AI-Assisted Microendoscopy for the Early Detection of Esophageal Cancer

摘要：食道癌是全球癌症相关死亡的第六大常见原因。而当 食道鳞状细胞瘤(ESCN)在全球范围内是一个重大的负担，在某些情况下，这些疾病得不到足够的服务 地理区域(南美、东非、伊朗东部、中国北部)的 由于缺乏内窥镜筛查能力而导致的发病率和死亡率。而内窥镜检查与卢戈尔氏病 染色内窥镜或“数字”染色内窥镜对筛查显示出高敏感性(&gt；95%)，特异性为 由于炎症性区域的混淆，结果很差(60%)和假阳性比比皆是。因此，标准的 护理内窥镜检查产生许多不必要的活检，增加了内窥镜筛查和治疗的风险和成本 监视系统。 在我们正在进行的R01项目中，我们开发并验证了一种移动式高分辨率显微内窥镜(MHRME) 用于筛查和监测ESCN。尽管≥延迟了2年，这特别影响了 中国网站，我们完成了：(1)一项随机对照临床试验(美国和中国；n=918) 接受食道鳞状细胞瘤筛查或监测的患者的视觉解读， (2)用于肿瘤图像自动检测的深度学习软件算法，以及(3)试点试验(n=41) 软件辅助的mHRME在巴西的应用。这项试验显示定性(视觉)的特异性更高 由专家翻译，但不是新手，在监视ARM(100%比19%，p&lt；0.05)。在筛选臂中， 诊断率(肿瘤活检/总活检)增加了3.6倍(8%至29%)；16%的患者 正确地避免了任何活检；18%的患者改变了临床计划。在一项单臂先导研究中，我们还评估了 41名巴西受试者的基于人工智能的移动HRME(AI-mHRME)。这项研究(2022年1月) 证实定量解释(AI-mHRME)使诊断效率提高了一倍，改善了内窥镜医生的 信心，并有显著的临床影响(改变临床计划在%)。我们最初的深度学习算法 敏感度/特异度为100%/85%。参与的临床医生一致表示，他们支持人工智能引导的 但对其执行情况表示关切。 在这一竞争更新中，我们将基于这些宝贵的全球数据来优化AI-mHRME并进行评估 它在不同种族和社会经济背景的人群中的临床影响和实施潜力 美国和巴西。将使用利益相关者参与的方法来评估障碍、可接受性、 在ESCN管理中使用AI-mHRME的适宜性和可行性，并确定上下文因素 影响领养。获得的数据将有助于实施和传播创新的、人工智能辅助的 在不同人群和其他癌症中的癌症筛查策略。

Abstract: Esophageal cancer is the 6th most common cause of cancer-related mortality worldwide. While esophageal squamous cell neoplasia (ESCN) carries a significant global burden, those in certain underserved geographic regions (South America, eastern Africa, eastern Iran, northern China) have particularly high incidence and mortality rates due to lack of endoscopic screening capacity. While endoscopy with Lugol’s chromoendoscopy or “digital” chromoendoscopy has shown high sensitivity (>95%) for screening, specificity is poor (<60%) and false-positive results abound due to confounding inflammatory areas. As a result, standard of care endoscopy produces many unnecessary biopsies, increasing risk and cost of endoscopic screening and surveillance. In our ongoing R01 project, we developed and validated a mobile, high-resolution microendoscope (mHRME) for screening and surveillance of ESCN. Despite ≥2 years of COVID delays, which especially impacted the Chinese sites, we completed: (1) a randomized, controlled clinical trial (USA and China; n=918) of mHRME with visual interpretation in patients undergoing screening or surveillance for esophageal squamous cell neoplasia, (2) deep-learning software algorithms for automated detection of neoplastic images, and (3) a pilot trial (n=41) of the software-assisted mHRME in Brazil. The trial revealed higher specificity for qualitative (visual) interpretation by experts but not novices, in surveillance arm (100% vs. 19%, p<0.05). In the screening arm, diagnostic yield (neoplastic biopsies/total biopsies) increased 3.6 times (8 to 29%); 16% of patients were correctly spared any biopsy; and 18% had a change in clinical plan. In a single-arm pilot study, we also evaluated an artificial intelligence-based mobile HRME (AI-mHRME) in 41 Brazilian subjects. This study (January 2022) confirmed that quantitative interpretation (AI-mHRME) doubled diagnostic yield, improved endoscopists’ confidence, and had significant clinical impact (change of clinical plan in 64%). Our initial deep-learning algorithm had a sensitivity/specificity of 100%/85%. Participating clinicians uniformly said they favored an AI-guided approach but expressed concerns about its implementation. In this competing renewal, we will build on this valuable global data to optimize an AI-mHRME and evaluate its clinical impact and implementation potential in ethnically and socioeconomically diverse populations in the USA and Brazil. A stakeholder-engaged approach will be used to evaluate barriers, acceptability, appropriateness, and feasibility of using AI-mHRME in ESCN management and to determine contextual factors influencing adoption. Data obtained will facilitate implementation and dissemination of innovative, AI-assisted cancer screening strategies in diverse populations and other cancers.