【作者】 罗丹；

【导师】 汤炜；

【作者基本信息】 四川大学 ， 口腔医学（专业学位）， 2022， 硕士

【摘要】 目的:针对目前颌骨囊性病变CBCT标注数据集缺乏和手动分割CBCT图像耗时耗力的问题,本研究建立带有标签的颌骨囊性病变术前和术后CBCT标注数据集,在此基础上采用深度学习方法建立颌骨囊性病变术前CBCT图像自动分割模型和术后CBCT图像自动分割模型,实现术前囊性病变的快速定位和精确测量、术后剩余囊腔和成骨区域的自动识别和术后疗效的实时评估。方法:1.建立颌骨囊性病变CBCT标注数据集首先基于四川大学华西口腔医院病理科的病理诊断,收集诊断为颌骨囊性病变患者的临床信息、术前和术后的CBCT数据。然后确定标注标签,术前的标签为背景(0)和囊性病变区域(cyst area,CA,1),术后的标签为背景(0)、剩余囊腔区域(residual cyst area,RCA,1)和成骨区域(osteogenesis area,OA,2)。最后将术前和术后的CBCT数据分别导入3D Slicer标注系统中进行标注,标签数据和原始CBCT数据以nrrd格式保存,完成带有标签的颌骨囊性病变术前CBCT标注数据集和术后CBCT标注数据集的建立。为了建立规范化的标注流程和获得高质量的标签,所有CBCT图像由一位有3年以上临床工作经验的口腔专科医生进行标注,并且由一位经验丰富的专家(>25年临床工作和科研经历)进行检查和纠正,获得最终的标注数据集,并以此为“金标准”用于训练和评估模型。2.建立基于深度学习的颌骨囊性病变术前CBCT图像自动分割模型本研究选择nnU-Net模型用于分割训练,同时采用五折交叉验证对颌骨囊性病变术前CBCT标注数据集进行分配,建立可自动分割和定位颌骨囊性病变的口腔颌面部医学图像处理模型。本研究采用的模型的具体结构按照输入的维度大小可以分为两种,一种是对CBCT数据进行逐层(slice)处理的nnU-Net 2D模型(输入数据尺寸为H*W的矩阵),另一种是对CBCT数据进行三维块(patch)处理的nnU-Net 3D模型(输入数据尺寸为H*W*D的矩阵)。采用Dice相似系数(DSC)、Jaccard指数、平均表面距离、表面重叠度指标来评估模型的分割性能,为了进一步测试模型的性能,从术前CBCT标注数据集中随机选取10例CBCT数据让另外三名高年资口腔医生进行手动分割,计算高年资口腔医生手动分割的DSC、Jaccard指数、平均表面距离、表面重叠度并取他们的平均值,然后将模型自动分割结果与高年资口腔医生的手动分割结果进行统计学分析。3.建立基于深度学习的颌骨囊性病变术后CBCT图像自动分割模型术后CBCT图像分割研究依然采用nnU-Net 2D模型和nnU-Net 3D模型及五折交叉验证,但术后模型建立的细节和术前模型不同,主要区别在于:(1)术前的模型的类别数目为2,术后模型的类别数目为3,因此两个模型最后的分割层的通道数目不一致。(2)术前模型采用随机初始化,而术后模型是采用继承实验二中术前模型参数的方式开始分割训练。最终建立可自动分割和定位剩余囊腔和成骨区域的口腔颌面部医学图像处理模型。客观评价指标和模型性能评估方法与实验二的一致。结果:1.本研究共收集了113例颌骨囊性病变患者的临床信息、75例术前CBCT数据和58例术后CBCT数据,采用3D Slicer软件成功建立了带有标签的颌骨囊性病变术前CBCT标注数据集和术后CBCT标注数据集,其中术前CBCT标注数据集包含75例CBCT数据和7639个带有标签的断层图片,术后CBCT标注数据集包含58例CBCT数据和6638个带有标签的断层图片。2.本研究基于nnU-Net模型,采用块采样策略、超参数设置、学习率调整策略、训练早停策略,建立了两种术前CBCT图像自动分割模型,一种为2D模型,另一种为3D模型,其中3D模型的分割性能优于2D模型,3D模型的病变区域的平均DSC、平均Jaccard指数、平均表面距离和平均表面重叠度分别为0.935±0.056、0.881±0.079、0.548±1.774(mm)和0.958±0.064;3D模型的自动分割效果与高年资口腔医生手动分割效果进行统计学分析,结果显示本课题的3D模型的病变区域的平均DSC和平均Jaccard指数高于高年资口腔医生的相应指标(P<0.05),两种方法的其他评价指标无统计学差异(P>0.05);关于模型分割的特点,从分割速度来看,本课题的模型分割速度快,分割一个CBCT数据大约需要30s;从病变特点来看,首先本课题的模型可分割六种类型的囊性病变,包括根端囊肿(平均DSC为0.936±0.022)、含牙囊肿(平均DSC为0.915±0.090)、牙源性角化囊肿(平均DSC为0.944±0.021)、鼻腭囊肿(平均DSC为0.919±0.033)、血外渗性囊肿(平均DSC为0.956±0.012)、成釉细胞瘤(平均DSC为0.952±0.005);然后本课题的模型不仅可分割出大而明显的囊性病变(最大病变的体积为37232.125 mm~3),还可分割出小而隐匿的囊性病变(最小病变的体积为245.5 mm~3);最后本课题的模型还可精确分割跨中线的囊性病变。3.针对多类分割具有挑战性和成骨区域预测困难的问题,本研究在继承实验二中术前模型参数的情况下,建立了两种基于nnU-Net的术后CBCT图像自动分割模型,一种为2D模型,另一种为3D模型,其中3D模型的分割性能优于2D模型,3D模型的剩余囊腔区域的平均DSC、平均Jaccard指数、平均表面距离和平均表面重叠度分别为0.821±0.115、0.709±0.135、0.534±0.645(mm)和0.877±0.123,3D模型的成骨区域的平均DSC、平均Jaccard指数、平均表面距离和平均表面重叠度分别为0.685±0.114、0.531±0.126、0.744±0.786(mm)和0.834±0.105;3D模型的自动分割效果与高年资口腔医生手动分割效果进行统计学分析,结果显示对于剩余囊腔区域,这两种方法的分割效果无统计学差异(P>0.05),对于成骨区域,本课题的3D模型的DSC和Jaccard指数低于高年资口腔医生的相应指标(P<0.05),两种方法的其他评价指标无统计学差异(P>0.05);关于术后模型分割的特点,从分割区域来看,本课题的模型对剩余囊腔的分割效果优于成骨区域的分割效果;从复查时间来看,本研究的模型对术后不同复查时间的CBCT图像上的剩余囊腔的分割效果差别不大,对成骨区域的分割,复查1-3月、4-6月、7-12月的分割性能逐渐提高;关于本课题的模型在刮治术术后疗效分析中的应用,模型预测的成骨比例和剩余囊腔比例的变化趋势与“金标准”的一致,即在刮治术后3、6、12个月成骨比例逐渐增加,剩余囊腔比例逐渐减少,表明本课题的模型具有临床上辅助术后疗效分析的应用价值。结论:本课题建立了带有标签的颌骨囊性病变术前CBCT标注数据集和术后CBCT标注数据集,并在这两个数据集的基础上,成功创建了基于nnU-Net的颌骨囊性病变CBCT图像自动分割模型。该模型对术前和术后的CBCT图像均可快速分割,可大大提高临床诊疗效率。对于术前CBCT图像上的囊性病变区域,该模型自动分割的准确性高,相当于高年资口腔医生的分割水平,可用于临床上精确定位常见囊性病变和快速测量病变体积。对于术后CBCT图像,本课题的模型可实现对剩余囊腔和成骨区域的快速识别和自动定位,可用于临床上辅助分析术后疗效、动态监测病变转归,但未来还需进一步的研究来提高成骨区域的分割效果,以实现精准测量。通过纵向比较不同复查时间的成骨情况有助于发现复发灶,但对于复发灶的内部特征(如HU值、形态等)的细微变化还有待进一步研究,以实现早期发现复发灶。更多还原

