基于2D深度学习网络的全心脏MR图像分割

Cardiac MR image segmentation based on 2D deep learning network

传统机器学习方法在心脏MR图像分割中存在分割精度较差、计算复杂度高,特别是难以同时分割左、右心室及心肌等问题.提出了将改进的全卷积神经网络自动分割方法用于心脏MR图像分割.在训练网络下采样与上采样路径中加入批归一化层,保持每层网络大小与维度一致,使用较高学习率训练网络,加速收敛,降低过拟合.结合像素交叉熵损失函数与Dice损失函数作为新的组合加权损失函数,提高分割精度.实验结果表明,能实现较好的分割精度和计算复杂度,且能同时分割出心脏图像中的左、右心室和心肌.

Because of the poor segmentation effect,high computational complexity of the traditional machine learning method in cardiac MR image segmentation,an automatic method of segmentation of cardiac image based on improved fully convolutional neural network is proposed.In the down-sampling and up-sampling path of training network,batch normalization layer is added,it keeps the scale of each layer and dimension,high learning rate is used to accelerate the network convergence and more weight interfaces are fall into the data distribution which used for reducing network overfitting.A novel loss function which combined cross-entropy with Dice loss leads to qualitative improvement in cardiac segmentation.The experimental results show that this method can achieve better segmentation accuracy and computational complexity and segment the left ventricle,right ventricle and myocardium well.