基于影像组学的脑脊液细胞分类方法

Radiomics-based cerebrospinal fluid cell classification

目的:基于影像组学构建出更为高效、准确的脑脊液细胞判别模型。方法:回顾性收集3331张脑脊液细胞显微图像,其中吞噬细胞167张、单核细胞332张、淋巴细胞1081张、中性粒细胞1751张。首先在显微图像上分割出细胞核、细胞核凸包区域和细胞核凸包区域的部分细胞质,然后设计3种细胞核形状特征,即圆度、凸度、坚固性。针对细胞核、凸包区域和凸包区域的部分细胞质设计48种颜色特征。基于细胞核凸包区域提取4 676种纹理特征。结果:共提取了4 727个影像组学特征,在经过ANOVA和LASSO特征选择之后,保留了519个特征,且形状特征和颜色特征都得到了较高比例的保留(100.0%, 66.7%)。特征选择之后,利用SMOTE数据增强和SVM分类器在测试集上进行预测,各项评价指标Accuracy、Sensitivity、Specificity、Precision、F1_score、AUC高达0.953、0.948、0.990、0.961、0.955、0.996。结论:本文提出的新的细胞显微图像特征提取方案和分类模型对细胞分类问题非常有效,且避免了细胞质分割的难题,无需分割细胞,只需分割细胞核和细胞核凸包区域就可以得到较好的分类结果。

Objective To build a more efficient and accurate cerebrospinal fluid cell discrimination model based on radiomics.Methods A total of 3 331 microscopic images of cerebrospinal fluid cells were retrospectively collected, including 167 of phagocytes, 332 of monocytes, 1 081 of lymphocytes, and 1 751 of neutrophils. After segmenting the nucleus, nucleus convex hull and some cytoplasm of the nucleus convex hull in the microscopic images, 3 kinds of nuclear shape features, including roundness, convexity and firmness, 48 color features of the nucleus, nucleus convex hull and some cytoplasm of the nucleus convex hull, and 4 676 texture features of nucleus convex hull were extracted. Results A total of 4 727 radiomic features were obtained. After ANOVA and LASSO feature selection, only 519 features were retained, and both shape features and color features were retained in a high proportion(100.0%, 66.7%). After feature selection, SMOTE data enhancement and SVM classifier were used for the prediction on the test set. The results showed that the accuracy, sensitivity, specificity, precision,F1_score, and AUC were as high as 0.953, 0.948, 0.990, 0.961, 0.955, and 0.996, respectively. Conclusion The proposed feature extraction scheme and classification model for cell microscopic image is effective for cell classification, and avoid the cytoplasm segmentation. It is not necessary to segment cells, but only to segment the nucleus and the nucleus convex hull to obtain better classification results.