基于DenseUnet模型的核磁共振图像下肛提肌的自动分割

Automatic segmentation of levator ani muscle in MRI images based on DenseUnet model

目的基于盆底的核磁共振图像(MRI),构建深度学习自动分割模型,进行盆底MRI图像的智能分割研究,以提高肛提肌分割效率和精准度。方法以DenseUnet模型为主干,搭建一个主要由编码器模块、上下文提取模块和解码器模块3部分组成的网络结构;在上下文提取模块中,通过使用空洞卷积和金字塔池化模块克服Unet较少利用上下文信息及不同感受野下的全局信息的缺点,通过19例患者的MRI数据,包括14例正常女性盆底MRI影像、1例盆腔脏器脱垂1度(pelvic organ prolapse degree 1,POP1)患者和2例盆腔脏器脱垂2度(POP2)患者作为训练集,并使用1例正常女性盆底MRI影像和1例POP2女性盆底MRI影像进行验证。结果构建模型能够自动、有效地分割盆底MRI图像中的肛提肌,通过验证,测试集中肛提肌总的平均相似性系数值为77.1%,平均豪斯多夫距离值为16 mm,平均对称面距离值为0.9 mm。其中正常志愿者肛提肌的平均相似性系数值为81.2%,POP2肛提肌的平均相似性系数值为74.5%。结论构建的DenseUnet模型分割精度优于Unet、ResUnet和Unet++3个经典的网络模型,在MRI图像下肛提肌的自动分割任务中具有较强的实用价值。

Objective To construct a deep learning automatic segmentation model based on the magnetic resonance image(MRI) of the pelvic floor, and make the intelligent segmentation of the pelvic floor MR image so as to reduce the work intensity of doctors and improve the segmentation efficiency and accuracy of levator ani muscle. Methods Based on DenseUnet model, a network structure composed of encoder module, context extraction module and decoder module was established;In the context extraction module, we used dilated convolution and pyramid pooling module to overcome the disadvantages that Unet uses less context information and global information under different receptive fields. We employed the MRI data of 19 patients, including 14 normal cases, 1 case of grade 1 pelvic organ prolapse(POP1) and 2 cases of grade 2 pelvic organ prolapse(POP2) as training sets. One normal pelvic floor MRI image and 1 POP2 pelvic floor MR image were used for verification. Results The model can segment levator ani muscle in pelvic floor MR image automatically and effectively. Through verification, the average similarity coefficient of levator ani muscle in the test set is 77.1%, the average Hausdorff distance is 16 mm, and the average symmetry plane distance is 0.9 mm. The average similarity coefficient of levator ani muscle in normal volunteers was 81.2%, and that of POP2 female pelvic floor levator ani muscle was 74.5%. Conclusion The segmentation accuracy of DenseUnet model is better than that of Unet, ResUnet and Unet++. It has a strong practical value in the automatic segmentation task of levator ani muscle in MRI images. Through the automatic segmentation of levator ani muscle, the repetitive work of doctors is reduced, and the work efficiency is improved. At the same time, it also provides an alternative for the intelligent auxiliary diagnosis and treatment of pelvic organ prolapse.