Cell Nuclei Segmentation in Cryonuseg dataset using Nested Unet with EfficientNet Encoder

Tuan Le Dinh, Suk-Hwan Lee, Seong-Geun Kwon, Ki-Ryong Kwon

Abstract: Cell Nuclei Segmentation is one of the important stages in clinical research. Many applications in medical treatment, drug discoveries, and disease diagnosis can be benefited from cell nuclei segmentation results. Although it plays a significant role, the task of cell nuclei segmentation seems tedious and time-consuming because of the large number of cells in one histopathology image and the manual nature of the task. The segmentation task of nuclei cells can be automated by computer programs which help to save lots of time and work for histopathologists and also produce stable results, preventing mistakes made by humans. Recently, with the emergence of deep learning methods, many segmentation tasks were done by neural nets with superb output and outperforming many traditional methods. In this paper, we run experiments with CryonuSeg dataset using Nested Unet, a variant of the Unet neural net for medical and biomedical image segmentation and combine the model with EfficientNet as the encoder of the network. The modified network results surpassed many other state-of-the-art techniques on medical image segmentation tasks with the Dice score = 0.929, AJI=0.604, and PQ=0.503.