【摘要】 目的基于深度学习方法提出一种稠残U-net神经网络,探讨其在放疗定位CT上自动预测甲状腺轮廓的可行性,以减少放疗中甲状腺所受辐射剂量,降低甲减发生率。方法在U-net网络中引入残差机制和稠密连接机制建立一种稠残U-net网络。选取76名患者定位CT图像的甲状腺切片制作数据集,随机划分为训练集58例、验证集9例和测试集9例,对稠残U-net进行训练、验证和测试,得到稠残U-net自动预测甲状腺的结果。通过戴斯相似性系数(Dice)、杰卡德相似系数(Jaccard))和豪斯多夫距离(HD)等评价指标来评估其分割性能。结果稠残U-net预测甲状腺的Dice值为0.86±0.09、Jaccard值为0.78±0.12、HD值为2.52±0.61,且预测的轮廓边界与专家勾画的标准边界非常接近。结论本文提出的稠残U-net能在定位CT图像上较为精准地预测甲状腺轮廓,且证明在卷积神经网络中引入残差机制和稠密连接机制能提高其分割性能。更多还原

【Abstract】 Objective Based on the deep learning method, a dense-residual U-net neural network is proposed to explore the feasibility of automatically predicting the contour of the thyroid on the CT images of radiotherapy positioning, so as to reduce the radiation dose to the thyroid during radiotherapy and reduce the incidence of hypothyroidism.Methods Residual mechanism and dense connection mechanism were introduced into U-net to establish dense-residual U-net neural network.The thyroid tissue structures of 76 patients were selected from localized CT images to make a data set, which were randomly divided into a training set of 58 patients, a verification set of 9 patients and a test set of 9 patients.The dense-residual U-net neural network was trained, verified and tested, and the thyroid contour predicted by neural network were obtained.The segment performance of dense-residual U-net were evaluated by dice, jaccard and HD.Results The thyroid’s dice was 0.86±0.09,the jaccard was 0.78±0.12,and the HD was 2.52±0.61 by dense-residual U-net, and the predicted contour boundary was very close to the standard drawn by experts.Conclusions The dense-residual U-net proposed in this study can accurately predict the thyroid contour on localized CT images, and it is proved that the residual mechanism and dense connection mechanism can improve performance of convolutional neural network in medical image segmentation.更多还原