A Novel Nuclei Segmentation Algorithm Based on Convolutional Neural Network

Abstract

With the continuous improvement of medical diagnostic techniques, the segmentation of the nucleus has important diagnostic significance in the examination of clinical pathology. However, as the clinical proportion increases, the traditional segmentation technique (manual segmentation) becomes a cumbersome and unstable method because of its time consuming and the technical level of the pathologists. On the other hand, the detection of nuclear equipment required for manual segmentation is extremely expensive and the requirements of the operating environment are badly demanding. Compared with traditional image processing technology, deep learning based segmentation technology can better summarize the appearance of nuclei. At present, the most prominent deep learning semantic segmentation network architecture is U-Net. Here, in this paper we propose a novel convolutional neural network combined with the existing segmentation technologies to independently learn the pixel-level prior information in the nuclei. The special connection method adopted by its internal structure can extract the scale information of different network layers. The proposed architecture is compared with the classical U-Net and its transforms on nuclei dataset from kaggle data science bowl 2018 dataset fixes. Experiments show that the proposed method shows higher performance than other networks in accuracy, Jaccard index and F1 score. Our network has more prominent advantages for the treatment of irregular nuclei.