基于深度学习的自动分割方法对头颈部危及器官勾画的评价

Evaluation of deep learning-based automatic segmentation on the delineation for head and neck OAR

目的:探讨与评价AccuContour软件基于深度学习的自动分割方法对头颈部危及器官(OAR)自动勾画结果的准确性。方法:选取20例鼻咽癌患者的CT图像资料,应用基于深度学习的机器算法模型AccuContour软件,对图像的19个OARs(脑干、脊髓、双侧眼球、双侧晶体、双侧视神经、垂体、双侧腮腺、口腔、下颌骨、双侧颌下腺、双侧颞颌关节、甲状腺和喉)进行自动勾画,勾画结果定义为深度学习勾画的体积(V_(DL)),对其进行手动修改后,将修改结果定义为参考勾画(V_(ref)),采用相似性系数(DSC)、Jaccard系数(JAC)、平均Hausdorff距离(AHD)、体积差异比(ΔV)、最大剂量(D_(max))或平均剂量(Dmean)评价勾画结果。结果:除垂体、双侧颞颌关节及双侧视神经外,其他器官的DSC平均值均>0.7,JAC平均值均>0.6;各器官AHD均值最大为1.91 mm,最小为0.23 mm。双侧眼球、双侧晶体、下颌骨、双侧颌下腺、喉和甲状腺在剂量评估中表现优异,以ΔD_(max)评估的器官中,绝对平均值最大为5.81 Gy,以Dmean评估的器官中,绝对平均值最大为3.77 Gy。结论:AccuContour软件基于深度学习的自动分割方法在头颈部OAR自动勾画中能够获得较好的结果,可以用来提高临床工作效率。

Objective:To investigate and evaluate the deep learning-based automatic segmentation of AccuContour software on the delineation for organs at risk(OARs)of head and neck.Methods:The computed tomography(CT)images of 20 patients with nasopharyngeal carcinoma were selected.A deep learning-based machine algorithm model(AccuContour,Manteia)was applied to conduct automatic delineation for 19 OARs(brain stem,spinal cord,bilateral eyeballs,bilateral lenses,bilateral optic nerves,pituitary,bilateral parotid glands,oral cavity,mandible,bilateral submandibular glands,bilateral temporomandibular joints,thyroid and larynx).The delineation results were defined as automatic delineation(V_(DL)),and they were defined as referred delineation(V_(ref))after they were manually modified.The dice similarity coefficient(DSC),Jaccard coefficient(JAC),average Hausdorff distance(AHD),diversity ration of volume(ΔV),the _(max)imum dose(D_(max))or mean dose(Dmean)were adopted to evaluate the results of delineation.Results:Except for pituitary,bilateral temporomandibular joints and bilateral optic nerves,the average DSC of other organs was greater than 0.7,and the average JAC was more than 0.6.The _(max)imum value of AHD mean of each organ was 1.91mm,and the minimum value was 0.23mm.The appearances of the bilateral eyeballs,bilateral lenses,mandible,bilateral submandibular glands,larynx and thyroid performed excellent in dose assessment.In the organs that were assessed byΔD_(max),the _(max)imum absolute mean was 5.81 Gy.In the organs that were assessed byΔDmean,the _(max)imum absolute mean was 3.77 Gy.Conclusion:The deep learning-based automatic segmentation of AccuContour software can obtain better results in the automatic delineation for head and neck OAR,which can is used to improve clinical work efficiency.