【作者】 刘敏；

【导师】 魏丽娟；

【作者基本信息】 吉林大学 ， 临床医学硕士（专业学位）， 2024， 硕士

【摘要】 目的:胃间质瘤(gastric stromal tumors,GST)和平滑肌瘤(gastric leiomyomas,GLs)是胃黏膜下肿瘤(submucosal tumors,SMTs)的两种主要病理类型,两者的生物学行为差异显著,治疗方式和预后也截然不同,术前需要准确鉴别。超声内镜(endoscopic ultrasonography,EUS)是临床评估SMTs的首选检查方式,但准确区分GST和GLs仍然存在一定困难。近期,人工智能(artificial intelligence,AI)在GST与GLs鉴别诊断方面表现出了良好的效能,但相关研究主要是基于较大直径的SMTs数据,对于小直径SMTs,其影像学特征相对较少,AI模型能否表现出优异的诊断效能仍未可知。因此,本研究旨在构建一个基于EUS图像的小GST和GLs的AI鉴别诊断模型,并评估其在临床实践中的应用价值,以辅助临床医师实现GST和GLs的早期诊断和合理化治疗。方法:回顾性收集2018年1月至2022年9月期间在吉林大学第二医院接受EUS检查,并经术后病理学、免疫组织化学检查诊断为GST或GLs的患者资料,包括一般临床特征资料及EUS图像。其中,2018年1月至2021年3月入院患者的资料作为模型训练集,2021年4月至2022年9月入院患者的资料作为验证集(验证集~1:肿瘤直径<1cm;验证集~2:肿瘤直径<2cm)。(1)对训练集中患者临床特征进行单因素和多因素Logistic回归分析,筛选鉴别GST与GLs的预测因素,构建临床鉴别诊断模型,并在训练集和验证集中分别检验模型诊断效能。(2)在原始EUS图像中标记感兴趣区域(ROI),训练集图像进行数据增强处理用于训练AI模型,构建基于卷积神经网络(CNN)的鉴别诊断模型(EUS-AI),并在训练集和验证集中分别检验模型诊断效能。(3)在验证集中,将内镜医师的诊断结果与EUS-AI模型进行对比分析。结果:(1)临床特征单因素分析结果显示,肿瘤长径/短径(LD/SD)(OR=0.614,95%CI:0.489～0.770,P=0.001)、年龄(OR=1.019,95%CI:1.012～1.026,P=0.000)、CEA(OR=1.117,95%CI:1.027～1.215,P=0.032)和糖尿病(OR=1.710,95%CI:1.175～2.489,P=0.020)指标在GST和GLs组间存在统计学差异。GST组LD/SD更低,年龄更大,CEA水平更高,且有着更高的糖尿病患病率。多因素分析结果显示,低LD/SD(OR=0.709,95%CI:0.568～0.884,P=0.011)、高龄(OR=1.013,95%CI:1.005～1.020,P=0.008)是GST诊断的独立预测因素。将LD/SD、年龄、CEA和糖尿病4个指标纳入临床鉴别诊断模型的构建,其中随机森林模型(RF)的诊断效能较好,其在训练集、验证集~1、验证集~2中的AUC值分别为0.901、0.693和0.599。(2)基于Dense Net201、Res Net50和VGG19构建EUS-AI鉴别诊断模型,在训练集和验证集中分别检验3种模型的诊断效能。基于单张EUS图像的诊断结果显示,Dense Net201模型的整体诊断表现优于其他2种模型,其在训练集、验证集~1、验证集~2中的诊断AUC值分别为0.949、0.837和0.844。基于肿瘤个体的诊断结果显示,Res Net50模型在训练集、验证集~1、验证集~2中的AUC值分别为0.994、0.911和0.915,诊断效能优于其他2种模型。与临床模型(RF)相比,EUS-AI模型(Res Net50)的诊断性能更好。(3)低年资医师在验证集~1和验证集~2中的诊断AUC值分别为0.586和0.658,而高年资医师的诊断AUC值分别为0.621和0.702。2位内镜医师在验证集~2中的诊断表现均显著优于验证集~1,低年资医师的诊断AUC值略低于高年资医师,但两者差异无统计学意义(P>0.05)。当与EUS-AI模型进行比较时,2位医师的诊断AUC值均显著低于EUS-AI模型(P<0.05)。结论:(1)纳入LD/SD、年龄、CEA、糖尿病指标构建的临床特征模型对于小GST和GLs的鉴别诊断效果欠佳,临床应用存在一定局限性。(2)基于ResNet50构建的EUS-AI模型可以有效区分小GST,尤其是微小GST和GLs(直径<1cm),其中以肿瘤个体为单位的诊断效果优于单张图像的诊断。(3)与内镜医师相比,EUS-AI模型的诊断效能更高,该模型可以辅助临床医师在术前准确鉴别小GST与GLs,有助于制定合理治疗策略,具有良好的应用前景。更多还原

