基于3D ERA U-Net的MRI海马体分割

MRI Hippocampus Segmentation Based on 3D ERA U-Net

海马体萎缩率是痴呆症等神经系统类疾病的早期诊断标志,因此精确地对海马体进行分割有利于准确测算海马体萎缩率,从而更高效地辅助医生预诊断病人的健康状态。传统图像分割技术在分割体积小、形状复杂的海马体时效果并不理想,因此本文提出了一种3D Efficient Residual Attention U-Net模型(3D ERA U-Net)。首先,利用改进的残差注意力机制将高层特征信息与低层特征信息融合,丰富二者的信息交流,使模型高质量地学习海马体细节信息。其次,在编码器和解码器过渡区加入一个扩张模块,在解码之前强化模型的感受野,对MR图像这样的三维图像起到了联系全局信息的作用。最后,改进了一种复合损失函数,加强关注边界分割。将该模型在ADNI数据集上进行实验,相比较其他海马体模型,整体精度提高,左右海马体平均Dice系数为89.38%、Precision为88.44%、Recall为89.29%,优化效果提升明显。

The rate of hippocampal atrophy is a marker of early diagnosis of neurological diseases such as de-mentia. Therefore, accurate segmentation of the hippocampus is beneficial to accurately measure the rate of hippocampal atrophy, so as to more efficiently assist doctors to predict the health status of patients. Traditional image segmentation techniques are not effective in segmenting the small volume and complex shape of the hippocampus, so this paper proposes a 3D Efficient Residual At-tention U-Net model (3D ERA U-Net). Firstly, the improved residual attention mechanism was used to integrate the high-level feature information with the low-level feature information to enrich the information exchange between the two, so that the model could learn the details of the hippocam-pus in high quality. Secondly, a Dilation Module is added to the transition area between encoder and decoder to strengthen the receptive field of the model before decoding, which plays a role in con-necting global information for 3D images such as MR Images. Finally, a compound loss function is improved to pay more attention to boundary segmentation. The model was tested on the ADNI data set. Compared with other models, the overall accuracy was improved, with the average Dice coeffi-cient of the left and right hippocampus being 89.38%, the Precision 88.44%, and the Recall 89.29%. The optimization effect was significantly improved.