Edge Preserved U-Net for chromosome edge detection

Karyotype is a visual representation on individual chromosome. Boundary detection is the major challenging and important step before chromosome classification and arrangement. We propose an Edge Preserved U-Net (EPU-Net) which is a deep learning architecture. EPU-Net has preprocessing block, a segmentation-Net which is an integration of guided filter, Sobel edge detector. The proposed architecture has convolution layer, pooling layer and upsampling operation with skip connection for preserving the spatial and background information. Edge Preserving Skip Layer (EPSL) and Edge Preserving Block (EPB) with dilated convolutions are introduced to preserve boundary details and improve feature extraction at multiple scale. The performance of the proposed architecture is evaluated on custom-designed dataset. Experimental results show that EPU-Net delivers better segmentation accuracy as 99.80%, IoU score of 99.60%, recall of 99.89%, specificity of 99.71% and structure similarity index metric of 0.9985. The results demonstrate that the proposed method is robust and achieves substantially better performance compared to existing state-of-the-art segmentation techniques.