【Abstract】 Objective:Gastric stromal tumors(GST)and gastric leiomyomas(GLs)are the two main pathological types of submucosal tumors(SMTs).It is crucial to accurately identify them before surgery due to their significantly different biological behaviors,treatment methods,and prognosis.Endoscopic ultrasonography(EUS)is the preferred method for clinical evaluation of SMTs.However,accurately distinguishing between GST and GLs remains challenging.Recent studies have shown that artificial intelligence(AI)can effectively distinguish between GST and GLs.However,the studies are primarily based on data from large-diameter SMTs.It is uncertain whether the AI model can achieve the same level of diagnostic performance for small-diameter SMTs due to the limited imaging features available.Therefore,the objective of this study is to construct an AI model for differential diagnosis between small GST and GLs based on EUS images and evaluate its application value in clinical practice,in order to assist clinicians in achieving early diagnosis and appropriate treatment of GST and GLs.Methods:This study retrospectively collected the data of patients who underwent EUS examination at the Second Hospital of Jilin University from January 2018 to September 2022 and were diagnosed with GST or GLs by post-operative pathology and immunohistochemistry,including general clinical features and EUS images.Of them,data from patients admitted from January 2018 to March 2021 were used as the model training set,and those from April 2021 to September 2022 as the validation set(validation set ~1:tumor diameter<1 cm;validation set ~2:tumor diameter<2 cm).(1)Univariate and multivariate logistic regression analyses were performed on the clinical features data of the patients in the training set to identify predictive factors for distinguishing GST from GLs,constructing a clinical differential diagnosis model,and testing its diagnostic efficacy in both the training and validation sets.(2)Mark the region of interest(ROI)in the original EUS image,perform data augmentation on the training set images for training the AI model,and then construct a convolutional neural network(CNN)-based differential diagnosis model(EUS-AI),and test its diagnostic efficacy in both the training and validation sets.(3)In the validation sets,the diagnostic results of endoscopists were compared with the EUS-AI model.Results:(1)Univariate analysis of clinical features showed statistically significant differences between GST and GLs groups in LD/SD(OR=0.614,95%CI:0.489-0.770,P=0.001),age(OR=1.019,95%CI:1.012-1.026,P=0.000),CEA levels(OR=1.117,95%CI:1.027-1.215,P=0.032),and diabetes(OR=1.710,95%CI:1.175-2.489,P=0.020).The GST group had lower LD/SD,older age,higher CEA levels,and higher prevalence of diabetes.Multivariate analysis revealed that low LD/SD(OR=0.709,95%CI:0.568-0.884,P=0.011)and older age(OR=1.013,95%CI:1.005-1.020,P=0.008)were independent predictors of GST diagnosis.Variables of LD/SD,age,CEA,and diabetes were selected for inclusion in the construction of the clinical differential diagnosis model,among which the Random Forest model(RF)had better diagnostic efficacy,and its AUC in the training set,validation set ~1,and validation set ~2 were 0.901,0.693,and 0.599 respectively.(2)The EUS-AI differential diagnosis model was constructed using Dense Net201,Res Net50,and VGG19,and the diagnostic efficacy of the three models was tested in both the training and validation set.The diagnostic results from the EUS image showed that the overall performance of the Dense Net201 model was superior to the other two models,with AUCs of 0.949,0.837,and 0.844 in the training set,validation set ~1,and validation set ~2 respectively.The tumor diagnostic results indicated that the Res Net50 model had AUCs of 0.994,0.911,and 0.915 in the training set,validation set ~1,and validation set ~2 respectively,demonstrating better diagnostic efficacy than the other models.Compared to the clinical model(RF),the EUS-AI model(Res Net50)performed better in differential diagnosis.(3)The AUCs of the junior endoscopists in the validation set ~1and validation set~2 were 0.586 and 0.658,respectively,while the senior endoscopists’AUCs were 0.621and 0.702 respectively.Both endoscopists performed significantly better in validation set ~2 than in validation set ~1,and the AUCs of the junior endoscopists were slightly lower than those of the senior endoscopists,but there was no statistically significant difference between them(P>0.05).When compared to the EUS-AI model,both endoscopists had significantly lower AUCs than the EUS-AI model(P<0.05).Conclusion:(1)The clinical differential diagnosis model incorporating the variables LD/SD,age,CEA,and diabetes demonstrates suboptimal diagnostic efficacy for distinguishing between small GST and GLs,with certain limitations in clinical application.(2)The EUS-AI model constructed based on Res Net50 can effectively distinguish small GST,especially micro-GSTs,from GLs,with superior diagnostic efficacy for per-tumor compared to per-image diagnosis.(3)Compared to endoscopists,the EUS-AI model has a higher diagnostic efficiency,which can assist clinicians in accurately distinguishing small GST from GLs preoperatively,thereby formulating reasonable treatment strategies and exhibiting promising application prospects.更多还原