【作者】 刘玲；

【导师】 凌亭生；

【作者基本信息】 南京中医药大学 ， 中西医结合临床（专业学位）， 2024， 硕士

【摘要】 目的:1.本研究采用CNN的Unet类网络,Transformer类型的网络,以及CNN和视觉Transformer的混合网络开展胃黏膜肠上皮化生内镜图像分割实验,建立可视化胃黏膜肠上皮化生诊断模型。2.本研究利用智能舌诊采集及分析仪器收集胃黏膜肠上皮化生患者的舌象图片,并进行客观分析,为中医舌诊理论辨治胃黏膜肠上皮化生提供客观依据,为中西医结合临床辨治胃黏膜肠上皮化生提供新的思路和方法。方法:1.胃黏膜肠上皮化生分割模型研究所有研究数据来自于2022年12月1日至2023年8月在南京中医药大学附属医院江苏省中医院紫东院区消化内镜中心收集经病理证实后的胃黏膜肠上皮化生患者联动成像模式(LCI)下拍摄的胃镜图片,本研究所采集图片均在富士公司ELUXEO 7000系统下所拍摄。使用X-AnyLabeling2.0软件对所采集的内镜图片进行标注工作,并过Excel软件按样本编号对数据将数据集随机划分为训练集和测试集,比例为9:1。本研究使用四个Nvidia GTX 4090 GPU,并利用PyTorch 2.0来训练深度学习网络。此外,本研究还使用当下前沿算法建立胃黏膜肠上皮化生分割模型,这些算法包括CCNet、DANet、DeepLabV3+、K-net、Mask2Former、PSAnet、SegFormer、UNet、PVT。最后使用模型分割准确度、精确度、召回率、f1分数和IoU等指标评价各算法模型分割的性能。2.胃黏膜肠上皮化生分布特征的客观化研究应用智能化舌象采集仪器拍摄2023年4月27日至2023年08月2日期间于江苏省中医院紫东院区消化内镜中心行胃镜检查且经病理诊断为胃黏膜肠上皮化生的患者的舌象图片,并运用人工智能舌诊分析仪对所采集的舌象图片进行统一分析。同时收集纳入患者的一般资料信息、胃镜图片及内镜黏膜表现。最后运用SPSS26.0软件进行数据统计分析,分析胃黏膜肠上皮化生患者的舌象分布特点以及胃黏膜肠上皮化生患者内镜下表现与舌象分布的相关性。结果:1.胃黏膜肠上皮化生分割模型研究本研究在2022年12月1日至2023年8月期间共收集522名胃黏膜肠上皮化生患者胃镜图片。经质控后筛选出符合研究要求的内镜图片共计1776张,分为训练集与测试集。分割模型的准确度、召回率、F1分数和IoU分别为88.9%、89.57%、78.92%、65.17%,本研究提出的算法模型在所有指标上都优于目前已公开发表的开源模型。并且该模型在4090 GPU上实现了 36 FPS的预测速度,能够实时处理内镜视频并自动划定胃黏膜肠上皮化生区域。2.胃黏膜肠上皮化生内镜下表现与舌象的相关性研究本研究共纳入287名经病理证实为胃黏膜肠上皮化生的患者,其中男性158例(56.80%),女性129例(43.20%),男女比为1.22:1。其中胃黏膜肠上皮化生的高危发病年龄段为59-68岁(33.45%)。EGGIM分级中,非广泛性胃黏膜肠上皮化生132例(45.99%),广泛性胃黏膜肠上皮化生155例(54.01%)。木村-竹本分型中,闭合型269例(96.8%),开放型24例(3.2%)。Hp感染方面,阳性患者65例(23.4%),胃癌风险的内镜表现方面,具有皱襞肿大的有44例(15.82%),弥漫性发红的有36例(12.9%),并发早期胃癌的患者有22例(7.9%)。本次研究发现GIM患者中常见的4种舌色比例占比依次为淡红舌270例(94.1%)>红舌 10 例(3.5%)>淡白舌 9 例(2.1%)>绛舌 1 例(0.3%),本次研究共统计了3种苔色,其占比分别为:淡黄苔181例(63.1%)>白苔105例(36.6%)>黄苔1例(0.3%)。本次研究共统计了 6种苔质,其占比依次为:厚腻苔224例(78.1%)>厚腐苔41例(14.3%)>薄苔11例(3.8%)>厚苔8例(2.8%)>薄腻苔3例(1.0%)。苔质中还包含了剥落苔,其中存在剥落苔患者共10例(3.5%)。本次研究共统计了种6种特殊舌形,包括齿痕舌、老嫩舌、点刺舌、裂纹舌、瘀点舌、胖瘦舌,其中齿痕舌共计225例(78.4%),老舌96例(33.5%),嫩舌27例(9.4%)。裂纹舌72例(25.1%),点刺舌29例(10.1%),瘀点舌 74 例(25.8%)。胖舌 8 例(2.8%),瘦舌 9(3.1%)。此外,研究还统计了胃黏膜肠上皮化生患者舌苔津液分布与舌下脉络分布结果,在舌苔津液方面,苔滑的患者67例(23.3%),苔润的患者209例(72.8%),苔燥的患者11例(3.9%)。具有舌下有瘀象表现的患者共86例(30%)。结论:(1)本研究所建立的胃黏膜肠上皮化生分割模型可以辅助内镜医生在LCI模式下进行GIM的实时诊断,提高其准确率及工作效率。(2)本研究在对胃黏膜肠上皮化生患者舌象信息进行客观分析后认为该类患者主要以淡红舌、厚腻苔,伴有齿痕舌为主要特征。(3)胃黏膜肠上皮化生患者舌象特征变化在内镜下胃黏膜肠上皮化生范围、萎缩范围中无显著差异,但分布频率有一定趋向性。(4)在胃黏膜肠上皮化生伴有Hp阳性患者中,尤其是内镜下具有弥漫性发红表现时,患者出现黄苔、点刺舌的频率更高,具有统计学差异(p<0.05)更多还原

【Abstract】 Objective:The study aims to enhance the diagnostic accuracy of gastrointestinal metaplasia(GIM)by employing Convolutional Neural Network(CNN)Net class,Transformer-type networks,and a hybrid of CNN and visual Transformer to establish a visual GIM diagnostic model.Additionally,tongue images of GIM patients are collected using an intelligent tongue diagnosis collection instrument,and a comprehensive and objective analysis is performed using a tongue image intelligent analysis instrument.This