【作者】 刘静；

【导师】 左秀丽；

【作者基本信息】 山东大学 ， 内科学（消化系病）， 2024， 博士

【摘要】 研究背景:结直肠癌位居2022年中国癌症死亡原因的第四位,近年来发病率呈年轻化趋势。随着结直肠癌筛查和监测策略的全球推广,结肠镜检查(结直肠癌筛查的金标准)的数量正呈指数级增长。规范的结肠镜检查是促进结肠癌早期诊断和早期治疗的关键。鉴于不同层级医院的基础条件、不同医师在结肠镜退镜技巧、病变识别及认知方面存在明显差异,以致部分结直肠腺瘤被内镜医师漏检,进一步削弱了内镜检查对结直肠癌的保护作用。目前,世界各地区的多项内镜相关指南及共识均推荐进行标准化操作,用以优化内镜检查质量及病变检出率等关键指标。基于标准化操作的内镜检查质量控制可降低内镜医师操作间的异质性,进而改善肿瘤性病变的检出率及患者预后。虽然传统的人工目检策略在内镜质控方面卓有成效,但需要耗费较高的卫生人力成本,在我国临床内镜实践中推动落地难度较大。至此,亟需开发一种行之有效的内镜质控策略用以提高结直肠腺瘤检出率,改善我国的肠镜检查质量。内镜质量控制的实质是在确保安全的前提下,实现对内镜操作全流程的实时监督和反馈,为患者提供高效、高质量的内镜检查,改善恶性疾病的预后。随着人工智能(artificial intelligence,AI)技术的快速迭代,包括随机森林、支持向量机和深度卷积神经网络(deep convolutional neural network,DCNN)在内的算法在医疗影像识别领域的高效分析得到了消化内镜医师的广泛关注。前期我们团队围绕内镜质量控制指标,开发了一套能在术中实时反馈内镜检查质量的智能质量控制系统(automatic quality control system,AQCS),协助医师提高结直肠腺瘤的检出。考虑到前期开发过程中可能受患者群体及内镜医师经验等多种因素影响而造成数据集偏移,迫切需要优化AQCS并评估其在复杂临床实际中的应用价值。因此,本研究围绕AQCS在结肠镜质量控制方面开展了系统研究,通过三个部分逐层对AQCS进行了模型和数据集的优化,并在真实世界数据及多中心随机对照盲法试验中进一步验证了 AQCS在结肠镜检查中的质控效能。第一部分基于DCNN的AQCS模型优化与测试研究目的:通过算法更新及数据集扩充,对AQCS系统进行优化及性能验证。研究方法:本研究部分将针对AQCS的病灶识别模型、肠道清洁度评分模型及回盲部识别模型分别进行优化与验证。收集山东大学齐鲁医院由宾得或奥林巴斯系列结肠镜拍摄的白光图像。数据集内所有图像由两名具有5年以上结肠镜操作经验的消化内科专家标注、审核并确认。若数据标注出现异议,需由另一名经验丰富的内镜专家对图像进行独立审核标记,并确定最终标签。图像数据集按照不同质控模型的训练目的,分为训练集、验证集和测试集:利用训练集构建初步深度学习模型后,根据验证集进一步调整超参数及模型评估,最终在测试集内对病灶识别模型、肠道清洁度评分模型及回盲部识别模型分别进行效能验证。研究结果:本部分经筛选后共收集结肠镜白光图片26266张。其中,病灶识别模型的准确度、敏感性、特异性及其 95%CI 分别为 99.0%(97.8%-99.1%)、98.4%(97.6%-99.2%)和98.1%(96.1%-99.8%)。肠道清洁度评分模型的准确度、敏感性、特异性及其95%CI分别为 95.4%(92.5%-97.3%)、93.3%(92.4%-94.3%)和 99.0%(98.6%-99.4%)。回盲部识别模型的准确度、敏感性、特异性及其95%CI分别为99.3%(99.0%-99.6%)、98.7%(98.5%-99.3%)和99.7%(99.6%-99.9%)。在病灶识别模型中,识别结直肠息肉的受试者工作特征(receiver operating characteristic,ROC)曲线下面积为0.99,识别结直肠早癌的ROC曲线下面积为1.00,识别正常图像的ROC曲线下面积为1.00。在肠道清洁度评分模型中,根据波士顿肠道准备评分量表(Boston bowel preparation scale,BBPS)评为0至3分的ROC曲线下面积均高于0.97,识别反光辅助类别及模糊辅助类别的ROC曲线下面积均高于0.99。在回盲部识别模型中,识别回盲部的ROC曲线下面积为0.99。研究结论:1、基于DCNN的AQCS优化模型能够实时识别回盲部、评估肠道清洁度以及辅助结直肠早癌和息肉检出。2、通过算法更新及数据集扩充优化后的AQCS模型,在测试集中的回盲部识别、肠道清洁度评估、结直肠早癌和息肉检出方面达到90%以上的准确度、敏感性以及特异性,为该系统在临床实践中的下一步应用奠定了前期基础。第二部分 AQCS优化模型对新手内镜医师筛查结肠镜检查质量的影响研究目的:AQCS优化模型辅助新手内镜医师进行筛查结肠镜的有效性和安全性尚不清楚。本部分通过对大数据进行分析,探讨应用智能质量控制系统是否能改善新手内镜医师的结直肠腺瘤检出能力。研究方法:纳入2020年10月至2022年10月间在山东大学齐鲁医院门诊进行筛查结肠镜的40-75岁患者进行分析。纳入已常规部署AQCS优化模型的内镜操作间,且该系统可经由新手内镜医师自由开关。排除标准包括:检查指征为诊断性结肠镜或随访监测结肠镜者,有炎症性肠病史、遗传性结直肠癌综合征及其各类息肉病、活检禁忌症及严重合并症者,孕期及哺乳期患者,寻求治疗性肠镜者,以及非新手医师操作的结肠镜检查。研究期间,若同一患者进行多次肠镜检查,仅纳入该时期的首次筛查肠镜进行分析。本研究的主要结局是结直肠腺瘤检出率(adenoma detection rate,ADR)。次要结局包括进展期结直肠癌检出率、结直肠息肉检出率(polyp detection rate,PDR)、进展期腺瘤检出率(advanced adenoma detection rate,AADR)、广基锯齿状病变(sessile serrated lesion,SSL)检出率、近端结肠腺瘤检出率、非肿瘤性息肉检出率、结肠镜平均腺瘤检出数(adenoma per colonoscopy,APC)、结肠镜平均息肉检出数(polyp per colonoscopy,PPC)、结肠镜平均进展期腺瘤检出数(advanced adenoma per colonoscopy,AAPC)、结肠镜平均近端结肠腺瘤检出数、肠道准备合格率及BBPS评分、进镜时间、退镜时间及不良事件。该研究已获山东大学齐鲁医院医学伦理委员会(2021285)批准。研究结果:本研究共纳入4013例经由新手医师操作的筛查结肠镜检查,AQCS设备主动开启率为56.1%。常规检查组及AQCS辅助检查组的ADR分别为12.9%和18.6%(P<0.001),其中,AQCS辅助检查组的近端结肠腺瘤检出率为10.6%,高于常规检查组的7.6%(P=0.001)。两组的PDR、AADR、进展期结直肠癌及SSL检出率无统计学差异。相比于常规检查组,AQCS辅助检查组的非肿瘤性息肉检出率增加2.1%(7.7%vs 5.6%,P=0.008)。同时,AQCS辅助检查组提高了 APC(0.22vs0.15,P<0.001)和结肠镜平均近端结肠腺瘤检出数(0.12 vs 0.09,P=0.002),但对AAPC和PPC无明显改善。