基于FC_DenseNet深度学习网络自动分割肺癌放疗中的危及器官

目的:建立一种基于密集连接深度学习的端到端胸部CT图像危及器官自动分割方法,提供一个高精度的自动分割模型,减轻医师临床勾画的工作强度。方法:收集36例肺癌患者CT图像,27例作为训练集,随机取6例作为验证集进行交叉验证,测试集为9例,训练时间约为5 h,完成了左肺、右肺、脊髓、心脏4个危及器官的自动分割,并使用Dice系数、HD95距离与平均表面距离(ASD)3个指标对测试集进行测试。结果:测试集的分割结果显示,与U-Net与ResNet50相比,FC_DenseNet网络在Dice值、HD95、ASD指标上表现较好,但是不同网络之间的分割结果并没有显著差异(P>0.05),FC_DenseNet网络Dice值最高是左肺为0.98,最低为心脏0.84。结论:本研究的结果表明,密集连接结构的深度学习模型能够较为准确地分割左右肺、脊髓、心脏4个危及器官,这种特征图复用的思想为基于深度学习的医学图像分割提供了新思路。

FC_Densenet-based autosegmentation of organs-at-risk in lung cancer radiotherapy

Objective To introduce a method for the end-to-end autosegmentation of organs-at-risk in chest CT images based on dense connection deep learning,and to provide a high-precision autosegmentation model for reducing doctors’workload.Methods The CT images of 36 lung cancer patients were analyzed in this study.Twenty-seven cases out of 36 cases were randomly selected as training set,6 cases as validation set for cross validation,and 9 cases as testing set.The training time was about 5 hours,and the segmentations of 4 organs-at-risk including the left and right lungs,spinal cord and heart were completed.The testing set was evaluated by Dice coefficient,HD95 distance and average surface distance.Results Compared with U-Net ResNet50,DenseNet was better in Dice coefficient,HD95 and average surface distance.However,there was no significant difference in segmentation results among 3 networks.The highest Dice coefficient of DenseNet was 0.98 for the left lung,while the lowest was 0.84 for the heart.Conclusion The left and right lungs,spinal cord and heart can be accurately segmented by dense connection deep learning model.The idea of feature map reuse provides a new idea for medical image segmentation based on deep learning.