Department of Computer Science and Engineering, Pondicherry University, Puducherry, India.
Department of Computer Science and Engineering, Community College, Pondicherry University, Puducherry, India.
J Neurosci Methods. 2019 Apr 15;318:84-99. doi: 10.1016/j.jneumeth.2019.01.011. Epub 2019 Feb 5.
As the medical images contain both superficial and imperceptible patterns, textures are successfully used as discriminant features for the detection of cancers, tumors, etc. NEW METHOD: Our algorithm selects the specific image blocks and computes the textures using the following steps: At first, the center image slice of the axes (sagittal, coronal and axial) is divided into small blocks and those which approximately resembles the regions of interest are marked. Then, all the marked blocks which are in the same location as in the center slice are collected from all the other slices, and the textures are computed per block on all the individual slices. The generated textures are then pipelined to a feature selection algorithm with bootstrapping to pick-out features of high relevance and less redundancy and are exhaustively analyzed with multiple feature selection techniques like fisher score, elastic net, recursive feature elimination and classification algorithms like random forest, linear support vector machines, and k-nearest neighbors algorithms.
This method is validated on baseline MR images of 812 subjects from Alzheimer's Disease Neuroimaging Initiative (ADNI) database. The results of binary classifications of different classes of Alzheimer's disease are also analyzed. The proposed features achieve the sensitivity/specificity of 89.58%/85.82% for AD/NC classification.
COMPARISON WITH EXISTING METHOD(S): The proposed textures extraction runs over two times faster than other texture processing methods used for AD classification.
This study identifies the proposed textures with regional atrophies that could be used as potential checkpoints for Alzheimer's disease classification.
由于医学图像既包含明显的模式，也包含难以察觉的模式，因此纹理被成功地用作癌症、肿瘤等检测的判别特征。
我们的算法选择特定的图像块，并使用以下步骤计算纹理：首先，将轴（矢状、冠状和轴）的中心图像切片划分为小的块，并标记出那些大致类似于感兴趣区域的块。然后，从所有其他切片中收集所有位于中心切片相同位置的标记块，并在所有单独的切片上计算每个块的纹理。生成的纹理随后通过带有自举的特征选择算法进行流水线处理，以挑选出具有高度相关性和较少冗余的特征，并通过多种特征选择技术（如 Fisher 得分、弹性网络、递归特征消除和分类算法，如随机森林、线性支持向量机和 k-最近邻算法）进行详尽分析。
该方法在来自阿尔茨海默病神经影像学倡议 (ADNI) 数据库的 812 名受试者的基线 MRI 图像上进行了验证。还分析了不同类别阿尔茨海默病的二分类结果。所提出的特征在 AD/NC 分类中达到了 89.58%/85.82%的灵敏度/特异性。
与用于 AD 分类的其他纹理处理方法相比，所提出的纹理提取速度快两倍。
这项研究确定了与区域性萎缩相关的所提出的纹理，这些纹理可作为阿尔茨海默病分类的潜在检查点。