【Abstract】 ObjectiveGiven the lack of datasets for cystic lesions of jaws and the time-consuming and labor-intensive problems of manual segmentation of CBCT images,this study established preoperative and postoperative CBCT annotated datasets for cystic lesions of jaws.On this basis,the deep learning method was used to establish the automatic segmentation model of preoperative CBCT images and the automatic segmentation model of postoperative CBCT images.It can achieve rapid localization and accurate measurement of cystic lesions in the preoperative CBCT images,automatic identification of residual cyst area and osteogenic area in the postoperative CBCT images,and real-time evaluation of postoperative efficacy.Methods1.To establish CBCT annotated datasets for cystic lesions of jawsFirstly,the clinical information,the preoperative CBCT data,and the postoperative CBCT data of patients with cystic lesions of jaws were collected based on pathological diagnosis in the Department of Pathology,West China Hospital of Stomatology,Sichuan University.Secondly,the labels were determined.That is,the preoperative labels were background(0)and cystic area(CA,1),and postoperative labels were background(0),residual cyst area(RCA,1),and osteogenesis area(OA,2).Finally,the preoperative and postoperative CBCT data were imported into the 3D Slicer system respectively to establish labels.After annotation,the labels and original CBCT data were saved in NRRD format to complete the establishment of the preoperative CBCT annotated dataset and postoperative CBCT annotated dataset for cystic lesions of jaws.In order to establish the standardized labeling process and get high-quality labels,all CBCT images were labeled by a dental surgeon with at least 3 y of clinical working experience,and all labels were checked and corrected by an experienced expert with more than 25 y of clinical work and scientific research experience.The ultimate annotated datasets which serve as the"gold standard"were used for training and evaluating a model.2.To establish an automatic segmentation model of preoperative CBCT image for cystic lesions of jaws based on deep learningThis study selected the nnU-Net model for segmentation training and adopted five-fold cross-validation for allocating the preoperative CBCT annotated dataset for cystic lesions of jaws.This study established oral and maxillofacial image processing models which can automatically segment and locate cystic lesions of jaws.The specific structure of the model adopted in this study can be divided into two types according to the dimension of the input.One is the nnU-Net 2D model for slice processing of CBCT data(the matrix of the input data size of H*W).The other is the nnU-Net 3D model for three-dimensional-patch processing of CBCT data(the matrix of input data size is H*W*D).Dice similarity coefficient(DSC),Jaccard index,mean surface distance,and surface overlap were used to evaluate the segmentation performance of the model.In order to further test the performance of the model,10 CBCT data were randomly selected from the preoperative CBCT annotated dataset for manual segmentation by the other 3 senior dental surgeons,and DSC,Jaccard index,mean surface distance,and surface overlap were calculated.The automatic segmentation results of the model and the manual segmentation results of the senior dental surgeon were statistically analyzed.3.To establish an automatic segmentation model of postoperative CBCT image for cystic lesions of jaws based on deep learningIn terms of model establishment and dataset allocation,the nnU-Net 