摘 要：对 3D U-NET 网络结构进行改进，提出一种 CT 影像中结节的自动分割方法。该项目在 3D U-Net 的基础上对其进行改进，改进的内容是卷积块操作采用 3×3×3，Stride=1，padding=same 的卷积，每个卷积后面相继增加 Batch Normalization、Relu 和 Dropout 操作，池化被卷积操作代替，同时加入 long skip connection 长链接，实现浅层、低水平、粗粒度特征传递下去而不消失，提升网络对形状在 10 mm 以下但亮度高结节的轮廓表示能力，同时扩大了感受野、加速了网络的收敛。实现对 CT 影像的自动、准确描述。
关键词：3D U-NET；CT 影像；长链接；感受野；浅层；低水平；粗粒度特征
DOI:10.19850/j.cnki.2096-4706.2021.21.027
基金项目：衢州市科技计划项目（2018K35）；衢州市科技计划项目（2019K01）；浙江省公益计划项目（LGF21F010002）；衢州学院大学生科技创新项目（Q20X050）
中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2021）21-0105-04
Research on Application of Improved 3D U-NET in CT Image Segmentation
LI Linjing1, HOU Junhao1, WU Jianfeng2, YANG Xiaojun3
(1.School of Electrical and Information Engineering, Quzhou University, Quzhou 324000, China; 2.Department of Nuclear Medicine,Quzhou People’s Hospital, Quzhou 324000, China; 3.Department of Radiology, Quzhou People’s Hospital, Quzhou 324000, China)
Abstract: The structure of 3D U-NET network is improved, and an automatic segmentation method of nodules in CT images is proposed. The project improves it on the basis of 3D U-NET. The improved content is that the convolution block operation adopts the convolution of 3×3×3, stripe=1 and padding=same. After each convolution, Batch Normalization, Relu and Dropout operations are added successively. Pooling is replaced by convolution operations. At the same time, long skip connection long links are added to realize the transmission of shallow layer, low level and coarse grained characteristics without disappearing, so as to improve the ability of the network to express the contour of nodules of shape less than 10 mm with high brightness, at the same time, it expands the receptive field and accelerates the convergence of the network. And then realize the automatic and accurate description of CT images.
Keywords: 3D U-NET; CT image; long link; receptive field; shallow layer; low level; coarse grained characteristic
参考文献：
[1] LIVNE M，RIEGER J，AYDIN O U，et al. A U-Net Deep Learning Framework for High Performance Vessel Segmentation in Patients With Cerebrovascular Disease [J]. Frontiers in Neuroscience，2019，13：97.DOI:10.3389/fnins.2019.00097.eCollection 2019.
[2] ÇIÇEK Ö，ABDULKADIR A，LIENKAMP S.S , et al. 3D U-Net：Learning Dense Volumetric Segmentation from Sparse Annotation. In：Ourselin S.，Joskowicz L.，Sabuncu M.，Unal G.，Wells W.（eds）Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016. MICCAI 2016. Lecture Notes in Computer Science，vol 9901. Springer，Cham. https://doi.org/10.1007/978-3-319-46723-8_49.
[3] 金征宇 . 人工智能医学影像应用：现实与挑战 [J]. 放射学实践，2018，33（10）：989-991.
[4] Roy A G ，Siddiqui S ，Plsterl S ，et al.‘Squeeze & Excite’ Guided Few-Shot Segmentation of Volumetric Images [J].Medical Image Analysis，2020，59（1）：234-238.
[5] 邹应诚 . 基于卷积神经网络的颅内动脉瘤检测方法研究[D]. 武汉：华中科技大学，2019.
作者简介：李林静（1976.03—），女，汉族，四川遂宁人，副教授，硕士，研究方向：图像理解。