【摘要】 针对医学伦理和人工标注的成本高昂,目前公开的数据集中标记的医学图像数据量不足的问题,许多研究者提出了不同的算法来增强医学数据.使用卷积神经网络(Convolutional Neural Networks, CNN)对阿尔茨海默症(Alzheimer’s Disease, AD)的诊断和研究只有在数据丰富时,才能得到更好的分析结果.本文提出一种基于形状增强和纹理插值的三维脑部MRI (Magnetic Resonance Imaging, MRI)数据增强算法,将脑部MRI图像的属性分解为形状和纹理,先通过GAN (Generative Adversarial Nets,GAN)增强MRI数据的形状,再通过三维薄板样条插值对所得的增强形状进行纹理插值,得到三维脑部MRI增强图像.通过实验可得生成数据的分布与真实数据的分布相似,验证了本文提出的数据增强算法的有效性.更多还原

【Abstract】 In response to the high cost of medical ethics and manual annotation, as well as the insufficient amount of labeled medical image data in publicly available datasets, many researchers have proposed different algorithms to enhance medical data. The diagnosis and research of Alzheimer’s Disease using Convolutional Neural Networks can only obtain better analysis results when the data is rich.This paper proposes a 3D brain MRI(Magnetic Resonance Imaging, MRI) data enhancement algorithm based on shape enhancement and texture interpolation, which decomposes the attributes of the brain MRI image into shape and texture. The shape of the MRI data is enhanced using GAN(Generative Adversarial Nets, GAN). The texture, performed on the enhanced shape, is obtained through 3D thin plate spline interpolation to obtain a 3D brain MRI enhanced image. The distribution of generated data is similar to that of real data through experiments, which verifies the effectiveness of the data augmentation algorithm proposed in this paper.更多还原