中位进镜时间在AQCS辅助检查组为7.38分钟,较常规检查组延长2.18分钟(P<0.001)。相较于常规检查组,AQCS干预延长了该组的中位退镜时间(13.82 mins vs 9.68 mins,P<0.001)和阴性结肠镜退镜时间(11.50 mins vs 8.40 mins,P<0.001)。两组在肠道准备合格率上无统计学差异(92.0%vs 90.4%,P=0.079)。然而,AQCS干预改善了总体及各肠段的BBPS评分(分别为P<0.001、P<0.001、P=0.003和P=0.011)。两组均未报告结肠镜检查相关的并发症。研究结论:1、相比于常规检查,应用AQCS优化模型进行辅助质控可以提高新手内镜医师在筛查结肠镜中的ADR及APC,为下一步将AQCS引入结肠镜医师培训路径提供了初步有利证据。2、AQCS优化模型通过改善肠道准备及延长退镜时间来优化新手内镜医师的结肠镜退镜表现。第三部分 AQCS优化模型提高结直肠腺瘤检出率的前瞻性、多中心验证研究目的:尽管智能质量控制系统已在单个学术医疗中心内验证了其改善结直肠腺瘤检出率的有效性及安全性,但尚缺乏多中心研究进一步评估AQCS优化模型的临床应用价值。本试验的主要目的是在日常结肠镜检查中明确实时AQCS优化模型是否可改善学术和非学术医疗中心内镜医师的ADR。研究方法:本研究是一项前瞻性、多中心、盲法随机对照试验,于中国6所医院的内镜中心实施,包括3个学术医疗中心和3个非学术医疗中心。通过按中心分层的区组随机化,将符合条件的患者被随机分配(1:1)接受标准结肠镜检查或AQCS辅助结肠镜检查。本研究对受试者、病理学家和数据分析人员设盲。AQCS可自动识别回盲部,智能记录并提示结肠镜的进镜及退镜时间;并通过监测退镜稳定性来提醒内镜医师平稳退镜,以减少黏膜观察盲区;实时识别当前视野的BBPS评分,当肠道准备不合格(BBPS评分<2分)时,AQCS可经语音提醒医师进行黏膜冲洗或粪水吸引,提高黏膜可视度;识别可疑病变并在第二观察屏显示标注病灶,音频提示医师及时干预。本研究的主要结局是ADR,并根据患者、协作中心及内镜医师等特征对ADR进行了亚组分析。次要结局指标包括:根据病变的位置、大小、形态及病理诊断分类的ADR水平、APC、BBPS评分、退镜观察时间及严重不良反应事件等。在研究开始前,本设计方案已在ClinicalTrials.gov完成注册(NCT04901130)。研究结果:研究总计纳入1254名患者,意向性分析显示,AQCS辅助检查组的ADR高于对照组(32.7%vs 22.6%,RR 1.60,95%CI 1.23-2.09,P<0.001)。AQCS 提升了学术医疗中心及非学术医疗中心的ADR(学术中心29.3%vs 20.8%,RR 1.58,95%CI 1.10-2.29,P=0.014;非学术中心 36.1%vs 24.5%,RR 1.74,95%CI 1.23-2.46,P=0.002)。AQCS改善了腺瘤检出水平较低(基线ADR<25%)及检出水平中等(基线ADR 25%-35%)内镜医师的 ADR(基线 ADR<25%者 30.0%vs 20.0%,RR 1.71,95%CI 1.24-2.35,P=0.001;基线 ADR25%-35%者 38.1%vs 27.7%,RR 1.61,95%CI 1.07-2.43,P=0.023)。当考虑最终病理组织学诊断时,相比于对照组,AQCS的干预提高了非进展期腺瘤的检出率(30.1%vs 21.2%,P=0.002),而对于进展期腺瘤及结直肠癌的检出并没有改善。特别指出,两组间非肿瘤性息肉检出率无统计学差异(23.9%vs21.2%,RR 1.10,95%CI 0.83-1.46,P=0.485),提示AQCS干预并未增加无效病变切除。其中,接受息肉切除术但病理证实无腺瘤、结直肠癌或广基锯齿状病变的患者分别占AQCS辅助检查组和对照组的 1 1.8%和 13.7%(RR 0.95,95%CI 0.67-1.36,P=0.780)。与此同时,当考虑病变大小时,AQCS辅助检查组可增加微小及大腺瘤的检出率(P=0.002,P=0.044)。对于病变位置方面,AQCS的干预提升了升结肠和横结肠的ADR(P=0.005,P=0.011),而似乎对回盲部、降结肠、乙状结肠及直肠的ADR没有改善。当考虑病变形态时,AQCS的干预可协助内镜医师检出更多的扁平及无蒂腺瘤(29.3%vs20.4%,RR 1.52,95%CI 1.16-2.00,P=0.003)。同时,AQCS 提升了 APC(0.86 vs 0.48;RR 1.50,95%CI 1.17-1.91,P=0.001)。AQCS的干预提高了腺瘤检出水平较低内镜医师的APC(0.82 vs 0.34,RR 1.71,95%CI 1.24-2.35,P=0.001)。在AQCS质控下,退镜观察时间和BBPS评分均提升(P<0.001,P=0.010)。两组均未发生严重不良反应。研究结论:1、通过前瞻性、多中心、盲法、随机对照试验验证,AQCS优化模型的多方面质控干预措施可通过延长退镜观察时间和提高肠道清洁度来改善学术及非学术医疗中心内日常结肠镜检查的ADR及APC,但应注意非肿瘤性病变的切除。2、建议腺瘤检出水平较低(基线ADR<25%)的内镜医师选择AQCS以提高其ADR及 APC。更多还原

【Abstract】 Background:Colorectal cancer(CRC)ranked fourth among the causes of cancer-related mortality in China in 2022,causing substantial burdens.Standardized endoscopy examination is a key element in promoting early diagnosis and treatment of tumors.Given the differences in endoscopist procedural skills and cognitive abilities across various levels of hospitals,a certain proportion of early-stage gastrointestinal cancers and precancerous lesions are missed during endoscopy procedures.Consequently,this significantly limits the role of screening endoscopy in gastrointestinal tumors prevention.Currently,multiple endoscopy guidelines and consensus statements recommend standardized procedures to optimize the quality of endoscopy procedures and key indicators such as adenoma