【摘要】 目的：阿尔茨海默病在临床诊断中因难以量化海马体萎缩程度而难以确诊，耽误了尽早治疗干预。提出采用基于深度学习的三维卷积神经网络对患者的脑核磁影像数据进行像素级语义分割，实现高效的自动化脑区分割，辅助医生临床诊断。方法：数据来自ADNI数据集，使用基于模板算法的自动化分割软件FreeSurfer标注训练数据。使用3D-UNet模型+Generalized Dice损失函数高效地拟合三维影像数据。结果：该模型分割结果相比模板法有更平滑的表面且不易受噪声影响，分割效率提升上百倍。针对分割结果进行了脑区体积的量化计算，并生成图表用于临床对体积和变化趋势的分析。结论：方法减少了人工标注，可以快速部署应用；具有比模板法高的分割效率且不失精度；直观的量化分析辅助临床诊断。更多还原

【Abstract】 Objective It is difficult to diagnose Alzheimer’s Disease in clinical diagnosis because it is difficult to quantify the degree of hippocampal atrophy, which delays the early treatment and intervention. In this paper, a threedimensional convolutional neural network based on deep learning is proposed to segment the patient’s brain MRI data at the pixel level, to achieve efficient automatic brain segmentation and assist doctors in clinical diagnosis. Methods The data in this paper are from ADNI dataset, and the training data are labeled by FreeSurfer, an automatic segmentation software based on template algorithm. 3D-UNet model + generalized dice loss function is used to fit 3D image data efficiently. Results Compared with the template method, the segmentation result of the model has a smoother surface and is not easily affected by noise, and the segmentation efficiency is improved by hundreds of times. According to the segmentation results, the volume of brain area is calculated quantitatively, and charts are generated for clinical analysis of volume and change trend. Conclusion This method reduces manual labeling and can deploy applications quickly; It is more efficient and accurate than the template method; Comprehensible quantitative analysis to assist clinical analysis.更多还原