TransEN U-Net: Enhance Image Segmentation of Cytomegalovirus Infected Cells in Histopathological Images

Abstract

Advances in histopathological image segmentation have a significant role in the diagnosis and begin treatment immediately including a study of Cytomegalovirus(CMV) for the tissues. Histopathological change with confirmation by immuno-histochemical or in situ hybridization study is a gold standard for diagnosis of CMV tissue infection. This required pathologists to identify the histopathological change which is time-consuming and can be missed in tissue with a subtle change. Automatic analysis of histopathological images with Deep Learning(DL) can help pathologists to diagnose CMV tissue infection more accurately. Typical issues with histopathological images which impede automatic analysis are the extremely enormous size of histopathological images more than 1 gigapixel, the limitations of GPU memory, and a limited number of histopathology images. Additionally, whole slide histopathological images are split huge images into multiple small image patches by cropping using the sliding window technique. In this paper, we propose TransEN U-Net which derives a benefit of a hybrid CNN-Transformer base on the U-shaped architecture for boosting the performance of segmentation of histopathology. The transformer encoder not only is able to the patches but also the relative self-attention mechanism in order to share information between sequences. Experiment results of segmenting images by the two-dimensional indicate that the TransEN U-Net can productively discriminate CMV viral inclusions including achieving higher values in terms of DSC score.