无先兆偏头痛3D-CNN辅助诊断算法

3D-CNN Auxiliary Diagnosis Algorithm for Migraine Without Aura

偏头痛是一种严重危害人类健康的脑疾病,其中无先兆偏头痛在临床中占比最多且诊断困难。当前无先兆偏头痛辅助诊断算法研究中,基于机器学习的脑影像功能连接分析方法是最主要的研究方向。由于此类方法多依赖于预定义的脑图谱模板,受模板选择主观因素及分类器性能影响,现有方法的智能化程度和准确率较低,难以满足临床及研究需求。基于设计的新型3D-CNN技术,提出了一种无先兆偏头痛智能辅助诊断算法MwoA3D-Net(3D convolutional neural network based diagnosis of migraine without aura)。该算法采用组信息指导的独立成分分析方法,生成被试的静息态脑网络,并以此作为输入训练MwoA3D-Net,实现对无先兆偏头痛患者与健康对照的自动诊断,可避免因先验模板不同导致的结果差异。在算法设计中引入3D数据增强、L1和L2正则化等一系列优化策略,可有效防止过拟合现象的发生。在60名无先兆偏头痛和65名健康被试数据集上的实验结果表明,MwoA3DNet的平均诊断准确率为98.40%,鲁棒性较高,且所选静息态脑功能网络均具有较强的辨识性,可作为无先兆偏头痛的潜在生物标志物用于个体化诊断。

Migraine is a brain disease that can seriously endanger human health.Migraine without aura is the most popular disease of migraine and is difficult to diagnose in the clinic.At present,the auxiliary diagnosis of migraine without aura based on functional connectivity analysis in brain images combined with machine learning algorithms is the most important research domain.However,those existing auxiliary diagnosis methods are hard to meet the clinical and research needs.Because the functional connectivity analysis mostly relies on the prior template,which is easily affected by subjective factors and the performance of the classifier,the intelligence and accuracy are still at a low level.To address this issue,an intelligent auxiliary diagnosis algorithm for migraine without aura based on improved 3 D convolutional neural network dubbed MwoA3 D-Net is proposed in this paper.To avoid the difference results caused by varying prior templates,a group information guided independent component analysis method is employed to obtain the resting-statebrain network for training the MwoA3 D-Net algorithm.Subsequently,the MwoA3 D-Net algorithm is applied to diagnose migraine without aura patients and healthy controls automatically.A series of strategies such as 3 D data augmentation and L1 and L2 regularization in the algorithm are introduced to effectively prevent the occurrence of overfitting.Experimental results on a dataset of 60 patients and 65 healthy controls show that MwoA3 D-Net has a highly robust performance,with an average diagnostic accuracy of 98.40%.Furthermore,the selected resting-state brain network has powerful identification and can be adopted as potential biomarkers of migraine without aura toward individualized diagnosis.