Publication

Abstract : Nucleus segmentation of H &E-stained (hematoxylin and eosin) histopathology images is a crucial step in developing a computer-aided diagnostic (CAD) system for cancer prediction and diagnosis. Nucleus segmentation technology has made it possible to subjectively and quantitatively analyze thousands and thousands of nuclei in H &E-stained histopathology images. The segmentation of variable touching nuclei, on the other hand, is a substantial issue during nuclei segmentation. This paper proposes a deep learning model for automatic nuclei segmentation. Since UNet performs well in medical image segmentation areas, a modified version of UNet architecture is utilized for the segmentation of nuclei. Generally, UNet architecture has a contracting path and an expansion path. In this paper, the contracting path or encoder of UNet is replaced by EfficientNetV2-L architecture. By making this change, the proposed model achieved the Jaccard index of value 0.85 along with a dice coefficient of 0.91 on the MoNuSeg dataset. After nuclei segmentation with the proposed model, a post-processing method is utilized to filter out noise in predicted masks images by using the watershed method.