基于全景病理图像细胞密度和异型特征的胶质瘤自动分级

Automated grading of glioma based on density and atypia analysis in whole slide images

胶质瘤是最常见的恶性脑肿瘤,其高低级别分类是制定治疗方案和预后的重要参考指标。临床中,脑胶质瘤的高低分级诊断通常由病理医生阅读全景病理图像(WSI)来完成,该任务繁琐且对医生经验要求较高。根据2016年第4版《中枢神经系统肿瘤WHO分类》标准,细胞的富集程度、核异型、坏死等现象与胶质瘤分级密切相关。受该标准启发,本文定量分析脑全景病理图像中细胞密度和异型特征,对胶质瘤进行高低级别自动分级。首先分析全局细胞密度定位感兴趣区域(ROI),提取全扫描图像的全局密度特征,然后对感兴趣区域提取局部密度特征和异型特征,最后利用特征选择并构建平衡权重的支持向量机(SVM)分类器,5折交叉验证的受试者工作特性曲线下的面积(AUC)为0.92±0.01,准确率(ACC)为0.82±0.01。实验结果表明,本文提出的感兴趣区域定位方法可快速有效地实现定位,构建的细胞密度和异型特征能够实现胶质瘤的自动分级,为临床诊断提供可靠依据。

Glioma is the most common malignant brain tumor and classification of low grade glioma(LGG) and high grade glioma(HGG) is an important reference of making decisions on patient treatment options and prognosis. This work is largely done manually by pathologist based on an examination of whole slide image(WSI),which is arduous and heavily dependent on doctors’ experience. In the World Health Organization(WHO) criteria,grade of glioma is closely related to hypercellularity, nuclear atypia and necrosis. Inspired by this, this paper designed and extracted cell density and atypia features to classify LGG and HGG. First, regions of interest(ROI) were located by analyzing cell density and global density features were extracted as well. Second, local density and atypia features were extracted in ROI. Third, balanced support vector machine(SVM) classifier was trained and tested using 10 selected features. The area under the curve(AUC) and accuracy(ACC) of 5-fold cross validation were 0.92 ± 0.01 and 0.82 ± 0.01 respectively. The results demonstrate that the proposed method of locating ROI is effective and the designed features of density and atypia can be used to predict glioma grade accurately, which can provide reliable basis for clinical diagnosis.