基于改进Res-Unet的脑肿瘤核磁图像分割算法

Improved Res-Unet-based NMR image segmentation algorithm for brain tumors

针对全卷积网络在脑肿瘤核磁图像分割中信息丢失严重,分割精度差等问题,提出一种基于改进Res-Unet模型的脑肿瘤核磁图像分割算法。通过将深度残差结构融入到U-Net网络得到具有104个卷积层的深层Res-Unet网络,并将dropout整合到网络中减少训练过拟合,提高了网络的深度,加强了网络对特征表达的准确度。最后引入注意力机制,充分利用脑肿瘤核磁图像的空间信息和上下文信息。该算法采用Dice系数等指标评价,肿瘤整体区域达到0.90分,肿瘤核心区域为0.83分,肿瘤增强区域为0.80分。相比传统网络分割模型,本算法具有更好的分割性能。

To address the problems of serious information loss and poor segmentation accuracy of full convolutional network in brain tumor MRI image segmentation,a brain tumor MRI image segmentation algorithm based on improved Res-Unet model is proposed.The deep Res-Unet network with 104 convolutional layers is obtained by incorporating the depth residual structure into the U-Net network,and the dropout is integrated into the network to reduce the training overfitting,improve the depth of the network,and enhance the accuracy of the network for feature representation.Finally,attention mechanism is introduced to make full use of the spatial and contextual information of brain tumor MRI images.The algorithm is evaluated by Dice coefficient and other indexes,and achieves 0.90 score for the overall tumor region,0.83 score for the tumor core region,and 0.80 score for the tumor enhancement region.Compared with the traditional network segmentation model,the algorithm in this paper has better segmentation performance.