文章摘要

文章信息/Info

Title:: Nuclear Segmentation of Clear Cell Renal Cell Carcinoma based on Deep Convolutional Neural Networks

作者:: 鲁浩达1△; 徐军1; 刘利卉1; 周超1; 周晓军2; 张泽林1; 1.南京信息工程大学，江苏省大数据分析技术重点实验室，南京210044；2.南京军区南京总医院，南京210044

Author(s):: LU Haoda1; XU Jun1; LIU Lihui1; ZHOU Chao1; ZHOU Xiaojun2; ZHANG Zelin1; 1.Nanjing University of Information Science and Technology，Jiangsu Key Laboratory of Big Data Analysis Technique，Nanjing 210044, China；2.Nanjing General Hospital，Nanjing 210044

Keywords:: Segmentation; Convolution neural network; Nuclei; Clear cell renal cell carcinoma; Pixel-wise

摘要:: 肾透明细胞癌病理图像中细胞核的形态和位置信息对肾癌的良恶性分级诊断具有重要意义，为提高肾透明细胞癌细胞核分割的质量，本研究提出了基于深度卷积神经网络的细胞核分割方法。首先，根据标定的病理图像中细胞核轮廓，构建细胞核分割样本集；然后，深度卷积神经网络通过隐式特征学习对细胞核分割模型进行训练，避免人为设计特征；最后，利用细胞核分割模型对病理图像进行逐像素分割。实验结果表明，深度卷积神经网络的细胞核分割算法在肾透明细胞癌细胞核分割的像素准确率高达90.33%，细胞核分割性能稳定, 深度卷积神经网络强大的鲁棒性和适应性使得肾透明细胞癌细胞核自动分割具有可能。

Abstract:: The shape feature and location information of clear cell renal cell carcinoma’s nucleus is important for the diagnosis of benign and malignant renal cell carcinoma. To improve nuclear segmentation accuracy, nuclear segmentation based on deep convolution neural networks was proposed. First, nuclear sample dataset was formed according to the nuclear contour labeled by pathologist. Then, deep convolution neural network extracted implicit nuclear feature instead of artificial nuclear characteristic and the nuclear segmentation model were trained. Finally, the nuclear segmentation model did nuclear segmentation by pixel-wise. The experimental results show that the clear cell renal cell carcinoma’s nucleus segmentation algorithm of deep convolution neural network is as high as 90.33% in the nucleus pixel accuracy and the nucleus segmentation performance is stable. The strong robustness and adaptability of deep convolution neural network makes nuclear auto-segmentation possible.