基于卷积神经网络的肠息肉分类方法的初步研究

目的通过开发一种基于卷积神经网络(convolutional neural network, CNN)的人工智能诊断模型,用于自动识别腺瘤和增生性息肉,以增加结直肠镜检查的统一性和客观性。方法回顾性收集南京医科大学第一附属医院内镜中心的白光内镜图像1 796张,其中正常结直肠图像692张,增生性息肉608张,腺瘤496张。在我们的分类系统上提出双流网络(dual-stream network),包括原始流和检测流,原始流的输入是完整的预处理肠镜图像,以用来学习全局特征,检测流重点关注病灶的局部特征。我们将数据集分为训练集、测试集和验证集进行10次交叉验证,以评估该方法的有效性。结果 CNN方法的正确率(accuracy)、准确率(precision)、召回率(recall)分别为96.9%、96.6%、96.7%,在正常肠镜图片、增生性息肉、腺瘤三类图像中的正确识别率分别为100%、95.1%和95.0%。结论 CNN系统对结直肠息肉的识别具有较高的特异性和灵敏性,可以帮助临床内镜医师快速诊断和识别结直肠息肉的类型。这表明我们的方法在临床上能够对肠息肉病变进行有效、准确、可靠的诊断。     

深度学习技术在提升结直肠息肉性质鉴别准确率中的应用

目的探讨采用深度学习技术提升内镜医师在窄带光成像（narrow band imaging,NBI）下判断结直肠息肉性质准确率的价值。方法收集武汉大学人民医院消化内镜中心结直肠息肉的NBI非放大图片并分为3个数据集,数据集1（2018年1月—2020年10月,1 846张非腺瘤性与2 699张腺瘤性息肉的NBI非放大图片）用来训练和验证结直肠息肉性质鉴别系统;数据集2（2018年1月—2020年10月,210张非腺瘤性息肉和288张腺瘤性息肉的NBI非放大图片）用来比较内镜医师及该系统息肉分型的准确性,同时比较4名消化内镜初学者在该系统的辅助下判断息肉性质的准确性是否有提升;数据集3（2020年11月—2021年1月,141张非腺瘤性息肉和203张腺瘤性息肉的NBI非放大图片）用来前瞻性测试该系统。结果该系统在数据集2中判断结直肠息肉的准确率为90.16%（449/498）,优于内镜医师。消化内镜初学者在有该系统的辅助下,息肉分型准确率显著提升。在前瞻性研究中,该系统的准确率为89.53%（308/344）。结论本研究开发的基于深度学习的结直肠息肉性质鉴别系统能够显著提升内镜医师初学者的息肉分型准确率。     

人工智能对内镜医师染色放大内镜下胃癌识别能力的影响研究

目的评估人工智能（artificial intelligence,AI）辅助胃癌诊断系统在实时染色放大内镜视频中对内镜医师识别胃癌能力的影响。方法回顾性收集2017年3月—2020年1月武汉大学人民医院和公开数据集中的早期胃癌和非癌染色放大内镜图片作为训练集和独立测试集,其中训练集包括4 667张图片（1 950张早期胃癌和2 717张非癌）,测试集包括1 539张图片（483张早期胃癌和1 056张非癌）。利用深度学习进行模型训练。前瞻性收集2020年6月9日—2020年11月17日来自北京大学肿瘤医院和武汉大学人民医院的100例患者的染色放大内镜视频（包含38例癌和62例非癌）作为视频测试集。纳入来自另外4家医院的4名不同年资内镜医师,分2次（无或有AI辅助）对视频测试集进行诊断,评估AI对内镜医师判断胃癌能力的影响。结果无AI辅助时,内镜医师诊断视频测试集中胃癌的准确率、敏感度和特异度分别为81.00%±4.30%、71.05%±9.67%和87.10%±10.88%;在AI辅助下,内镜医师辨认胃癌的准确率、敏感度和特异度分别为86.50%±2.06%、84.87%±11.07%和87.50%±4.47%,诊断准确率（P=0.302）和敏感度（P=0.180）较无AI辅助时均有提升。AI在视频测试集中辨认胃癌的准确率为88.00%（88/100）,敏感度为97.37%（37/38）,特异度为82.26%（51/62）,AI的敏感度高于内镜医师平均水平（P=0.002）。结论AI辅助诊断系统是染色放大内镜模式下辅助诊断胃癌的有效工具,可提高内镜医师对胃癌的诊断能力。它能实时提醒内镜医师关注高风险区域,以降低漏诊率。     

