内镜人工智能诊断辅助系统对胃局灶性病变检出的应用（含视频）

Application of an artificial intelligence⁃assisted endoscopic diagnosis system to the detection of focal gastric lesions (with video)

摘要:

目的 构建一个基于YOLO v3算法的实时人工智能诊断辅助系统，并评估其在胃镜检查中对胃局灶性病变检出的能力。方法 回顾性收集武汉大学人民医院消化内镜中心2019年6—11月胃镜检查的白光内镜图像5 488张（有、无胃局灶性病变的图像分别为2 733张、2 755张）及2020年5—6月期间92例行胃镜检查的受试者视频资料中288 168个清晰胃帧用于人工智能辅助系统测试；前瞻性收集2020年7月6日—11月27日及2021年5月6日—8月2日于武汉大学人民医院消化内镜中心接受胃镜检查的3 997例受检者的视频资料用于评估人工智能辅助系统在实际临床应用中的性能。当人工智能辅助系统识别到异常病灶时，以蓝色方框圈出病灶进行提示。对人工智能辅助系统识别胃局灶性病变的能力及其出现假阳性和假阴性的频率和原因进行统计分析。结果 图像测试集中，人工智能辅助系统“提示病灶”的准确率、灵敏度 、特异度、阳性预测值及阴性预测值分别为92.3%（5 064/5 488）、95.0%（2 597/2 733）、89.5%（2 467/2 755）、90.0%（2 597/2 885）和94.8%（2 467/2 603）。视频测试集中，人工智能辅助系统“提示病灶”的准确率、灵敏度、特异度、阳性预测值及阴性预测值分别为95.4%（274 792/288 168）、95.2%（109 727/115 287）、95.5%（165 065/172 881）、93.4%（109 727/117 543）和96.7%（165 065/170 625）。临床应用中，人工智能辅助系统对胃局灶性病变的检出率为93.0%（6 830/7 344）。共漏检胃局灶性病变514处，主要原因为微小糜烂灶（48.8%，251/514）、微小黄斑瘤（22.8%，117/514）和小息肉（21.4%，110/514）。平均每例上消化道内镜检查中，人工智能辅助系统的假阳性个数为2（1， 4）个，主要原因为正常黏膜皱襞（50.2%，5 635/11 225）、气泡和黏液（35.0%，3 928/11 225）、胃底液体（9.1%，1 021/11 225）。结论 在胃镜检查过程中应用人工智能辅助系统有助于胃局灶性病变的检出。

Abstract:

Objective To construct a real-time artificial intelligence (AI)-assisted endoscopic diagnosis system based on YOLO v3 algorithm, and to evaluate its ability of detecting focal gastric lesions in gastroscopy. Methods A total of 5 488 white light gastroscopic images (2 733 images with gastric focal lesions and 2 755 images without gastric focal lesions) from June to November 2019 and videos of 92 cases (288 168 clear stomach frames) from May to June 2020 at the Digestive Endoscopy Center of Renmin Hospital of Wuhan University were retrospectively collected for AI system test. A total of 3 997 prospective consecutive patients undergoing gastroscopy at the Digestive Endoscopy Center of Renmin Hospital of Wuhan University from July 6, 2020 to November 27, 2020 and May 6, 2021 to August 2, 2021 were enrolled to assess the clinical applicability of AI system. When AI system recognized an abnormal lesion, it marked the lesion with a blue box as a warning. The ability to identify focal gastric lesions and the frequency and causes of false positives and false negatives of AI system were statistically analyzed. Results In the image test set, the accuracy, the sensitivity, the specificity, the positive predictive value and the negative predictive value of AI system were 92.3% (5 064/5 488), 95.0% (2 597/2 733), 89.5% (2 467/ 2 755), 90.0% (2 597/2 885) and 94.8% (2 467/2 603), respectively. In the video test set, the accuracy, the sensitivity, the specificity, the positive predictive value and the negative predictive value of AI system were 95.4% (274 792/288 168), 95.2% (109 727/115 287), 95.5% (165 065/172 881), 93.4% (109 727/117 543) and 96.7% (165 065/170 625), respectively. In clinical application, the detection rate of local gastric lesions by AI system was 93.0% (6 830/7 344). A total of 514 focal gastric lesions were missed by AI system. The main reasons were punctate erosions (48.8%, 251/514), diminutive xanthomas (22.8%, 117/514) and diminutive polyps (21.4%, 110/514). The mean number of false positives per gastroscopy was 2 (1, 4), most of which were due to normal mucosa folds (50.2%, 5 635/11 225), bubbles and mucus (35.0%, 3 928/11 225), and liquid deposited in the fundus (9.1%, 1 021/11 225). Conclusion The application of AI system can increase the detection rate of focal gastric lesions.