approach provides a more standardized and objective tongue diagnosis in identifying and treating GIM based on traditional Chinese medicine theories and offers a novel thought process and methodology for the clinical identification and treatment of GIM in both Chinese and Western medicine.Methods:1.Research on gastrointestinal metaplasia segmentationAll study data derive from Linked Color Imaging(LCI)gastroscopy images of pathologically confirmed GIM patients collected at the the Gastrointestinal Endoscopy Center of the Zidong Hospital District of Jiangsu Provincial Hospital of Traditional Chinese Medicine from December 1,2022,to August 2023.All images are captured under the ELUXEO 7000 system of Fujifilm.The collected endoscopic images are labeled using X-AnyLabeling 2.0 software,and the dataset is randomly divided into a training set and a test set in a 9:1 ratio using Excel software.Four Nvidia GTX 4090 GPUs are used,and PyTorch 2.0 is utilized to train the deep learning network.This study also utilizes popular algorithms to build a GIM segmentation model,including CCNet,DANet,DeepLabV3+,K-net,Mask2Former,PSAnet,SegFormer,UNet,and PVT.The performance of each algorithm model segmentation is evaluated using accuracy,precision,recall,F1-score,and IoU.2.Objective research tongue characteristics of gastrointestinal metaplasiaIntelligent tongue collection instruments are used to collect tongue images of patients who undergo gastroscopy and are diagnosed with GIM at the the Gastrointestinal Endoscopy Center of the Zidong Hospital District of Jiangsu Provincial Hospital of Traditional Chinese Medicine from April 27,2023,to August 2,2023.These images are uniformly analyzed using an artificial intelligence tongue diagnostic analyzer.General data information,gastroscopic images,and endoscopic mucosal manifestations of the included patients are also collected.Finally,the characteristics of tongue image distribution in GIM patients and the correlation between endoscopic manifestations and tongue image are analyzed using SPSS26.0.Results:1.Research on gastrointestinal metaplasia segmentationA total of 522 gastroscopic images of GIM patients are collected between December 1,2022,and August 2023.Following quality control,1776 endoscopic images that meet the study requirements are obtained and divided into 178 cases in the training set and 1598 in the test set.The accuracy,recall,F1-score,and IoU of the segmentation model are 88.9%,89.57%,78.92%,and 65.17%,respectively.The proposed algorithmic model outperforms the currently published open-source models in all metrics.Moreover,the model achieves a prediction speed of 36 FPS on 4090 GPUs,enabling real-time processing of endoscopic videos and automatic