Objective To propose a novel blood cell classification framework based on spectral-spatial information extracted from medical hyperspectral image for performing blood cell classification and counting and improving classification accuracy and identification of abnormal cells with the unique features ofhyperspectral image.Methods The hyperspectral images of blood cells were obtained with a microscope and a hyperspectral camera to make sure the data had high spatial and spectral resolution.Dozens of successive narrow wavelength bands were included in the hyperspectral data,which showed the detailed spectral information about different substances.For blood cell classification,band selection was firstly applied to preserve the most informative bands,in which the useful spatial information was represented with Gabor filter.And then,several state-of-the-art pixel-based classifiers,such as sparse representation-based classification,support vector machine,and kernal-based extreme learning machine,were used to verify the extracted spectral-spatial features.Results The proposed classification framework fully utilizd the spatial information surrounding locations where pixels tend to be the same class,and the performance of the proposed classification framework for blood cell classification and counting was validated with medical hyperspectral data.Conclusion Even when the number of training samples varies,the proposed framework can achieve a significantly higher classification accuracy than these conventional pixel-wise solely with spectral information classifiers.%目的:提出一种新的从医学高光谱成像中提取基于光谱空间信息的血细胞分类框架.给出一个使用高光谱成像的血细胞分类与计数方法,能以其独特的特征改善分类精度和异常细胞的识别.方法:使用显微高光谱摄像仪采集高光谱血细胞图像,能够同时保证采集数据的高空间和光谱分辨率.高光谱数据中包含了几十个连续的窄波段,这样能够显示不同物质的光谱细节信息差异.对于血细胞分类,波段选择首先采用保存信息量最丰富的波段,然后用Gabor滤波器表示有用的空间信息.最后,使用一些先进的基于像素的分类器,例如稀疏表示分类器、支持向量机、核函数极限学习机,去验证所提取光谱空间特征.结果:本文提出的分类框架,充分利用了极有可能是同类像素的空间邻域信息,这已经经过了医学高光谱数据的验证.结论:实验结果表明此框架在不同训练样本数量的情况下,与传统单纯基于光谱信息分类方法相比,显著提升了分类精度.
展开▼