基于人工智能的自动内镜下病灶尺寸测量系统（含视频）

目的　开发一个基于人工智能的自动内镜下病灶尺寸测量系统,并测试其实时测量白光内镜下病灶尺寸的能力。方法　测量系统由3个模型组成：首先由模型1识别视频的连续图片中有无活检钳,有钳者标记钳叶轮廓;随后由模型2对有钳图片进行分类,分为张钳图片和未张钳图片;与此同时,模型3识别视频的连续图片中有无病灶,有病灶者标记边界;最后系统根据活检钳钳叶轮廓与病灶边界的像素对比,实时计算出病灶尺寸。数据集1由回顾性收集的武汉大学人民医院2017年1月1日—2019年11月30日4 835张图片组成,用于模型的训练和验证;数据集2由前瞻性收集的武汉大学人民医院内镜中心2019年12月1日—2020年6月4日检查拍摄的图片组成,用于测试模型分割活检钳边界和病灶边界的能力;数据集3由151个模拟病灶的302张图片组成,每个模拟病灶包括活检钳倾斜角度较大（与病灶垂直线夹角45°）和倾斜角度较小（与病灶垂直线夹角10°）情况下的图片各1张,用于测试模型在活检钳不同状态下测量病灶尺寸的能力;数据集4为视频测试集,由前瞻性收集的武汉大学人民医院内镜中心2019年8月5日—2020年9月4日检查拍摄的视频组成。以内镜医师复核后结果或内镜手术病理作为金标准,观察模型1识别有无活检钳的准确率、模型2分类活检钳状态（张钳或未张钳）的准确率和模型3识别有无病灶的准确率,用交并比（intersection over union,IoU）评价模型1的活检钳钳叶分割效果和模型3的病灶分割效果,用绝对误差和相对误差评价系统的病灶尺寸测量能力。结果　（1）数据集2共纳入1 252张图片,有钳图片821张（其中张钳图片401张、未张钳图片420张）、无钳图片431张;包含病灶图片640张、不包含病灶图片612张。模型1判断无钳图片433张（430张准确）、有钳图片819张（818张准确）,识别有无活检钳的准确率为99.68%（1 248/1 252）,以818张模型1准确判断有钳图片的数据统计模型1的活检钳钳叶分割效果,平均IoU为0.91（95%CI：0.90～0.92）。使用模型1准确判断的818张有钳图片评价模型2的活检钳状态分类准确率,模型2判断张钳图片384张（382张准确）、未张钳图片434张（416张准确）,模型2的活检钳状态分类准确率为97.56%（798/818）。模型3判断包含病灶图片654张（626张准确）、不包含病灶图片598张（584张准确）,识别有无病灶的准确率为96.65%（1 210/1 252）,以626张模型3准确判断有病灶图片的数据统计模型3的病灶分割效果,平均IoU为0.86（95%CI：0.85～0.87）。（2）数据集3中：活检钳倾斜角度较小状态下系统病灶尺寸测量的平均绝对误差为0.17 mm（95%CI：0.08～0.28 mm）,平均相对误差为3.77%（95%CI：0.00%～10.85%）;活检钳倾斜角度较大状态下系统病灶尺寸测量的平均绝对误差为0.17 mm（95%CI：0.09～0.26 mm）,平均相对误差为4.02%（95%CI：2.90%～5.14%）。（3）数据集4共纳入59例患者的59个内镜检查视频的780张图片,系统病灶尺寸测量的平均绝对误差为0.24 mm（95%CI：0.00～0.67 mm）,平均相对误差为9.74%（95%CI：0.00%～29.83%）。结论　基于人工智能的自动内镜下病灶尺寸测量系统可以实现内镜下对病灶尺寸的准确测量,有望提高内镜医师对病灶尺寸估计的准确率。     

