基于改进U-Net的人脑黑质致密部分割

Segmentation of Brain Substantia Nigra Pars Compacta Based on Improved U-Net

人脑黑质致密部分割能够为帕金森病的诊断提供一定依据。黑质致密部在人脑核磁共振成像中像素占比低、类间差异小,为提高计算机辅助诊断系统对人脑黑质致密部的分割精度,提出一种基于改进U形神经网络(U-Net)的人脑黑质致密部分割方法。为了提取更多有效的多尺度图像语义特征,结合U-Net的跨连接结构并采用多头注意力机制,同时融合基于Transformer编码器的高维语义编码模块以提取高维语义特征,避免浅层噪声对特征造成的影响。建立多任务模型并设计基于二维高斯核权重掩膜的损失函数,解决神经网络分割模型因多次下采样造成的不连续分割误差问题。构建包括140个帕金森病患者以及48个健康对照者的高精度核磁共振脑成像数据集进行实验,结果表明,相较常用的医疗影像分割方法 R2U-Net、HANet等,该方法的多任务分割效果取得明显提升,戴斯相关系数和AUC指标分别达到0.869 1和0.943 9,消融实验结果也验证了改进编码器和改进损失这2个模块的有效性。

The segmentation of brain Substantia Nigra pars compacta(SNpc) can provide the basis for Parkinson’s disease diagnosis. SNpc is a low-pixel-ratio object in Magnetic Resonance Imaging(MRI) of the human brain,and the pixels with different labels are similar.This study proposes a method based on an improved U-Shape Neural Network(UNet) to improve the segmentation accuracy of a computer-aided diagnosis system for brain SNpc. The cross-connection structure of U-Net is combined with the multihead attention mechanism to extract more effective multiscale image semantic features. The high-dimensional semantic encoding module based on a Transformer encoder is fused to extract highdimensional semantic features and eliminate the effect of shallow noise on features.The multitask model is established,and the loss function based on the two-dimensional Gaussian kernel weight mask is designed to solve the discontinuous segmentation error caused by multiple down-samples of the neural network segmentation model. A high-precision MRI brain imaging dataset comprising 140 Parkinson’s patients and 48 healthy controls was constructed for experiments. The results show that compared with the widely used medical image segmentation methods,such as R2U-Net and HANet,the multitask segmentation effect of this method improved significantly. The Dais phase relationship number and the Area Under the Curve(AUC)index reached 0.869 1 and 0.943 9,respectively.Furthermore,the results of the ablation experiment verify the effectiveness of the improved encoder and the improved loss modules.