Chosun University, Department of Computer Engineering, 309 Pilmun-daero, Dong-gu, Gwangju 501-759, Republic of KoreabChosun University, Center for Holographic Imaging Informatics, 309 Pilmun-daero, Dong-gu, Gwangju 501-759, Republic of Korea.
J Biomed Opt. 2016 Dec 1;21(12):126015. doi: 10.1117/1.JBO.21.12.126015.
The classification of erythrocytes plays an important role in the field of hematological diagnosis, specifically blood disorders. Since the biconcave shape of red blood cell (RBC) is altered during the different stages of hematological disorders, we believe that the three-dimensional (3-D) morphological features of erythrocyte provide better classification results than conventional two-dimensional (2-D) features. Therefore, we introduce a set of 3-D features related to the morphological and chemical properties of RBC profile and try to evaluate the discrimination power of these features against 2-D features with a neural network classifier. The 3-D features include erythrocyte surface area, volume, average cell thickness, sphericity index, sphericity coefficient and functionality factor, MCH and MCHSD, and two newly introduced features extracted from the ring section of RBC at the single-cell level. In contrast, the 2-D features are RBC projected surface area, perimeter, radius, elongation, and projected surface area to perimeter ratio. All features are obtained from images visualized by off-axis digital holographic microscopy with a numerical reconstruction algorithm, and four categories of biconcave (doughnut shape), flat-disc, stomatocyte, and echinospherocyte RBCs are interested. Our experimental results demonstrate that the 3-D features can be more useful in RBC classification than the 2-D features. Finally, we choose the best feature set of the 2-D and 3-D features by sequential forward feature selection technique, which yields better discrimination results. We believe that the final feature set evaluated with a neural network classification strategy can improve the RBC classification accuracy.
红细胞分类在血液学诊断领域，特别是血液疾病诊断中起着重要作用。由于在血液疾病的不同阶段，红细胞（RBC）的双凹形形状发生改变，我们认为红细胞的三维（3-D）形态特征比传统的二维（2-D）特征提供更好的分类结果。因此，我们引入了一组与 RBC 轮廓的形态和化学性质相关的 3-D 特征，并尝试使用神经网络分类器评估这些特征对 2-D 特征的区分能力。3-D 特征包括红细胞表面积、体积、平均细胞厚度、球形指数、球形系数和功能因子、MCH 和 MCHSD，以及从单细胞水平的 RBC 环截面提取的两个新特征。相比之下，2-D 特征是 RBC 投影表面积、周长、半径、伸长率和投影表面积与周长比。所有特征均从使用离轴数字全息显微镜和数值重建算法可视化的图像中获得，感兴趣的 RBC 类别包括双凹（甜甜圈形状）、平碟形、口形细胞和棘形细胞。我们的实验结果表明，与 2-D 特征相比，3-D 特征在 RBC 分类中更有用。最后，我们通过顺序前向特征选择技术选择 2-D 和 3-D 特征的最佳特征集，这可以产生更好的区分结果。我们相信，使用神经网络分类策略评估的最终特征集可以提高 RBC 分类的准确性。
