Objective: To construct the diagnostic model of superficial esophageal squamous cell carcinoma (ESCC) and precancerous lesions in endoscopic images based on the YOLOv5l model by using deep learning method of artificial intelligence to improve the diagnosis of early ESCC and precancerous lesions under endoscopy. Methods: 13, 009 endoscopic esophageal images of white light imaging (WLI), narrow band imaging (NBI) and lugol chromoendoscopy (LCE) were collected from June 2019 to July 2021 from 1, 126 patients at the Cancer Hospital, Chinese Academy of Medical Sciences, including low-grade intraepithelial neoplasia, high-grade intraepithelial neoplasia, ESCC limited to the mucosal layer, benign esophageal lesions and normal esophagus. By computerized random function method, the images were divided into a training set (11, 547 images from 1, 025 patients) and a validation set (1, 462 images from 101 patients). The YOLOv5l model was trained and constructed with the training set, and the model was validated with the validation set, while the validation set was diagnosed by two senior and two junior endoscopists, respectively, to compare the diagnostic results of YOLOv5l model and those of the endoscopists. Results: In the validation set, the accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) of the YOLOv5l model in diagnosing early ESCC and precancerous lesions in the WLI, NBI and LCE modes were 96.9%, 87.9%, 98.3%, 88.8%, 98.1%, and 98.6%, 89.3%, 99.5%, 94.4%, 98.2%, and 93.0%, 77.5%, 98.0%, 92.6%, 93.1%, respectively. The accuracy in the NBI model was higher than that in the WLI model (P<0.05) and lower than that in the LCE model (P<0.05). The diagnostic accuracies of YOLOv5l model in the WLI, NBI and LCE modes for the early ESCC and precancerous lesions were similar to those of the 2 senior endoscopists (96.9%, 98.8%, 94.3%, and 97.5%, 99.6%, 91.9%, respectively; P>0.05), but significantly higher than those of the 2 junior endoscopists (84.7%, 92.9%, 81.6% and 88.3%, 91.9%, 81.2%, respectively; P<0.05). Conclusion: The constructed YOLOv5l model has high accuracy in diagnosing early ESCC and precancerous lesions in endoscopic WLI, NBI and LCE modes, which can assist junior endoscopists to improve diagnosis and reduce missed diagnoses.
目的： 以人工智能深度学习的方法，构建基于YOLOv5l模型的内镜图像早期食管癌及癌前病变的识别模型，以提高内镜下早期食管癌及癌前病变的诊断水平。 方法： 收集2019年6月至2021年7月中国医学科学院肿瘤医院1 126例患者的白光成像(WLI)、窄带光成像(NBI)和卢戈液染色(LCE)的内镜食管图像13 009幅，包括低级别上皮内瘤变、高级别上皮内瘤变、限于黏膜层的食管鳞癌、良性食管病变及正常食管。通过计算机随机函数方法，分为训练集(1 025例患者的11 547幅图像)和验证集(101例患者的1 462幅图像)。以训练集训练、构建YOLOv5l模型，以验证集验证该模型，同时由2名高年资和2名低年资内镜医师分别对验证集进行诊断，比较YOLOv5l模型与内镜医师的诊断结果。 结果： 在验证集中，YOLOv5l模型在WLI、NBI和LCE模式下诊断早期食管癌及癌前病变的准确度、灵敏度、特异度、阳性预测值(PPV)和阴性预测值(NPV)分别为96.9%、87.9%、98.3%、88.8%和98.1%, 98.6%、89.3%、99.5%、94.4%和98.2%，93.0%、77.5%、98.0%、92.6%和93.1%。NBI模式下的准确度高于WLI模式(P<0.05)，LCE模式下的准确度低于WLI(P<0.05)。YOLOv5l模型在WLI、NBI和LCE模式下诊断早期食管癌及癌前病变的准确度与2位高年资内镜医师(分别为96.9%、98.8%和94.3%，97.5%、99.6%和91.9%；均P>0.05)相当，但明显高于2位低年资内镜医师(分别为84.7%、92.9%和81.6%，88.3%、91.9%和81.2%；均P<0.05)。 结论： 所构建的YOLOv5l模型在内镜WLI、NBI和LCE模式下诊断早期食管癌及癌前病变中具有较高的准确度，可以辅助低年资内镜医师提高诊断水平、减少漏诊。.
Keywords: Artificial intelligence; Diagnosis; Endoscope; Esophageal neoplasms, early; Precancerous lesions.