Abstract: In order to improve the accuracy and robustness of medical image segmentation, a medical image segmentation method based on improved convolutional neural network is proposed. Firstly, convolution neural network is used to segment the 2D slice sequence in coronal, sagittal and axial views, and then the segmentation results under the three views are integrated to obtain the final results. The convolution neural network is composed of encoding part, bidirectional convolution long short-term memory network (BDC-LSTM) and decoding part. In order to obtain multi-scale information and expand the receptive field of convolution layer, the encoding part uses asymmetric convolution layer and dilated convolution of different sizes. In addition, BDC-LSTM is used to fully extract the relevant information between the slice sequences in a single view between the encoding and decoding parts, thus improving the segmentation accuracy. Taking the hippocampus segmentation as an example, on the ADNI standard dataset, the similarity coefficient, sensitivity and positive prediction rate are used as the evaluation criteria, and the accuracy rates are 89.36%, 88.73% and 90.16%, respectively. The experimental results show that the proposed algorithm is more competitive in accuracy.

Key words: medical image segmentation, magnetic resonance imaging (MRI), convolutional neural network, long short-term memory (LSTM), multi-view ensemble

摘要： 为了提高医学图像分割的精确性和鲁棒性，提出了一种基于改进卷积神经网络的医学图像分割方法。首先采用卷积神经网络对冠状面、矢状面以及横断面三个视图下的2D切片序列进行分割，然后将三个视图下的分割结果进行集成，得到最终的结果。其中卷积神经网络由编码部分、双向卷积长短记忆网络（BDC-LSTM）和解码部分组成。为获取多尺度信息，扩大卷积层的感受野，编码部分使用不同大小的非对称卷积层和空洞卷积。此外，在编码和解码部分之间使用BDC-LSTM，充分挖掘单视图下切片序列间的相关信息，从而提高分割精度。以海马体分割为例，在ADNI标准数据集上，以相似性系数、灵敏度和阳性预测率作为评判标准，准确率分别达到了89.36%、88.73%和90.16%。实验结果表明，该算法在准确率上更具竞争力。

关键词: 医学图像分割, 磁共振成像（MRI）, 卷积神经网络, 长短记忆网络（LSTM）, 多视图集成