基于面部图像的有无早期肺癌风险分类研究  

Study on Risk Classification of Early Lung Cancer Based on Facial Images

作　　者：周孟齐 胡广芹[1] 林岚[1] 李斌[1] 张新峰 ZHOU Mengqi;HU Guangqin;LIN Lan;LI Bin;ZHANG Xinfeng(Faculty of Environment and Life,Beijing University of Technology,Beijing 100124,China;Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学环境与生命学部,北京100124 [2]北京工业大学信息学部,北京100124

出　　处：《中国医疗设备》2022年第11期52-56,84,共6页China Medical Devices

基　　金：国家重点研发计划(2018YFC1707705)。

摘　　要：目的通过对数据集进行目标区域分割、特征提取等操作,建立随机森林模型,以实现有无早期肺癌风险的分类研究。方法使用BiSeNet算法实现图像分割,并将分割后的图像转换到YCbCr颜色空间模型[亮度(Y)、蓝色分量(CB)、红色分量(CR)]上,通过CB与CR 2个分量的取值寻找非肤色点,对非肤色点采用9×9均值滤波器进行滤波,并在该颜色模型下提取颜色特征值,再将图像转换到灰度空间,在其灰度共生矩阵上获取其纹理特征值。将这些特征值作为输入构造随机森林分类模型,构造随机森林时使用ID3算法构造决策树,通过调整决策树个数和最大特征数寻找最优分类模型。结果BiSeNet面部图像分割准确率为96.25%;在YCbCr颜色空间上具有椭圆肤色聚类的特性,可以检测到非肤色点;经调整发现2个超参数决策树个数、最大特征数取值分别为30和4时,随机森林模型性能最优,其准确率能够达到87.34%。结论通过面部的颜色特征以及文理特征信息,可以进行早期肺癌的分类研究,经实验分析肺癌患者面部红色特征以及文理特征与未患肺癌相比,存在显著差异,有助于有无早期肺癌的分类判断,为临床上早期肺癌的发现提供辅助依据。Objective To establish a random forest model based on target region segmentation and feature extraction of the dataset so as to classify the risk of early lung cancer.Methods The BiSeNet algorithm was used to realize image segmentation,and the segmented image was converted to YCbCr color space model.The non-skin-color points were found through the values of CB and CR components,and the non-skin-color points were filtered by 9×9 mean filter.The color feature value was extracted under the color model,then the image was converted to gray space.The texture feature value was obtained on the gray level co-occurrence matrix.These eigenvalues were used as inputs to construct a random forest classification model,and ID3 algorithm was used to construct a decision tree.The optimal classification model was found by adjusting the number of decision tree and the maximum eigenvalues.Results The segmentation accuracy of BiSeNet facial image was 96.25%.In YCbCr color space,it had the feature of elliptical skin color clustering,and non-skin-color points could be detected.After adjusting the two parameters(the number of decision tree and the maximum eigenvalues),it is found that when the values of the two parameters were 30 and 4 respectively,the random forest model had the best performance,and the accuracy rate could reach 87.34%.Conclusion The early lung cancer can be classified according to the facial color features and texture features.The experimental analysis shows that the facial red features and texture features of the patients with lung cancer are significantly different from those of patients without lung cancer,which contribute to realize the classification and judgment of early lung cancer,and provide auxiliary evidence for the clinical discovery of early lung cancer.

关 键 词：面部图像 特征提取 肺癌风险 随机森林 

分 类 号：R734.2[医药卫生—肿瘤] R730.44[医药卫生—临床医学]