Development of manifold learning based texture analysis for lesions within noise included medical images

实验研究了基于流形学习的医学图像损伤检测方法。各种医学图像普遍存在噪声，因为成像条件和参数是固定的，以减轻患者的负担。然而，一些实验验证表明，包含噪声的医学图像对基于体素的纹理特征空间的低维流形学习和基于纹理特征的体素分类效果有限。此外，通过GGO体素分类和GGO结节区域检测实验，验证了基于流形的一类分类器和两类分类器级联分类的优越性。

Manifold learning based image analyses for detecting lesion in medical images were studied experimentally. Various medical images are noisy in common, because imaging conditions and parameters are fixed for reduction of patients’ burdens. However some experimental validation showed the noise included medical images has only a limited effect on learning low-dimensional manifold of voxel-wise texture feature space and voxel classification by the texture features. In addition, the benefit of the cascade classification by a manifold based one class classifier and a two class classifier was shown through experiments of the GGO voxel classification and the GGO nodule region detection.