Abstract:

A 3D Convolutional Neural Network（CNN） algorithm combining attention model is proposed to improve the accuracy of glioma segmentation. The image blocks of three different scales are inputted, and then after 9 convolutional layers and one classification layer, the 3 different classification results are obtained. The output results are obtained through multiply the classification results with the weights learned of the attention mechanism and add them pixel by pixel. In addition, the algorithm uses a hyperparameter loss function that combines the Dice loss function with the Focal loss function to improve the segmentation results. Experiments show that the Dice coefficient of the algorithm reaches 95.31%, 80.12%, and 82.25% in the whole region, core region and enhanced region, respectively. Compared with the existing glioma segmentation algorithm deepmedic, the Dice coefficient in the whole region, core region and enhanced region increased by 3%, 2%, and 6%, respectively. It has important clinical significance in the segmentation of Glioblastoma multiforme.

Key words: segmentation, glioblastoma multiforme, 3D convolutional neural network, attention, hyperparameter loss function

摘要：

为了提升脑胶质瘤分割精度，提出一种结合注意力机制的3D卷积神经网络算法。输入3个不同尺度的图像块，经过9个卷积层和1个分类层后得到3个不同的分类结果，将分类结果与注意力学习到的权重相乘并逐体素相加得到输出。此外该算法采用了一种混合Dice损失函数与Focal损失函数的超参数损失函数。实验表明，该算法的Dice系数在整体区域、核心区域以及增强区域分别达到了95.31%、80.12%、82.25%。与已有的一种脑胶质瘤分割算法deepmedic相比，整体区域、核心区域以及增强区域的Dice系数分别提升了3%、2%、6%。在脑胶质瘤分割方面，具有重要的临床意义。

关键词: 分割, 脑胶质瘤, 3D卷积神经网络, 注意力机制, 超参数损失函数