卷积神经网络识别正常及异常甲状腺超声图像的价值

The Value of the Convolutional Neural Network in the Identification of Normal and Abnormal Thyroid Ultrasound Images

目的利用卷积神经网络自动识别正常及异常甲状腺超声图。方法使用我院超声医学科2018年1月至2018年10月甲状腺超声图像资料,分为正常甲状腺、甲状腺局限性病变、甲状腺弥漫性病变、甲状腺弥漫合并局限性病变四类进行标注,以Mask R-CNN算法进行训练,在训练过程中加入基于级联网络改进。统计和汇总采用python完成,目标检测及分割采用平均精度及标准交并比评价,分类采用敏感度、特异性、准确度及一致性评价。结果共使用甲状腺图像47206幅,Mask R-CNN及基于级联网络结构的Mask R-CNN算法的平均精确度为79.21%、84.51%,标准交并比为87.78%、89.26%。在基于级联的Mask R-CNN算法中甲状腺超声图像分类效能较高,敏感度、特异度、准确度、一致性检验在正常甲状腺分别为83.98%、93.93%、91.84%、0.76,局限性病变中为85.09%、94.12%、90、36%、0.79,弥漫性病变中为86.00%、97.11%、94.29%、0.84,弥漫合并局限性病变中为82.99%、94.55%、93.16%、0.70。结论基于级联的Mask R-CNN算法对甲状腺超声图像的目标检测及分割能力较高,对于自动识别二维灰阶正常甲状腺、甲状腺局限性疾病、甲状腺弥漫性病变、甲状腺弥漫合并局限性疾病有较好的效果。

Objective To identify normal and abnormal thyroid ultrasound images by convolutional neural network automatically.Methods Analysis was performed on thyroid ultrasound data of ultrasound department of our hospital from January 2018 to October 2018.During the labeling process,the images were divided into four categories:normal thyroid,thyroid focal lesion,diffuse thyroid disease,diffuse with focal thyroid disease.Mask R-CNN was used as the main framework,and cascade network was added to improve the model.Python was used for statistics.To assess the performance of algorithm target detection and segmentation,mean average precision(MAP)and standard intersection over union(IOU)were adopted.Sensitivity,specificity,accuracy and consistency(kappa value)were used to evaluate classification.Results A total of 47,206 thyroid ultrasound images were used in this study.MAP of Mask R-CNN and Cascade Mask R-CN was 79.21%and 84.51%,IOU was 87.78%and 89.26%.The classification efficiency of the four categories of thyroid ultrasound images is excellent in Cascade Mask R-CNN,the sensitivity,specificity,accuracy and consistency of normal thyroid are 83.98%,93.93%,91.84%and 0.76 respectively,for focal thyroid disease were 85.09%、94.12%,90,36%,0.79,for diffuse thyroid disease were 86.00%,97.11%,94.29%and 0.84,for diffuse and focal thyroid disease were 82.99%,94.55%,93.16%and 0.70.Conclusion Cascade Mask R-CNN had higher target detection and segmentation capability in thyroid ultrasound image.It can be used to automatically identify the two-dimensional gray scale of normal thyroid,thyroid focal lesion,diffuse thyroid disease,diffuse with focal thyroid disease.