Biochemistry and Molecular Genetics, Biomedical Diagnostic Center, Hospital Clinic de Barcelona, Barcelona, Spain
School of Engineering, Science and Technology, Universidad del Rosario Facultad de Ciencias Naturales y Matemáticas, Bogota, Cundinamarca, Colombia.
J Clin Pathol. 2020 Oct;73(10):665-670. doi: 10.1136/jclinpath-2019-206419. Epub 2020 Mar 16.
Morphological recognition of red blood cells infected with malaria parasites is an important task in the laboratory practice. Nowadays, there is a lack of specific automated systems able to differentiate malaria with respect to other red blood cell inclusions. This study aims to develop a machine learning approach able to discriminate parasitised erythrocytes not only from normal, but also from other erythrocyte inclusions, such as Howell-Jolly and Pappenheimer bodies, basophilic stippling as well as platelets overlying red blood cells.
A total of 15 660 erythrocyte images from 87 smears were segmented using histogram thresholding and watershed techniques, which allowed the extraction of 2852 colour and texture features. Dataset was split into a training and assessment sets. Training set was used to develop the whole system, in which several classification approaches were compared with obtain the most accurate recognition. Afterwards, the recognition system was evaluated with the assessment set, performing two steps: (1) classifying each individual cell image to assess the system's recognition ability and (2) analysing whole smears to obtain a malaria infection diagnosis.
The selection of the best classification approach resulted in a final sequential system with an accuracy of 97.7% for the six groups of red blood cell inclusions. The ability of the system to detect patients infected with malaria showed a sensitivity and specificity of 100% and 90%, respectively.
The proposed method achieves a high diagnostic performance in the recognition of red blood cell infected with malaria, along with other frequent erythrocyte inclusions.
识别感染疟原虫的红细胞的形态是实验室实践中的一项重要任务。如今，缺乏能够将疟疾与其他红细胞内含物区分开来的特定自动化系统。本研究旨在开发一种机器学习方法，不仅能够区分寄生虫感染的红细胞，还能够区分其他红细胞内含物，如豪-乔二氏体（Howell-Jolly bodies）和帕彭海默氏体（Pappenheimer bodies）、碱性点彩以及覆盖在红细胞上的血小板。
使用直方图阈值和分水岭技术对来自 87 张涂片的总共 15660 个红细胞图像进行分割，从而提取出 2852 个颜色和纹理特征。数据集被分为训练集和评估集。训练集用于开发整个系统，在该系统中比较了几种分类方法，以获得最准确的识别。然后，使用评估集评估识别系统，执行两个步骤：（1）对每个单个细胞图像进行分类，以评估系统的识别能力；（2）分析整个涂片，以获得疟疾感染的诊断。
最佳分类方法的选择导致最终的顺序系统对红细胞内含物的六组具有 97.7%的准确率。该系统检测感染疟疾的患者的能力显示出 100%的敏感性和 90%的特异性。
该方法在识别感染疟疾的红细胞以及其他常见的红细胞内含物方面具有较高的诊断性能。
