计算机科学
特征(语言学)
人工智能
分割
棱锥(几何)
编码器
模式识别(心理学)
卷积神经网络
计算机视觉
图像分割
豪斯多夫距离
频道(广播)
数学
哲学
语言学
几何学
操作系统
计算机网络
作者
Yihan Zhang,Zhengyao Bai,Yilin You,Xuheng Liu,Xiao Xiao,Zhu Xu
标识
DOI:10.1007/978-3-031-46317-4_3
摘要
Automated segmentation of gliomas in MRI images is crucial for timely diagnosis and treatment planning. In this paper, we propose an encoder-decoder network for brain tumor MRI image segmentation that can address the problem of imbalanced data classification in brain tumors. Our method introduces a multi-path feature fusion module and a multi-channel feature pyramid module into the U-Net architecture to alleviate the problems of low segmentation accuracy for small targets and insufficient use of multi-scale information. The output feature maps from each level of the encoder are fed into the multi-path feature fusion module to achieve multi-scale feature fusion. Both the encoder and decoder parts of the proposed method use cascaded dilated convolutional modules, which can effectively extract additional feature information without increasing the number of parameters. In the BraTS'2019 training dataset of 50 samples, our method achieved Dice values of 0.8551, 0.8728, and 0.7906, and Hausdorff values of 2.5693, 1.5845, and 2.7284 for the entire tumor, tumor core, and enhanced tumor, respectively. The proposed encoder-decoder network shows significant performance advantages in brain tumor MRI image segmentation, especially in addressing the problems of low segmentation accuracy for small targets and insufficient use of multi-scale information. Moreover, our method can extract additional feature without increasing the number of parameters, demonstrating good practicality and versatility. These results suggest that our method can provide an effective solution for automated segmentation of MRI brain tumors.
科研通智能强力驱动
Strongly Powered by AbleSci AI