结肠镜人工智能辅助诊断模型的构建

目的 运用临床结肠镜检查图像和视频,构建结肠镜辅助诊断人工智能深度学习模型。 方法 收集浙江大学医学院附属第二医院内镜中心2014年至2018年的结肠镜图像60余万幅,内镜专家录制大量高质量的结肠镜手术操作视频,以此作为分析数据。训练集样本的每个细分类别图像由6位内镜专家阅片,讨论确定细分类别病变特征,并删减部分模糊和易混淆的分类图像,最终的阅片结果大约为4选1。后再由自主开发的软件逐一标注。采用公信力最高的Google公司TensorFlow平台,对其深度学习算法进行二次开发。 结果 经过机器训练结果与内镜专家结合病理的判断结果进行反复的对比分析,在实验室条件下,该模型对部分疾病（如结肠息肉）的灵敏度为99%。在临床结肠镜操作实验中,该模型对结肠息肉的灵敏度为9830%（4 187/4 259）,特异度为8810%（17 620/20 000）,诊断结肠息肉的总体准确率为9292%[2×（9830%×8810%）/(9830%+8810%)]。对溃疡性结肠炎的灵敏度为7832%（2 671/3 410）,特异度为6706%（13 412/20 000）。单张图像的诊断时长为（05±003）s,此时长为实时应用的时间,包括系统识别、视频图像中文字提示、后台记录和存储三个部分。 结论 本团队研发的人工智能辅助诊断模型能够识别的病灶有结肠息肉、结直肠癌、结直肠隆起性病变、结肠憩室、溃疡性结肠炎等。结肠病辅助诊断模型一方面能够指导肠镜初学者进行肠镜检查,另一方面提高了病灶检出率、并降低漏诊率,而且内镜中心整体的运行效率得以提升,有利于结肠镜检查的质量控制。     

基于深度学习的人工智能技术在早期胃癌诊断中的应用

目的 构建和验证一个用于早期胃癌自动识别的深度学习模型,旨在提高早期胃癌的识别和诊断水平。 方法 从长海医院消化内镜中心数据库选取2014年5月至2016年12月期间5 159张胃镜图像,其中包括早期胃癌1 000张,良性病变及正常图像4 159张。首先选取4 449张图像（其中早期胃癌图像768张,其他良性病变及正常图像3 681张）用于深度学习模型的训练。然后将剩余的710张图像用于模型的验证,同时再交给4名内镜医师进行诊断。最后统计相关结果。 结果 深度学习模型用于早期胃癌诊断的准确率89.4％（635／710）、敏感度88.8％（206／232）、特异度89.7％（429／478）,每张图像的诊断时间为（0.30±0.02）s,均优于相比较的4名内镜医师。 结论 本研究构建的深度学习模型用于早期胃癌的诊断具有较高的准确率、特异度和敏感度,可在胃镜检查中辅助内镜医师进行实时诊断。     

监测上消化道盲区智能内镜影像分析系统的构建及验证

目的　构建监测上消化道盲区的智能内镜影像分析系统,并验证其监测性能。方法　回顾性收集武汉大学人民医院消化内镜中心2016—2020年的上消化道内镜图片87 167张（数据集1）,其中训练集75 551张,测试集11 616张;回顾性收集来自武汉大学人民医院消化内镜中心2016—2020年的咽部图片2 414张（数据集2）,其中训练集2 233张, 测试集181张。分别构建上消化道盲区监测27分类模型（模型1,区分图像为咽部、食管、胃腔等27个解剖学部位）、咽部盲区监测5分类模型（模型2,区分上颚、咽后壁、喉部、左梨状窝、右梨状窝）。基于数据集1、2对上述模型进行训练和图片测试,基于keras框架的EfficientNet‑B4、ResNet50、VGG16模型进行训练。进一步回顾性收集来自武汉大学人民医院消化内镜中心2021年的完整上消化道内镜检查视频30个,在视频中测试模型2盲区监测性能。结果　模型1在图片中识别上消化道27个解剖学部位准确率的横向对比结果显示,EfficientNet‑B4、ResNet50、VGG16在上消化道盲区监测27分类模型的平均准确率分别为90.90%、90.24%、89.22%,其中EfficientNet‑B4模型的表现最优,EfficientNet‑B4模型各个部位监测的准确率介于80.49%~97.80%。模型2在图片中识别咽部5个解剖学部位准确率的横向对比结果显示,EfficientNet‑B4、ResNet50、VGG16在咽部盲区监测5分类模型的平均准确率分别为99.40%、98.56%、97.01%,其中EfficientNet‑B4模型的表现最优,其各个部位监测的准确率介于96.15%~100.00%;模型2在视频中识别咽部5个解剖学部位的总体准确率为97.33%（146/150）。结论　基于深度学习构建的可监测上消化道盲区的智能内镜影像分析系统,耦合了咽部盲区监测及食管、胃腔、十二指肠盲区监测功能,在静止图像及视频中均具有较高识别准确率,有望应用于临床辅助医生实现上消化道视野全覆盖。     

