摘要
针对人工判别胃癌病理图像对专业知识要求较高且费时费力的问题,提出一种基于深度学习的分割算法对胃癌病理图像的病变区域进行自动分割,为病理医生的工作提供更有依据的诊断指导。在已有的U-Net模型基础上,提出一种基于多尺度输入与特征融合的多输入融合网络(MIFNet)模型,通过对不同输入尺度特征进行自动提取和融合以实现对胃癌病理图像中病变区域的自动分割。多尺度输入数据能够帮助模型更有针对性地捕捉图像的局部和全局特征,特征融合策略能够兼顾模型对全局特征与局部特征的关注。在中国大数据人工智能创新创业大赛的"病理切片识别AI挑战赛"数据集上的实验结果显示,MIFNet在测试中的dice系数达到了81.87%,比U-Net和SegNet等模型提高了10%以上,模型的参数规模也大大下降。所以说,MIFNet模型在提高分割的准确度以及节省计算资源等方面都取得了更好的效果。
For artificially identifying images of stomach cancer pathology requires high expertise and is time-consuming,a depth learning based segmentation algorithm was proposed to automatically segment the lesion area of stomach cancer pathology images,provide more informed diagnostic guidance for the work of pathologists.Based on the U-Net model,a new deep learning model MIFNet(Multi-Input-Fusion Net)based on multi-scale input and feature fusion was proposed to segment stomach cancer pathological images automatically.Multi-scale input data can help deep learning model get the local and global features of the image pertinently;feature fusion strategy can focus on both global features and local features.The experimental results on the Pathology Slice Recognition AI Challenge dataset of Chinese Big Data Artificial Intelligence Innovation and Entrepreneurship Competition show that the dice coefficient of MIFNet reached 81.87%in the test,which is more than 10%higher than other models such as U-Net and SegNet,and the parameter size of the model is also greatly reduced.MIFNet model has achieved great results in improving the accuracy of segmentation and saving computing resources.
作者
张泽中
高敬阳
赵地
ZHANG Zezhong;GAO Jingyang;ZHAO Di(College of Information Science and Technology,Beijing University of Chemical Technology,Beijing 100029,China;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China)
出处
《计算机应用》
CSCD
北大核心
2019年第S02期107-113,共7页
journal of Computer Applications
基金
国家重点研发计划项目(SQ2017ZX106047)
北京市自然科学基金资助项目(5182018)
北京市自然科学基金重点项目(4161004)
北京市科技计划项目(Z171100000117001,Z161100000216143)
关键词
多输入融合网络
多尺度输入
特征融合
胃癌病理图像
图像分割
Multi-Input-Fusion Net(MIFNet)
multi-scale input
feature fusion
stomach cancer pathology image
image segmentation
