基于深度学习的人工智能诊断模型在食管早癌内窥镜筛查中的研究

Study on the application of artificial intelligence diagnostic model based on deep learning in endoscopic screening of early esophageal cancer

目的:研究基于深度学习的人工智能(AI)诊断模型应用于食管早癌内窥镜筛查。方法:利用Inception ResNet V2 CNN模型构建深度学习模型,并在Tensor Flow1.9.0框架下进行训练、验证及测试。选取医院收治的80例食管癌患者资料,另选同期在医院进行内窥镜检查的80例食管其他病变患者资料,收集160例患者内窥镜检查时获得的传统白光成像技术图像及窄带成像图像,按照7∶3的比例分为训练组(112例)和验证组(48例),建立基于深度学习的AI诊断模型,收集两组临床基本资料。分析食管早癌内窥镜下特征。以病理学检查结果为“金标准”,分析传统白光成像技术及窄带成像技术对食管早癌的诊断结果。观察不同性质食管病变患者镜下检查结果。分析基于卷积神经网络(CNN)的深度学习方法构建的早期食管癌AI诊断模型与内窥镜医师诊断结果之间的差异。采用受试者工作特征(ROC)曲线下面积(AUC)分析AI诊断模型与低资历内窥镜医师、高资历内窥镜医师诊断的效能。结果:训练组患者基于传统白光成像技术图像及窄带成像图像的病灶表面不光滑占79.46%(89/112),具有黏液附着占72.32%(81/112),伴有糜烂占64.29%(72/112),黏膜颜色发红占90.18%(101/112),上皮微细结构与腺管开口不规则或消失占90.18%(101/112),黏膜下微血管不规则或消失占96.43%(108/112)。经病理学检测结果发现食管早癌患者21例,经传统白光成像技术联合窄带成像技术检测发现食管早癌患者36例。验证组诊断显示,AI诊断模型及低资历内窥镜医师、高资历内窥镜医师诊断的曲线下面积分别为0.932、0.734和0.916;灵敏度分别为91.25%、42.38%和62.26%;特异度分别为97.74%、93.57%和98.14%;准确率分别为94.28%、87.74%和93.41%。AI诊断模型的AUC明显高于低资历内窥镜医师的诊断效能,其差异有统计学意义(Z=2.856,P<0.05)。结论:基于CNN的深度学习方法构建的食管早癌AI诊断模型对食管早癌的诊断具有一定的应用价值,其临床应用对于食管早癌的早期发现及时治疗,延长患者生存时间,改善患者预后意义重大。

Objective:To explore the application of artificial intelligence(AI)diagnostic model based on deep learning in endoscopic screening of early esophageal cancer.Methods:A total of 80 patients with esophageal cancer admitted to hospital were selected,and the data of other 80 patients with other esophageal lesions who underwent endoscopic examination during the same time were selected.The images,that obtained from conventional white light imaging and narrowband imaging,of 160 patients during endoscopic examination were collected.They were divided into a training group(112 cases)and a validation group(48 cases)in a 7∶3 ratio to construct AI diagnostic model based on deep learning,and the clinically basic data of the two groups were collected.The endoscopic features of early esophageal cancer were analyzed.Based on the pathological examination results as the"gold standard",the diagnostic results of the conventional white light imaging technique and narrowband imaging technique for early esophageal cancer were further analyzed.The results of endoscopic examination in patients with different properties of esophageal lesions were observed.And then,the difference between the constructed AI diagnosis model of early esophageal cancer based on the deep learning method of convolutional neural network(CNN)and the diagnostic results of endoscopic physicians was analyzed.The diagnostic efficiencies of the AI diagnostic model,the endoscopic physician with low qualification and the endoscopic physician with high qualification were analyzed by using the area under the curve(AUC)of receiver operating characteristics(ROC).Results:In all of patients,the proportion of the surface of the lesion without smooth in the images based on the conventional white light imaging technique and narrowband imaging technique was 79.46%(89/112),and that with mucus adhesion was 72.32%(81/112),and that with erosion was 64.29%(72/112),and that with red color mucosal was 90.18%(101/112),and that with irregular or disappearing micro vessels was 96.43%(108/112).A total of 21 patients with early esophageal cancer were found by pathological examination,and the conventional white light imaging technique combined with narrowband imaging technique confirmed 36 patients were early esophageal cancer.The diagnosis of validation group showed that the AUCs of the AI diagnostic model,the endoscopic physician with low qualification and the endoscopic physician with high qualification were respectively 0.932,0.734 and 0.916,and the sensitivities of them were respectively 91.25%,42.38%and 62.26%,and the specificities of them were respectively 97.74%,93.57%and 98.14%,and the accuracies of them were respectively 94.28%,87.74%,and 93.41%,respectively.The AUC of the diagnostic efficiency of the AI diagnostic model was significantly higher than that of the endoscopic physician with low qualification(Z=2.856,P<0.05).Conclusion:For early esophageal cancer,the AI diagnostic model of deep learning method based on the CNN has a certain of application value in diagnosing early esophageal cancer,which clinical application is of great significance in early detection and timely treatment of early esophageal cancer,prolonging survival time of patient and improving prognosis of patient.