全结肠染色内镜用于大肠癌高风险人群的临床价值评估

目的探讨全结肠靛胭脂染色内镜在大肠癌高风险人群肠镜检查中的应用价值。方法纳入500例准备接受结肠镜检查的大肠癌高危患者作为研究对象,采用随机数字表法将患者随机分为2组,对照组行常规肠镜检查,试验组行全结肠0.2%靛胭脂喷洒染色肠镜检查,主要对比分析2组间息肉、腺瘤患者的发现率。结果研究期间72例因肠道准备差、肠镜检查失败、炎症性肠病、缺血性肠病等原因被剔除,最终共428例纳入数据分析,其中试验组209例、对照组219例,2组基线资料具有可比性。与对照组比较,试验组息肉发现率更高[61.2%（128/209）比43.8%（96/219）,P＜0.001],腺瘤发现率更高[35.9%（75/209）比26.0%（57/219）,P=0.027];且直径＜5 mm息肉发现率试验组明显高于对照组 [27.8%（58/209）比15.5%（34/219）,P=0.002],直径＜5 mm腺瘤发现率试验组明显高于对照组[24.9%（52/209）比12.8%（28/219）,P=0.001]。结论全结肠靛胭脂染色内镜用于大肠癌高风险人群的肠镜检查具有较好的应用价值,有助于提高结肠息肉、腺瘤的检出,尤其是小的结肠病变。     

伺机性筛查新模式在结直肠肿瘤患者一级亲属中的探索性研究

目的探究结直肠病房筛查新模式在结直肠肿瘤患者一级亲属筛查的有效性。 方法采用结直肠肿瘤风险问卷评分、粪便潜血免疫化学检测（FIT）以及粪便多靶点FIT-DNA检测对2019年10月至2021年7月在中国医学科学院肿瘤医院结直肠外科就诊的结直肠癌及进展期腺瘤患者的一级亲属进行检测，根据检测结果将一级亲属进行筛查风险分层以及肠镜检查推荐分类，分析不同分层分类后一级亲属的肠镜依从率与病变检出率。 结果共250名受试者被纳入本研究。总体人群肠镜依从率为38.0%（95/250），肠镜病变检出率为9.5%（9/95）；高风险人群（A类推荐人群）肠镜依从率为78.9%（15/19），肠镜病变检出率为26.7%（4/15）；中风险人群（B类推荐人群）肠镜依从率为61.2%（30/49），肠镜病变检出率为16.7%（5/30）；低风险人群（C类推荐人群）肠镜依从率为27.5%（50/182），肠镜病变检出率为0（0/50）。 结论三种筛查方法联合使用可以高效精准地区分一级亲属的筛查风险，此方案是一个可以在病房开展的有效可行的结直肠肿瘤患者一级亲属人群的伺机性筛查新模式。     

