基于高光谱图像的淋巴瘤血细胞属性特征提取与自动识别研究

Automatic recognition of lymphoma blood cells based on the peripheral blood smear is the quickest way to screen for the disease, and is helpful to identify problems early in the clinic. For conventional technologies such as digital image analysis, it is difficult to collect all types of blood cell image samples and fundamentally realize the abnormal blood cell recognition due to the difference, the morphological diversity and the uncertainty of blood cell images. Compared with digital images, hyperspectral imaging technology with the high-resolution spectral information can break through the existing automatic recognition methods which simulate the manual microscopic examination and can radically improve the recognition accuracy of abnormal blood cells. However, hyperspectral images have high dimensionality, the spectral overlap and weak signal, and are influenced by multiple interference factors. The urgent problem to be solved at present is how to extract the features which reflect the essential properties of the recognition target and build an automatic recognition model. Based on previous studies on blood cell classification and recognition and hyperspectral image analysis, we firstly extract the property features of lymphoma cells from the cell level and the sub-cell level respectively, analyze the effectiveness of extracted features by using single-sample model and multi-sample mixed model in multiple aspects, and establish the dimensionality reduction and feature extraction method. Then, the reference spectra and multiple constraints are introduced to optimize the spectral unmixing method, which can improve the problems such as the determination of the number of pure spectra and the lack of a fixed coordinate system. The proposed method can realize the automatic feature extraction and enhance the reliability of the extracted characteristic spectra. Finally, the influence of the interference target feature on the model accuracy is analyzed, and we further work on improving the recognition accuracy by incorporating other features. Based on these, the automatic recognition model of lymphoma blood cells with a high precision is built. The implementation of the project is helpful to expand our hyperspectral image database of blood smears. The above-mentioned research results can promote the clinical application of rapid lymphoma screening and play a very important role in perfecting the theoretical and methodological system of automatic blood cell recognition based on hyperspectral imaging. Moreover, the proposed methods may also be used as a reference for other pathological image analysis based on hyperspectral imaging.

基于外周血涂片的淋巴瘤血细胞自动识别能够最为快速地实现该疾病的筛查，有助于临床早期发现问题。面对血细胞图像的差异性、形态多样性与不确定性，构建覆盖各种类型的样本非常困难，采用血细胞数字图像分析等常规技术难以从根本上解决淋巴瘤血细胞类似的异常血细胞的识别问题。高光谱成像技术相比传统数字图像增加了高分辨光谱信息，能够突破模拟人工镜检的自动识别方法，从根本上改善异常细胞识别的精度。然而，高光谱图像面临着维数灾难、光谱重叠、信号微弱以及多种干扰因素的影响，如何提取反映目标本质属性且有利于血细胞识别的特征，建立淋巴瘤血细胞自动识别模型是目前迫切需要解决的问题。申请人在前期血细胞分类识别以及高光谱图像分析等研究的基础上，从细胞和亚细胞两个层次提取淋巴瘤血细胞的特征，并基于单样本模型与多样本混合模型多角度分析提取特征的有效性，建立血细胞高光谱图像降维与特征提取方法；在此基础上，通过参考光谱与多约束条件的引入优化光谱解混方法，改善光谱解混中纯净光谱个数确定、缺乏固定的坐标系等问题，实现反映目标本质属性且有利于血细胞分类识别的特征光谱自动提取，提高特征光谱的可靠性；进一步深入分析其他干扰目标特征的引入对模型精度的影响以及如何通过融入其他特征提高识别精度，建立高精度淋巴瘤血细胞自动识别模型。通过项目的实施，进一步完善血细胞涂片的高光谱图像数据库，所提出的方法与建立的模型能够促进基于外周血细胞的淋巴瘤快速筛查、初诊方法的临床应用。项目的研究成果对构建基于高光谱成像的血细胞自动识别的理论与方法体系具有重要的作用。此外，项目所提出的方法对其他病理组织切片高光谱成像分析同样具有参考与借鉴作用。

基于外周血涂片的淋巴瘤血细胞自动识别能够最快速地实现该疾病的筛查，有助于临床早期发现问题。高光谱成像技术相比传统数字图像增加了高分辨光谱信息，能够突破模拟人工镜检的自动识别方法，从根本上改善异常细胞识别的精度。然而，如何实现高光谱图像的降维，提取有利于血细胞识别的特征是目前迫切需要解决的问题。申请人在前期血细胞分类识别以及高光谱图像分析等研究的基础上，开展了以下工作：(1)建立了基于多元曲线分辨-交替最小二乘法的高光谱图像降维与特征提取方法，实验结果验证了该方法提取血细胞特征的有效性；(2)通过引入参考光谱矩阵优化特征提取方法迭代过程，提高目标光谱特征的准确度，实现目标特征的自动提取；(3)提出基于层次化注意力机制的三维卷积神经网络的分类方法实现血涂片高光谱图像中白细胞的分类，分析了光谱特征与分类性能之间的相关性，为后续异常血细胞的自动识别打下基础；(4)基于淋巴瘤患者血液涂片数字图像数据集，建立了基于Faster R-CNN的淋巴瘤血细胞检测模型；(5)基于高光谱图像数据集，将空间特征与光谱特征结合起来，建立了正常和异常白细胞的高精度识别模型，有助于实现淋巴瘤早期诊断。项目研究成果可促进淋巴瘤快速筛查的临床应用，完善基于高光谱成像的血细胞自动识别的理论与方法体系。