detection rate(ADR).Quality control of endoscopy procedures could significantly reduce inter-operator variability among endoscopists,thereby improving the detection rates of neoplastic lesions.However,these quality control strategies,based on traditional manual inspection,entail high demands on medical care costs related to physician manpower,making them difficult to implement effectively in clinical practice in China.Thus,there is an urgent need to develop a practical and effective quality control strategy to improve the detection rates of early-stage gastrointestinal cancers and precancerous lesions,thereby improving the quality of digestive endoscopies.The essence of endoscopic quality control lies in ensuring safety and achieving real-time supervision and feedback throughout the entire procedure.This aims to provide patients with efficient and high-quality endoscopy examinations,ultimately contributing to the improvement of malignant disease prognosis.Recently,with the rapid advancement of artificial intelligence(AI),algorithms including random forests,support vector machines,and deep convolutional neural networks(DCNN)have garnered widespread attention from gastroenterologists in the field of medical image recognition.Consequently,we have developed an automatic quality control system(AQCS)based on DCNN,focusing on quality control for colonoscopy procedures.AQCS provides real-time feedback to endoscopists regarding the quality of colonoscopy procedures,thereby assisting physicians in improving the detection of early-stage cancers and precancerous lesions.Although AQCS has shown commendable performance in a single-center,randomized controlled trial,it is essential to consider that during the initial development phase,various factors such as patient demographics,hospital characteristics,and endoscopist experience may have led to dataset biases.Therefore,the value of AQCS in complex clinical settings needs to be further evaluated.Thus,this study focuses on the optimization and application of AQCS for colonoscopy quality control.Part 1 Optimization and testing of the AQCS model based on DCNNObjective:To optimize and evaluate the performance of the AQCS in colonoscopy quality control.Methods:In this study,we optimized DCNN models for detecting colorectal lesions,evaluating bowel preparation,and identifying cecum.White light images captured by Pentax or Olympus series colonoscopes at Qilu Hospital of Shandong University were retrospectively collected.All images in the dataset were reviewed and confirmed by at least two experienced colonoscopists(≥5 years),and the image was labeled according to the purposes of different quality-control models.In case of disagreement,another experienced endoscopist would independently review and label the images.Following image preparation,the datasets were partitioned into train,validation,and test sets based on the objectives of different quality control models.After obtaining preliminary deep learning models using the training set,further optimization of parameters was conducted