基于深度学习的超声内镜分站和胰腺分割识别系统

目的 尝试构建1个基于深度学习的内镜超声检查（endoscopic ultrasonography,EUS）质量控制系统,并验证其价值。 方法 从武汉大学人民医院消化内镜中心数据库中,回顾性收集2016年12月—2019年12月间的269个EUS检查资料,分为：（1）用于训练模型的训练数据集A,包含205个检查,其中16 305张图像用于分类训练,1 953张图像用于分割训练;（2）用于评估模型性能的测试数据集B,包含44个检查,其中1 606张图像用于分类验证,480张图像用于分割验证;（3）用于内镜医师与模型进行比较的数据集C,包含20个检查,共150张图像。EUS专家（具有10年以上的EUS操作经验）甲和乙通过讨论对训练集A和测试集B、C的所有图像进行分类和标注,其结果用作金标准。EUS专家丙和高年资EUS医师（具有5年以上的EUS操作经验）丁、戊对数据集C中的图像进行分类和标注,其结果用于与深度学习模型进行比较。主要观察指标包括分类的准确率、分割的Dice（F1指数）和一致性分析的Kappa系数。 结果 在测试数据集B中,模型分类的平均准确率为94.1%,胰腺分割的平均Dice为0.826,血管分割的平均Dice为0.841。在数据集C中,模型的分类准确率达到90.0％,专家丙、高年资医师丁和戊分别为89.3％、88.7％和87.3％;模型的胰腺和血管分割Dice系数分别为0.740和0.859,专家丙分别为0.708和0.778,高年资医师丁分别为0.747和0.875,高年资医师戊分别为0.774和0.789,模型与专家的水平相当。一致性分析结果显示,模型与内镜医师之间达成了较好的一致性（Kappa系数分别为：模型与专家丙间0.823、模型与高年资医师丁间0.840、模型与高年资医师戊间0.799）。 结论 基于深度学习的EUS分站和胰腺分割识别系统可以用于胰腺EUS的质量控制,具有与EUS专家相当的分类和分割识别水平。     

Comparison of virtual colonoscopy with conventional colonoscopy in detection of colorectal polyps.

BACKGROUND/AIMS: To determine the sensitivity and specificity of multidetector computed tomography-based virtual colonoscopy for colorectal polyp detection by using conventional colonoscopy as the reference standard. METHODS: 48 patients with high risk for colorectal cancer underwent virtual colonoscopy followed by conventional colonoscopy. Examination results were compared with conventional colonoscopy, which served as the gold standard. RESULTS: Virtual colonoscopy correctly depicted 19 of 22 polyps (sensitivity, 86%) that were detected in conventional colonoscopy. All 4 polyps that were greater than 10 mm in size (100%), 6 of 7 polyps 6-9 mm in size (85%), and 9 of 11 polyps 5 mm in size or smaller (81%) were correctly depicted with virtual colonoscopy. Virtual colonoscopy had an overall sensitivity of 86% and specificity of 98%. CONCLUSION: Multidetector computed tomography-based virtual colonoscopy has excellent sensitivity for the detection of clinically important colorectal polyps.     

结直肠息肉发病风险预测模型构建及其临床价值

目的探讨行结肠镜检查者罹患结直肠息肉的相关危险因素并构建预测模型。 方法回顾性分析2019年1月至2021年10月于南京中医药大学第二附属医院行结肠镜检查者共1 671例的临床资料。根据结肠镜结果将968例结直肠息肉患者纳入息肉组，703名无息肉病变的患者纳入无息肉组。收集患者年龄、性别、身高、体重、吸烟史、饮酒史、实验室检查结果和既往肠镜检查结果等多种因素，分析影响结直肠息肉发生的相关危险因素。应用R语言建立预测结直肠息肉发生风险的列线图模型，用Bootstrap法进行模型内部验证，采用列线图验证曲线及ROC曲线评价列线图的预测性能。 结果息肉组患者年龄（t=151.531，P＜0.001）、男性比例（χ²=50.843，P＜0.001）、长期吸烟史比例（χ²=5.034，P=0.013）、BMI（t=0.813，P＜0.001）、既往息肉史（χ²=8.323，P=0.004）高于无息肉组，差异有统计学意义。多因素Logistic回归模型分析结果显示：年龄，性别，BMI，长期吸烟，既往息肉大小、病理类型及生长位置为结直肠息肉发病的独立危险因素（均P＜0.05）。ROC曲线显示AUC为0.908，敏感性和特异性分别为76.9%和83.2%。 结论年龄，性别，长期吸烟史，BMI，既往息肉大小、病理类型及生长位置为结直肠息肉发病的独立危险因素。该研究所建立的列线图模型具有良好的区分度和准确度，可为直观、个体化地分析结直肠息肉发生风险，甄别高危人群，为临床制订筛查方案提供参考依据。     

