基于卷积神经网络的直肠癌靶区及危及器官自动勾画

Automatic delineation of rectal cancer target volume and organs at risk based on convolutional neural network

目的实现直肠癌靶区和正常组织的自动勾画,提高临床工作效率。方法采用基于卷积神经网络的深度学习方法,架构神经网络,学习并实现自动勾画,比较自动勾画与人工勾画的差异。结果210例直肠癌患者随机分组为190例训练集,20例验证集。测量单个患者完整勾画耗时约10s,CTV的平均Dice为0.87±0.04,其余正常组织的平均Dice均>0.8,CTV的HD指数为25.33±16.05,MDA指数为3.07±1.49,JSC指数为0.77±0.07。结论使用基于全卷积神经网络的深度学习方法可以实现直肠癌靶区的自动勾画,提高工作效率。

Objective To realize automatic delineation of rectal cancer target volume and normal tissues and improve clinical work efficiency.Methods The deep learning method based on convolutional neural network was adopted to construct neural network,learn and realize automatic delineation,and compare the differences between automatic delineation and manual delineation.Results Two hundred and ten cases with rectal cancer were randomly assigned to a training set of 190 and a validation set of 20.The complete delineation of a single case took about 10s;the average Dice of CTV was 0.87±0.04;the average Dice of other normal tissues was bigger than 0.8;the Hausdorff distance(HD)index of CTV was 25.33±16.05;the mean distance to agreement(MDA)index was 3.07±1.49,and the Jaccard similarity coefficient(JSC)index was 0.77±0.07.Conclusion The deep learning method based on full convolutional neural network can realize the automatic delineation of rectal cancer target volume and improve work efficiency.