| 摘 要: 针对脑肿瘤分割研究中对肿瘤边缘分割不够精确，分割目标较小而背景因素过大的问题，提出改进的V-C-Net卷积神经网络模型。在原始的V-Net模型基础上，加入CBAM混合注意力机制，使模型更关注脑肿瘤图像的病变部分，结合三种损失函数优点并与训练次数间建立联系提出改进的联合损失函数，并对核磁共振脑肿瘤3D多模态医学图像做重叠分块预处理，利用改进的模型对BraTs数据集进行分割实验，并与FCN全卷积网络、3D-Unet神经网络、传统V-Net神经网络模型的分割效果做对比分析。模型在测试集上的Dice系数(DSC)、交并比(IoU)、敏感度(Sensitivity)、精确率(Precision)及豪斯多夫距离(HD)五个性能指标平均值分别为90.78%、89.68%、91.70%、96.48%、0.451，实验结果表明改进的V-C-Net模型对脑肿瘤病变部分分割性能更优。 |
| 关键词: 脑肿瘤分割；V-C-Net卷积神经网络；CBAM混合注意力机制；联合损失函数；重叠分块预处理 |
|
中图分类号: TP39
文献标识码: A
|
| 基金项目: 甘肃省科技计划项目(定西地区农村电子商务营销综合能力提升)(20CX9NA095). |
|
| Multi-level Segmentation Experiment of Brain Tumor Images based on Improved V-C-Net Convolutional Neural Network |
|
MO Xiuyuan1, WU Lili2, LU Zhixiang3
|
( 1. Faculty of Science, Gansu Agricultural University, Lanzhou 730000, China ;
2. College of Information Science and Technology, Lanzhou 730000, China ;
3. Key Laboratory of Ecohydrology of Inland River Basin, Northwest Institute of Eco -Environment and Resources Chinese Academy of Sciences, Lanzhou 730000, China)
2863216812@qq.com; wull@gsau.edu.cn; lzhxiang@lzb.ac.cn
|
| Abstract: Aiming at the problems of inaccurate tumor edge segmentation, small segmentation target and excessive background factors in brain tumor segmentation, this paper proposes an improved V-C-Net convolutional neural network model. On the basis of the original V-Net model, CBAM (Convolutional block attention module) attention mechanism is added to make the model pay more attention to the pathological part of the brain tumor image. Combined with the advantages of three loss functions and the establishment of links between the three loss functions and the training times, an improved joint loss function is proposed, and the preprocessing of overlapped blocks is conducted for the 3D multimodal medical image of brain tumor in MRI (Nuclear Magnetic Resonance). The improved model is used to perform the segment test on the BraTs dataset, and its segmentation results are compared with those of FCN full convolution network, 3D-Unet network and the original V-C-Net network model. The average values of the five performance indexes of the model on the test set, namely, Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Sensitivity, Precision and Hausdorff distance (HD), are 90.78%, 89.68%, 91.70%, 96.48% and 0.451 respectively. The experimental results show that the improved V-C-Net model has better segmentation performance for brain tumor lesions. |
| Keywords: brain tumor segmentation; V-C-Net convolutional neural network; CBAM attention mechanism; joint loss function; preprocessing of overlapped blocks |