Physician-patient and patient-family communication after colonoscopy

Objectives:A personal or family history of colorectal adenomas increases the risk of colorectal cancer (CRC). We aimed to compare physicians' communication with polyp patients vs. non-polyp patients, assess whether polyps or CRC family history were associated with physician-patient communication, and describe patients' disclosure of colonoscopy and polyp diagnosis to their relatives.Methods:Four hundred nine patients completed an online survey regarding physician-patient communication of colonoscopy results, perceived personal and familial risk of polyps and CRC, and disclosure of colonoscopy results to relatives.Results:Six percent of participants reported that their physicians discussed familial risks. Polyp diagnosis and family history predicted physician-patient discussions about familial CRC risks. Polyp diagnosis predicted physician-patient discussions of future surveillance. Twenty-two percent of patients told none of their relatives that they had a colonoscopy. Family history, gender, and education were associated with patient-family communication.Conclusions:There is room for improvement in physician-patient and patient-family communication following colonoscopy.     

Role of computed tomographic colonoscopy of postoperative surveillance in patient with gastric cancer

AIM: To examine the diagnostic yield of colorectal neoplasia at computed tomographic colonoscopy (CTC) as well as the feasibility of contrast enhanced CTC in patients with gastric cancer. METHODS: To examine the incidence of colon polyp we selected postoperative 188 gastric cancer patients, which we refer to as the 'colon polyp survey group'. To examine the feasibility of CTC for early detection of colon cancer or advanced colon adenoma, we selected 47 gastric cancer patients (M:F 29:18, mean age 53.8 years), which we call the 'CT colonoscopy group'. All the 47 patients underwent successive CTC and colonoscopy on the same day. RESULTS: Totally 109 colon polyps were observed from 59 out of 188 gastric cancer patients, the incidence rate of colon polyps in gastric cancer patients being 31.4%. The sensitivity of CTC in detecting individuals with at least 1 lesion of any size was 57.1%, the specificity was 72.7%, the positive predictive value was 47.1%, and the negative predictive value was 71.9%. When the cutoff size was decreased to 6 mm, the sensitivity and specificity were 80.0% and 92.9%, respectively, with positive and negative predictive values of 57.1% and 97.5%, respectively. Only one patient was classified as false negative by virtual colonoscopy. CONCLUSION: The diagnostic yield of colorectal polyp was 31.4% in patients with gastric cancer, and contrast enhanced CTC is an acceptable tool for the detection of synchronous colorectal advanced adenoma andpostoperative surveillance of gastric cancer patients.     

Artificial intelligence technologies for the detection of colorectal lesions: The future is now

Several studies have shown a significant adenoma miss rate up to 35% during screening colonoscopy, especially in patients with diminutive adenomas. The use of artificial intelligence(AI) in colonoscopy has been gaining popularity by helping endoscopists in polyp detection, with the aim to increase their adenoma detection rate(ADR) and polyp detection rate(PDR) in order to reduce the incidence of interval cancers. The efficacy of deep convolutional neural network(DCNN)-based AI system for polyp detection has been trained and tested in ex vivo settings such as colonoscopy still images or videos. Recent trials have evaluated the real-time efficacy of DCNN-based systems showing promising results in term of improved ADR and PDR. In this review we reported data from the preliminary ex vivo experiences and summarized the results of the initial randomized controlled trials.     

A novel summary report of colonoscopy: timeline visualization providing meaningful colonoscopy video information

Purpose

The colonoscopy adenoma detection rate depends largely on physician experience and skill, and overlooked colorectal adenomas could develop into cancer. This study assessed a system that detects polyps and summarizes meaningful information from colonoscopy videos.

Methods

One hundred thirteen consecutive patients had colonoscopy videos prospectively recorded at the Seoul National University Hospital. Informative video frames were extracted using a MATLAB support vector machine (SVM) model and classified as bleeding, polypectomy, tool, residue, thin wrinkle, folded wrinkle, or common. Thin wrinkle, folded wrinkle, and common frames were reanalyzed using SVM for polyp detection. The SVM model was applied hierarchically for effective classification and optimization of the SVM.

