Alférez S, Merino A, Bigorra L, Rodellar J
Matematica Aplicada III, Technical University of Catalonia, Barcelona, Spain.
Department of Hemotherapy-Hemostasis, Hospital Clinic, Barcelona, Spain.
Int J Lab Hematol. 2016 Apr;38(2):209-19. doi: 10.1111/ijlh.12473.
The objective was to advance in the automatic, image-based, characterization and recognition of a heterogeneous set of lymphoid cells from peripheral blood, including normal, reactive, and five groups of abnormal lymphocytes: hairy cells, mantle cells, follicular lymphoma, chronic lymphocytic leukemia, and prolymphocytes.
A number of 4389 images from 105 patients were selected by pathologists, based on morphologic visual appearance, from patients whose diagnosis was confirmed by all the remaining complementary tests. Besides geometry, new color and texture features were extracted using six alternative color spaces to obtain rich information to characterize the cell groups. The recognition system was designed using support vector machines trained with the whole image set.
In the experimental tests, individual sets of images from 21 new patients were analyzed by the trained recognition system and compared with the true diagnosis. An overall recognition accuracy of 97.67% was achieved when the cell screening was performed into three groups: normal lymphocytes, abnormal lymphoid cells, and reactive lymphocytes. The accuracy of the whole experimental study was 91.23% when considering the further discrimination of the abnormal lymphoid cells into the specific five groups.
The excellent automatic screening of the three groups of normal, reactive, and abnormal lymphocytes is useful as it discriminates between malignancy and not malignancy. The discrimination of the five groups of abnormal lymphoid cells is encouraging toward the idea that the system could be an automated image-based screening method to identify blood involvement by a variety of B lymphomas.
目的是推进对外周血中一组异质性淋巴细胞的基于图像的自动表征和识别，这些淋巴细胞包括正常、反应性以及五组异常淋巴细胞：毛细胞、套细胞、滤泡性淋巴瘤、慢性淋巴细胞白血病和幼淋巴细胞。
病理学家根据形态视觉外观，从105名患者中挑选出4389张图像，这些患者的诊断通过所有其余补充检查得以确认。除了几何特征外，还使用六种替代颜色空间提取新的颜色和纹理特征，以获取丰富信息来表征细胞组。识别系统采用支持向量机进行设计，并使用整个图像集进行训练。
在实验测试中，训练好的识别系统对21名新患者的单独图像集进行了分析，并与真实诊断结果进行了比较。当将细胞筛选分为三组：正常淋巴细胞、异常淋巴细胞和反应性淋巴细胞时，总体识别准确率达到了97.67%。当进一步将异常淋巴细胞细分为特定的五组时，整个实验研究的准确率为91.23%。
对正常、反应性和异常淋巴细胞这三组进行出色的自动筛选很有用，因为它能区分恶性和非恶性。对五组异常淋巴细胞的区分令人鼓舞，表明该系统有望成为一种基于图像的自动筛选方法，用于识别各种B淋巴瘤的血液受累情况。
