome disentangling procedure,evaluate the shape of the object to find out the overlapping regions.U-Net based CNN model achieved 97% Dice Similarity Coefficient score for the binary mask of chromosomes.The rest of the algorithm achieved 97.8% accuracy in terms of the number of correctly segmented chromosomes,by extracting 6532 out of6678 chromosomes correctly.