Results

The mean classification accuracy according to type was over 93%; sensitivity was over 87%. The mean sensitivity for polyp detection was 82.1%, and the positive predicted value (PPV) was 39.3%. Polyps detected using the system were larger (6.3?±?6.4 vs. 4.9?±?2.5 mm; P?=?0.003) with a more pedunculated morphology (Yamada type III, 10.2 vs. 0%; P?<?0.001; Yamada type IV, 2.8 vs. 0%; P?<?0.001) than polyps missed by the system. There were no statistically significant differences in polyp distribution or histology between the groups. Informative frames and suspected polyps were presented on a timeline. This summary was evaluated using the system usability scale questionnaire; 89.3% of participants expressed positive opinions.

Conclusions

We developed and verified a system to extract meaningful information from colonoscopy videos. Although further improvement and validation of the system is needed, the proposed system is useful for physicians and patients.

     

Computed tomographic colonography compared with colonoscopy in patients at increased risk for colorectal cancer

BACKGROUND & AIMS: To date, computed tomographic (CT) colonography has been compared with an imperfect test, colonoscopy, and has been mainly assessed in patients with positive screening test results or symptoms. Therefore, the available data may not apply to screening of patients with a personal or family history of colorectal polyps or cancer (increased risk). We prospectively investigated the ability of CT colonography to identify individuals with large (>or=10 mm) colorectal polyps in consecutive patients at increased risk for colorectal cancer. METHODS: A total of 249 consecutive patients at increased risk for colorectal cancer underwent CT colonography before colonoscopy. Two reviewers interpreted CT colonography examinations independently. Sensitivity, specificity, and predictive values were determined after meticulous matching of CT colonography with colonoscopy. Unexplained large false-positive findings were verified with a second-look colonoscopy. RESULTS: In total, 31 patients (12%) had 48 large polyps at colonoscopy. This included 8 patients with 8 large polyps that were overlooked initially and detected at the second-look colonoscopy. In 6 of 8 patients, the missed polyp was the only large lesion. With CT colonography, 84% of patients (26/31) with large polyp(s) were identified, paired for a specificity of 92% (200-201/218). Positive and negative predictive values were 59%-60% (26/43-44) and 98% (200-201/205-206), respectively. CT colonography detected 75%-77% (36-37/48) of large polyps, with 9 of the missed lesions being flat. CONCLUSIONS: CT colonography and colonoscopy have a similar ability to identify individuals with large polyps in patients at increased risk for colorectal cancer. The large proportion of missed flat lesions warrants further study.     

Prospective Study Comparing Multislice CT Colonography With Colonoscopy in the Detection of Colorectal Cancer and Polyps

PURPOSE: Multislice CT colonography is an alternative to colonoscopy. The purpose of this study was to compare multislice CT colonography with colonoscopy in the detection of colorectal polyps and cancers. METHODS: Between June 2000 and December 2001, 45 males and 35 females (median age, 68 (29–83) years) with symptoms of colorectal disease were studied prospectively. All patients underwent multislice CT colonography and colonoscopy, and the findings were compared. RESULTS: Colonoscopy was incomplete in 18 (22 percent) patients because of obstructing lesions or technical difficulty, and multislice CT colonography was unsuccessful in 4 (5 percent) because of fecal residue. Colonoscopy was normal in 26 patients and detected 29 colorectal cancers and 33 polyps in 35 patients, diverticulosis in 16 patients, and colitis in 3 patients. Multislice CT colonography identified 28 of 29 colorectal cancers with one false negative and one false positive (sensitivity, 97 percent; specificity, 98 percent; positive predictive value, 96 percent; negative predictive value, 98 percent). Multislice CT colonography identified all 12 polyps measuring 10 mm in diameter (sensitivity, 100 percent), 5 of 6 measuring 6 to 9 mm in diameter (sensitivity, 83 percent), 8 of 15 polyps 5 mm (sensitivity, 53 percent), and false-positive for 8 polyps. The overall sensitivity was 74 percent and specificity 96 percent. The positive predictive value for polyps was 88 percent, and the negative predictive value was 90 percent. Multislice CT colonography also detected 5 of 16 patients with diverticulosis (sensitivity, 31 percent; specificity, 98 percent) and colitis in 2 of 3 patients (sensitivity, 67 percent; specificity, 100 percent). In ten (13 percent) patients, extracolonic findings on multislice CT colonography altered management and included five patients with colorectal liver metastases. In 15 (19 percent) patients, there were incidental findings that did not demand further investigation. CONCLUSIONS: The results from this study indicate that the efficacy of multislice CT colonography in the detection of colorectal cancers and polyps 6 mm is similar to colonoscopy. Multislice CT colonography allows clinical staging of colorectal cancers, outlines the whole length of the colon in obstructing carcinoma when colonoscopy fails, and can identify extracolonic causes of abdominal symptoms.     