delineation of mucosal intestinal epithelialized areas.2.Objective research tongue characteristics of gastrointestinal metaplasiaA total of 287 patients with pathologically confirmed gastrointestinal metaplasia are enrolled in this research,of which 158(56.80%)are males and 129(43.20%)are females,with a male to female ratio of 1.22:1.Gastrointestinal metaplasia is found in patients with a maximum age of 80 years and a minimum age of 35 years.The mean age of onset of GIM is 57.99±10.13 years,with the mean age of onset for males being 57.29± 10.17 years and females(57.91 ±9.58 years).Age stratification of the patients shows that the high-risk age group for GIM is 59-68 years(33.45%).In EGGIM classification,132 cases(45.99%)are non-extensive GIM,155 cases(54.01%)are extensive GIM and the highest number of 3 points is 22.64%.In the KimuraTakemoto classification,there are 269 cases(96.8%)of closed type and 24 cases(3.2%)of open type.A total of 6 special tongue shapes are counted in this study,including dentate tongue,old and young tongue,punctate tongue,fissured tongue,petechiae tongue,and fat and thin tongue.Among 287 GIM patients included in this study,dentate tongue is found in a total of 225 cases(78.4%),old tongue is found in 96 cases(33.5%),and young tongue is found in 27 cases(9.4%).Cracked tongue is found in 72 cases(25.1%),punctured tongue in 29 cases(10.1%),and petechial tongue in 74 cases(25.8%).Fat tongue is found in 8 cases(2.8%)and thin tongue in 9(3.1%).In addition,the study also counts the results of the distribution of the tongue moss and fluid and the distribution of the sublingual veins.In terms of the tongue moss and fluid,there are 67 cases(23.3%)of patients with slippery moss,209 cases(72.8%)of patients with moist moss,and 11 cases(3.9%)of patients with dry moss.Patients with the manifestation of stasis under the tongue total 86 patients(30%).Conclusion:1.The GIM model developed in this study can assist endoscopists in real-time endoscopic diagnosis of GIM,improving the accuracy and efficiency of diagnosing GIM under LCI mode.2.After a comprehensive and objective analysis of the tongue information of GIM patients,it was found that GIM patients were primarily characterized by a pale red tongue and thick,greasy moss,accompanied by teeth marks.3.There was no significant difference in the distribution of tongue images in patients with GIM concerning the endoscopic range of GIM and the range of atrophy,but there was some convergence in the frequency of distribution.4.In GIM patients with Hp infection,especially when endoscopically exhibiting diffuse reddening manifestations,patients had a higher frequency of yellow moss and punctate tongue,which was statistically significant(P<0.05).更多还原