using the validation set.Finally,we assessed the performance of the lesion detection model,bowel cleanliness model,and cecum identification model via the test set.Results:After screening,a total of 26,266 white-light colonoscopy images were collected.The accuracy,sensitivity,and specificity of the lesion detection model were 99.0%,98.4%,and 98.1%,respectively.For the bowel cleanliness model,the accuracy,sensitivity,and specificity were 95.4%,93.3%,and 99.0%,respectively.Regarding the cecum identification model,the accuracy,sensitivity,and specificity were 99.3%,98.7%,and 99.7%,respectively.In the lesion detection model,the area under the receiver operating characteristic(ROC)curve for identifying colorectal polyps was 0.99,while for identifying early colorectal cancer and normal images,the areas under the ROC curves were 1.00.In the bowel cleanliness model,the areas under the ROC curves for identifying Boston bowel preparation scale(BBPS)scores of 0 to 3 were all higher than 0.97,and for identifying reflection-assisted and blur-assisted categories,the areas under the ROC curves were both higher than 0.99.For the cecum identification model,the area under the ROC curve for identifying the cecum was 0.99.Conclusions:1 The DCNN-based,real-time,optimized AQCS model could identify the cecum,assess bowel cleanliness,and assist the detection of early colorectal cancer and polyps.2 After optimizing the DCNN model,the AQCS achieved good accuracy,sensitivity,and specificity in quality control for identifying cecum,assessing bowel cleanliness,and detecting colorectal lesions.Part 2 Effect of optimized AQCS on novice-performed screening colonoscopyObjective:Evidence on the real-world effectiveness and safety of the optimized AQCS in assisting novices in screening colonoscopies is limited.This study aimed to investigate whether the implementation of AQCS could improve the adenoma detection of novices.Methods:Patients aged 40-75 who underwent screening colonoscopy at the outpatient clinic of Qilu Hospital of Shandong University from October 2020 to October 2022 were analyzed.Endoscopy suites installed with the optimized AQCS were included,and the AQCS could be freely switched on and off by novices.Exclusion criteria included indications for diagnostic or surveillance colonoscopies,patients with a history of hereditary CRC syndrome and inflammatory bowel disease,biopsy contraindications and severe comorbidities,pregnancy or lactation,those seeking therapeutic colonoscopies,and colonoscopies performed by experts.During the implementation period,only the first colonoscopy record per patient was included.The primary outcome was ADR.Secondary outcomes included polyp detection rate(PDR),advanced adenoma detection rate(AADR),detection rates of advanced CRC,sessile serrated lesion(SSL)and non-neoplastic polyps,proximal ADR,adenoma per colonoscopy(APC),polyp per colonoscopy(PPC),advanced adenoma per colonoscopy(AAPC),proximal APC,bowel preparation adequacy rates,BBPS score,intubation and withdrawal time,and adverse events.The study protocol