2D model and nnU-Net 3D model were used in the study of postoperative CBCT data,which was the same as the study of preoperative CBCT data segmentation for cystic lesions of jaws.However,the details of the postoperative model are different from those of the preoperative model.The main differences are as follows:(1)the number of categories of the preoperative model is 2,while the number of categories of the postoperative model is 3,so the number of channels in the last segmentation layer of the two models is inconsistent.(2)The preoperative model was randomly initialized,while the postoperative model began segmentation training by inheriting the preoperative model parameters in Experiment 2.With the above methods,medical image processing models of an oral and maxillofacial region were established,which could automatically segment and locate the residual cyst area and osteogenesis area.The objective evaluation index and model performance evaluation method are consistent with experiment 2.Results1.In this study,a total of 113 patients with cystic lesions of jaws,75preoperative CBCT data,and 58 postoperative CBCT data were acquired.The preoperative and postoperative CBCT annotated datasets of cystic lesions of jaws were successfully established.The preoperative CBCT annotated dataset included 75CBCT data and 7639 labeled tomography images,and the postoperative CBCT annotated dataset included 58 CBCT data and 6638 labeled tomography images.2.Based on the nnU-Net model and adopting patch sampling strategy,hyperparameter setting,learning rate adjustment strategy,and training early stop strategy,this study established two preoperative CBCT image automatic segmentation models.One was the 2D model,and the other was the 3D model.The segmentation performance of the 3D model was better than that of the 2D model.The mean DSC,mean Jaccard index,mean surface distance,and mean surface overlap of 3D model lesions were 0.935±0.056,0.881±0.079,0.548±1.774(mm),and 0.958±0.064,respectively.Statistical analysis showed that the mean DSC and Mean Jaccard index of the 3D model were higher than those of the senior dental surgeon(P<0.05),and there was no statistical difference between the two methods in other evaluation indexes(P>0.05).As for the characteristics of model segmentation,from the perspective of segmentation speed,the segmentation speed of the model trained in this study was fast,and it took about 30s to segment a CBCT scan.From the characteristics of lesions,firstly,the model of this study could segment 6 types of cystic lesions,including radicular cyst(mean DSC 0.936±0.022),dentigerous cyst(mean DSC 0.915±0.090),odontogenic keratocyst(mean DSC 0.944±0.021),nasopalatine cyst(mean DSC 0.919±0.033),extravasation cyst(mean DSC 0.956±0.012),ameloblastoma(mean DSC 0.952±0.005);Secondly,the model of this study could not only segment the large and obvious cystic lesions(the volume of the largest lesion is 37232.125 mm~3)but also could segment the small and hidden cystic lesions(the volume of the smallest lesion is 245.5 mm~3);Finally,the model could accurately segment the cystic lesions across the midline.3.Given the challenges of multiclass segmentation and the difficulty in predicting osteogenic regions,this study established two automatic segmentation models of postoperative CBCT images based on nnU-Net,inheriting the preoperative model parameters in Experiment 2.One was the 2D model,and