CT图像下结合RCNN与U-Net的肺实质自动分割方法
2021-09-06毕蓉蓉王进科
毕蓉蓉 王进科
摘 要:针对肺CT实质自动分割问题,提出了一种改进的U-Net模型。该模型充分结合了U-Net网络和循环卷积神经网络(RCNN)的优势,可以提取图像中更加细小的特征,在拥有较少数量网络参数的情况下获得良好的性能。该模型首先在编码和解码单元中,使用带有残差单元的循环卷积层代替U-Net模型中的前向卷积层。然后,从基本网络的端到端使用了有效的特征累积方法。最后,从原始的U-Net模型中删除了裁剪和复制单元,仅保留了级联操作。为了验证提出方法的性能,将提出的模型与当前最先进的模型在LUNA16数据集上进行了实验比较,实验结果显示,提出的分割模型在分割精度方面体现出一定的优越性。
关键词:
RCNN;U-Net;医学图像分割;肺CT
DOI:10.15938/j.jhust.2021.03.011
中图分类号: TP391.4
文献标志码: A
文章编号: 1007-2683(2021)03-0074-09
Automatic Lung Segmentation Using RCNN Combing U-Net in CT
BI Rong-rong, WANG Jin-ke
(Rongcheng Campus, Harbin University of Science and Technology, Rongcheng 264300, China)
Abstract:Aiming at the problem of automatic lung CT segmentation, an improved U-Net model is proposed. The model fully utilizes the advantages of U-Net model and Recurrent Convolutional Neural Network(RCNN), and it helps extract very low-level features and ensure better performance with fewer numbers of network parameters. Firstly, in the encoding and decoding units, a recurrent convolution layer replaced the forward convolutional layers of the original U-Net models. Secondly, an effective feature accumulation method was utilized from one part of the network to the other. Finally, the cropping and copying units were removed from the original U-Net model and only the concatenation operations were retained. To evaluate the performance of the proposed scheme, the comparative segmentation experiments with state-of-the-art models were conducted in the LUNA16 datasets. The experimental results proved a superiority of the proposed method on segmentation accuracy.
Keywords:RCNN; U-Net; medical image segmentation; lung CT
0 引 言
医学图像处理技术可以帮助病人获得更快更好的诊断,CT成像因其高性价比被广泛应用于临床的辅助诊断。然而,手动分割的过程缓慢且繁琐,而自动的CT图像分割技术可以减少人为主观错误,降低时间成本和误诊率。因此,研究可靠的鲁棒的自动算法,对于高效、精准的医疗诊断具有重要的临床应用价值。
近年来,深度学习(DL)在图像分类[1]、分割[2]、检测[3]等方面得到了广泛的应用,VGG[4]、GoogleNet[5]、ResidualNet[6]等著名的深度卷积神经网络模型相继被提出。基于DL的方法在图像分类、图像分割方面都表现出了优越的性能,然而,在大多数情况下,基于DL的方法在对大规模的数据集分類时(如ImageNet),分类任务的输出是仍然是单个标签概率值。……