was approved by the Medical Ethics Committee of Qilu Hospital of Shandong University(2021285).Results:A total of 4013 screening colonoscopies performed by novices were included.The activation rate of AQCS was 56.1%.The ADRs of the conventional group and the AQCS-assisted group were 12.9%(227/1761)and 18.6%(418/2252),respectively(P<0.001).The proximal ADR of the AQCS-assisted group was 10.6%,which was significantly higher than the 7.6%of the conventional group(P=0.001).However,there was no significant difference in PDR,AADR,SSL and advanced CRC detection rates between the two groups.The detection rate of non-neoplastic polyps was significantly higher in AQCS-assisted group(7.7%vs 5.6%,P=0.008).APC and proximal APC in AQCS-assisted group were higher than those in conventional group(APC 0.22 vs 0.15,P<0.001;proximal APC 0.12 vs 0.09,P=0.002).We did not identify significant increases in AAPC and PPC.Intubation time was significantly longer compared with the conventional group,at 7.38 mins(median)and 5.20 mins(median),respectively.The implementation of AQCS significantly prolonged the median withdrawal time(13.82 mins vs 9.68 mins,P<0.001)and the median withdrawal time for negative colonoscopies(11.50 mins vs 8.40 mins,P<0.001).There was no significant difference in bowel preparation adequacy rates(92.0%vs 90.4%,P=0.079).However,the implementation of AQCS improved the mean overall BBPS score and subscores(P<0.001,P<0.001,P=0.003,and P=0.011).No serious adverse events occurred in either group.Conclusions:1 Compared with the conventional examination,the implementation of optimized AQCS improved the ADR of novices in screening colonoscopies.2 The optimized AQCS improved the performance of novices in screening colonoscopies by improving the quality of bowel preparation and prolonging withdrawal time.Part 3 Multicenter prospective validation of the optimized AQCS for improving detection rate of colorectal adenomasObjective:Although AQCS has verified its effectiveness and safety in improving the detection rate of colorectal adenomas in a single academic medical center,there is still a lack of multicenter validation to further evaluate its clinical application value.The primary outcome was to determine whether the real-time,optimized AQCS improves ADR among endoscopists at academic and non-academic medical centers during routine colonoscopies.Methods:This prospective,multicenter randomized controlled trial was performed in 6 centers(3 academic and 3 non-academic centers)in China.Eligible patients were randomly assigned(1:1)to either standard colonoscopy(SC)or AQCS-assisted colonoscopy via block randomization stratified by center.Intervention allocation was masked from patients,pathologists and statistical analysts,but endoscopists were aware of group assignment.In the AQCS group,AQCS was turned on before intubation and started supervising at withdrawal phases once the cecum was identified.Alongside the original videos,four additional visual and audio notices were fed back to endoscopists in real-time:(1)timer on a second