the other was the 3D model.The segmentation performance of the 3D model was better than that of the 2D model.The mean DSC,mean Jaccard index,mean surface distance,and mean surface overlap of the residual cystic area of the 3D model were0.821±0.115,0.709±0.135,0.534±0.645(mm),and 0.877±0.123,respectively.The mean DSC,mean Jaccard index,mean surface distance,and mean surface overlap of the osteogenesis area of the 3D model were 0.685±0.114,0.531±0.126,0.744±0.786(mm),and 0.834±0.105,respectively.Statistical analysis showed that there was no statistical difference between the automatic segmentation effect and the manual segmentation effect for the residual cystic area(P>0.05).For the osteogenesis area,the DSC and Jaccard index of the 3D model were lower than the corresponding indexes of the senior dental surgeon(P<0.05).There was no statistical difference in other evaluation indexes between the two methods(P>0.05).As for the characteristics of postoperative model segmentation,from the perspective of the segmentation area,the segmentation effect of the model for the residual cystic area is better than that for the osteogenesis area.From the view of follow-up time,the segmentation effect of the model for the residual cyst area at different times after surgery has little difference.For the osteogenesis area,the segmentation performance gradually improves in 1-3 months,4-6 months,and 7-12 months after surgery.About the application of the model in the postoperative efficacy analysis,the trend of our model is consistent with the"gold standard"in the proportion of the osteogenesis area and the proportion of the residual cystic area.That is,the proportion of the osteogenesis area gradually increased and the proportion of the residual cystic area gradually reduced.It showed that our model has the application value of assisting postoperative efficacy analysis.ConclusionsIn this study,preoperative and postoperative CBCT annotated datasets were established.On the basis of these datasets,the automatic segmentation model of CBCT images for cystic lesions of jaws based on nnU-Net was successfully established.The model can rapidly segment preoperative and postoperative CBCT images,which can greatly improve the efficiency of clinical diagnosis and treatment.For cystic lesion areas on preoperative CBCT images,the segmentation accuracy of our model is high,which is equivalent to the segmentation accuracy of a senior dental surgeon and can be used to accurately locate common cystic lesions,rapidly segment and accurately measure lesion volume in clinical practice,greatly simplify image processing and improve clinical diagnosis and treatment on the premise of ensuring accuracy.For postoperative CBCT images,our model can achieve rapid recognition and automatic localization of residual cystic area and osteogenesis area,which can be used in clinically assisting analysis of postoperative efficacy and real-time monitoring of disease outcomes.However,further research is needed to improve the segmentation effect of residual cystic area and osteogenesis area in the future,so as to achieve accurate measurement.Longitudinal comparison of osteogenesis at different follow-up times is helpful for the detection of recurrent foci,but the subtle changes in internal characteristics(such as HU score and morphology)of recurrent foci need to be further studied to achieve early detection of recurrent foci.更多还原