阿尔兹海默症(Alzheimer's Disease,AD)是一种患病率很高的神经退行性疾病,严重影响老年人生活.磁共振成像(Magnetic Resonance Imaging,MRI)能够无创获取大脑的形态结构揭示脑部的病理改变,是目前 AD诊断的主要手段.深度学习在图像处理上具有强大的特征提取和建模能力,使用深度学习方法处理 MRI进行 AD的自动诊断具有巨大的应用价值.对于三维的脑图像,病变的大小和位置具有随机性和关联性,局部细节特征和全局的长程依赖信息都很重要.针对此类问题,提出了一种基于注意力的结合 3D CNN和 Transformer的端到端网络用以分类 AD病人和正常人.采用 3D CNN提取深层语义特征图,经多尺度特征加权的注意力编码后由 Transformer全局建模得到分类结果.在 AD数据集和公开的 3D医学分类数据集上验证,分类指标准确性、敏感性和特异性均有所提升.在 AD分类任务上准确性达到 95%,模型的注意力图突出了额叶、后扣带皮质等疾病相关区域.结果显示该方法具有较好的分类性能,可以作为一种自动、有效、便捷的 AD及其他医学任务辅助诊断方法.
Classification of Alzheimer's Disease Based on Attention and Transformer
Alzheimer's Disease(AD)is a neurodegenerative disease with high prevalence,which seriously affects the life of the elderly.Magnetic Resonance Imaging(MRI)can non-invasively obtain the morphological structure of the brain and reveal the pathological changes of the brain,which is currently the main means of AD diagnosis.Deep learning has powerful feature extraction and modeling capabilities in image processing,and the use of deep learning methods to process MRI for automatic diagnosis of AD has great application value.For three-dimensional brain images,the size and location of lesions are random and correlated,and local detailed features and global long-range dependency information are important.An attention based end-to-end network combining 3D CNN and Transformer is proposed to classify AD patients and normal individuals in response to such issues.Firstly,3D CNN is used to extract deep semantic feature-maps,which are then subjected to multi-scale feature weighted attention encoding and globally modeled by Transformer to obtain classification results.The method is validated on the AD dataset and publicly available 3D medical classification datasets.It is shown that the accuracy,sensitivity,and specificity are improved.The accuracy on the AD classification task reaches 95%,and the attention maps of the model highlight the disease-related areas such as the frontal lobe and the posterior cingulate cortex.The results show that the method has good classification performance and can be used as an automatic,effective,and convenient method for auxiliary diagnosis of AD and other medical tasks.
国家科技期刊平台
NSTL
万方数据
维普