high-definition monitor;(2)prompts for controlling withdrawal speed and reexamining certain segments when unsteady or fuzzy frames identified continuously by AQCS;(3)prompts for cleaning mucosa or suctioning liquid pools when suboptimal cleansing(BBPS score<2)was recognized;and(4)green tracking box on the monitor indicating lesions location.The primary outcome was the ADR,and subgroup analyses of ADR were conducted based on the characteristics of patients,centers and endoscopists.Secondary outcomes included AADR,proximal ADR,adenoma per colonoscopy(APC),BBPS score,withdrawal time without intervention,and adverse events.This protocol was registered on ClinicalTrials.gov before embarking on this study(NCT04901130).Results:A total of 1254 patients were included in the study.Intention-to-treat analysis showed that the ADR of the AQCS group was significantly higher than that of the SC group(32.7%vs 22.6%,RR 1.60,95%CI 1.23-2.09,P<0.001).AQCS improved the ADR in both academic and non-academic medical centers(academic centers 29.3%vs 20.8%,RR 1.58,95%CI 1.10-2.29,P=0.014;non-academic centers 36.1%vs 24.5%,RR 1.74,95%CI 1.23-2.46,P=0.002).AQCS significantly increased ADR of the lower-level detectors(30.0%vs 20.0%,RR 1.71,95%CI 1.24-2.35,P=0.001)as well as of the medium-level detectors(38.1%vs 27.7%,RR 1.61,95%CI 1.07-2.43,P=0.023).Significant differences were only found in the histology subgroups with non-advanced adenomas(30.1%vs 21.2%,P=0.002).The detection rate of non-neoplastic polyps between the two groups was not significantly different(23.9%vs 21.2%SC,P=0.485).Of these,patients who had undergone polypectomy but with no pathological proven adenoma,CRCs or SSLs accounted for 11.8%and 13.7%of the AQCS group and SC group(RR 0.95,95%CI 0.67-1.36,P=0.780).ADR in other histology subgroups(advanced adenoma,SSL or CRC)were not significantly different.The detection rate of adenomas was significantly higher in the AQCS group when considering shape(flat or sessile),size(diminutive and large),and location(ascending and transverse colon).At the same time,APC in AQCS group was higher than that in SC group(0.86 vs 0.48,RR 1.50,95%CI:1.17-1.91,P=0.001).Of note,APC in AQCS group was higher than those in SC group among the lower-level detectors(0.82 vs 0.34,RR 1.71,95%CI 1.24-2.35,P=0.001),but not among the medium-level detectors(0.94 vs 0.73,RR 1.16,95%CI 0.79-1.71,P=0.450).This procedure-incorporated AQCS could prolong withdrawal time without intervention and improve Boston bowel preparation scale score.No serious adverse events occurred.Conclusions:1 In this prospective,multicenter randomized controlled trial,the multifaceted quality improvement intervention utilizing optimized AQCS improved ADR and APC for routine colonoscopies in both academic and non-academic centers by extending withdrawal time without intervention and improving bowel cleanliness,but attention should be paid to avoiding unnecessary removal of non-neoplastic lesions.2 For lower-level detectors(baseline ADR<25%),AQCS was recommended to improve their ADR and APC.更多还原