变分自编码的分层解耦卷积脑肿瘤分割网络

A Variational Autoencoding-Hierarchical Decoupled Convolutional Brain Tumor Segmentation Network

三维卷积神经网络处理图像分割精度高,可以保留更多空间信息,有效解决标签不平衡问题,但存在参数量大的缺点.针对目前三维脑肿瘤分割网络内存资源占用大、硬件设备要求高、计算效率低的问题,将传统3D UNet网络中的3D卷积替换为分层解耦卷积,能够降低空间环境的计算复杂度和内存占用量,在不提高计算量的前提下显著提高分割精度,提高网络性能.为解决传统自编码器不能自主生成数据的问题,使用结合深度学习和统计学习的变分自编码器,在编码器结果中加入高斯噪声,使得编码器对结果具有鲁棒性,在编码器中加入概率分布防止过拟合,提高算法的泛化性能.采用三线性插值在三维离散采样数据的张量积网格上进行线性插值,有效避免线性方程组不断增大导致计算时间过长的问题.通过对损失函数加权混合,避免梯度弥散时出现学习速率下降现象,解决小区域分割不平衡问题,减少局部性能最优,使网络保持较高运算速度的同时有效提高分割精度,在有限内存空间最大化网络特征提取能力.在脑肿瘤公开数据集BraTS2019上的实验结果表明,该网络在增强型肿瘤、全肿瘤、肿瘤核心上的Dice值分别可达78.02%、90.05%和83.14%,参数量仅为0.30×10^(6),能够准确、高效地分割出脑肿瘤中各病灶区域,节约硬件设备的算力和内存资源,为临床应用提供可能性.

Three-dimensional convolutional neural networks have the advantages of high segmentation accuracy,can effectively solve the unbalanced label problem,and retain more spatial information.However,they involve complex calculations and several parameters.In view of the 3D brain tumor segmentation network’s high memory resource usage,high-hardware requirements,and the problem of low efficiency of computation,we propose a pseudo 3D UNet module to replace the traditional 3D convolution with a novel hierarchical decoupled convolution module,which significantly reduces the usage of memory and computation,minimizes the perception while reducing the computational complexity of the space environment,segments the 3D stereo image at once,and significantly improves the segmentation accuracy without increasing the computational cost.To solve the problem that the hidden layer output of the traditional autoencoder is easily to be chaotic and cannot generate data independently,we propose a variational autoencoder that combines deep learning and statistical learning.Gaussian noise is added to the encoder result to make the encoder robust to the result,and the probability distribution is added to the encoder to prevent over-fitting and improve the generalization performance of the algorithm.We use trilinear interpolation to perform linear interpolation on the tensor product grid of 3D discrete sampling data,which effectively avoids the long calculation time caused by the continuous increase in the linear equation system during the calculation.By weighing and mixing the three-loss functions,this study prevent the learning rate decrease caused by the gradient dispersion in a gradient descent calculation,solves the imbalance of small region segmentation,and ensures that the algorithm can effectively improve the segmentation accuracy while maintaining a high computing speed.This study uses the brain tumor open dataset BraTS2019 for testing,and the average Dice_ET,Dice_WT,and Dice_TC reached 78.02%,90.05%,and 83.14%,respectively,with only 0.30×10^(6)of the parameters.The results show that this network can accurately and efficiently segment each lesion region of a brain tumor,save the computing power and memory resources of hardware equipment,and provide the potential for clinical application.