人工智能技术在结肠镜退镜速度实时监控中的应用

目的构建一种基于计算机视觉的结肠镜退镜速度实时监控系统,并验证其可行性和性能。方法从武汉大学人民医院消化内镜中心数据库选取2018年5—10月期间的35938张肠镜图片和63个结肠镜检查视频。肠镜图片分成体外/体内/不合格和回盲部/非盲肠两个数据集,分别从第一个、第二个数据集中选取3594张和2000张图片用于深度学习模型的测试,其余图片用于训练模型;选取3个结肠镜检查视频资料评价实时监控系统自动监控退镜速度的可行性,剩余60个结肠镜检查视频资料用于评估实时监控系统的性能。结果深度学习模型对于结肠镜检查图片分类识别体外/体内/不合格图片的准确率分别为90.79%(897/988)、99.92%(1300/1301)、99.08%(1293/1305),总体准确率为97.11%(3490/3594);分类识别回盲部/非盲肠图片的准确率分别为96.70%(967/1000)、94.90%(949/1000),总体准确率为95.80%(1916/2000)。在其可行性评价方面,3个结肠镜视频资料显示退镜速度与图片处理间隔时间呈线性关系,提示该监控系统可在结肠镜退出过程中自动监控退镜速度。在其性能评价方面,结肠镜退镜速度实时监控系统正确预测了所有60个肠镜检查的开始时间和结束时间,分析显示结肠镜平均退镜速度和退镜时间呈明显负相关(R=-0.661,P<0.001),退镜时间不足5 min、5~6 min和超过6 min视频的平均退镜速度的95%置信区间分别为43.90~49.74、40.19~45.43和34.89~39.11,故将39.11设为安全退镜速度,将45.43设为预警退镜速度。结论构建的结肠镜退镜速度实时监控系统可用于实时监控结肠镜退镜速度,可在结肠镜检查中辅助内镜医师进行实时监测,以提高结肠镜检查质量。     

Prospective comparison of contrast enhanced CT colonography and conventional colonoscopy for detection of colorectal neoplasms in a single institutional study using second-look colonoscopy with discrepant results

BACKGROUND: Colorectal cancer is the second leading cause of death from cancer in Western countries. Early detection by colorectal cancer screening can effectively cut its mortality rate. CT colonography represents a promising, minimally invasive alternative to conventional methods of colorectal carcinoma screening. AIMS: The purpose of this prospective single institutional study was to compare the abilities of routine clinical CT colonography and conventional colonoscopy to detect colorectal neoplasms using second-look colonoscopy to clarify discrepant results. PATIENTS AND METHODS: CT colonography was performed in 100 symptomatic patients using contrast enhanced multidetector CT followed by conventional colonoscopy on the same day. If results were discrepant, a second-look colonoscopy was performed after unblinding. CT colonographic findings were compared with those of conventional colonoscopy. RESULTS: Conventional colonoscopy found 122 colorectal neoplasms in 49 patients. The overall sensitivity of CT colonography at detecting patients with at least one polyp 6 mm or larger was 76% and its specificity was 88%. Its by-patient sensitivity for polyps 10 mm or larger was 95% and its specificity was 98%. By-polyp sensitivities were 71% for polyps 10 mm or larger, and 61% for polyps 6 mm or larger. A second-look colonoscopy was performed in 19 patients and two initial false-positive findings of CT colonography were reclassified as true-positive. For conventional colonoscopy, this produced a by-polyp sensitivity of 94% for detection of lesions 6 mm and larger. CONCLUSIONS: CT colonography had both a high by-patient sensitivity and specificity for detection of clinically important colorectal neoplasms 10 mm or larger. This suggests that CT colonography has the potential to become a valuable clinical screening method for colorectal neoplasms.