Development and Validation of an Artificial-Intelligence-enabled Portable Colposcopy Device for Optimizing Triage Alternatives for HPV-based Cervical Cancer Screening

摘要 宫颈癌是世界范围内妇女死亡的第二大原因。令人担忧的是，85%的死亡发生在 低收入和中等收入国家（LMIC），因为它们缺乏基于细胞学的 筛查、转诊阴道镜诊断和专家医生，这些都大大减少了疾病 高收入国家（HICs）。高度敏感的人乳头瘤病毒（HPV）检测已经有效 在直接与治疗相结合的情况下，降低宫颈癌的发病率和死亡率;然而， 大多数感染HPV的女性没有宫颈癌前病变，这使得HPV检测成为一种不好的分诊检测， 过度治疗会带来风险， 出血和不育。 阴道镜检查后活检，首选的分流 在大多数LMIC环境中，由于阴道镜和病理设施的成本， 处理和解释活检结果。更糟糕的是，在LMIC环境中， 采用多次就诊模式进行子宫颈癌筛查。醋酸目视检查（VIA），世界 卫生组织建议在HPV检测后进行分流检测，其敏感性和特异性差异很大 这取决于供应商的培训水平。在本提案中，我们提出了一个单次访问模型， 宫颈癌的诊断和治疗需要两个主要的技术工具， 实现这个模型：一种低成本的方法来执行子宫颈成像和机器学习算法， 在没有提供者的情况下自动诊断。我们以前开发了袖珍阴道镜， 已经显示出与标准阴道镜检查的高度一致性，其成本仅为标准阴道镜检查的一小部分，并在数千例 几乎每个大陆的女性。我们现在正在开发一种最先进的卷积 神经网络（CNN），称为阴道镜自动风险评估（CARE），用袖珍阴道镜训练 图像来自动化诊断过程。我们目前的原型算法已经非常成功， 回顾性地对袖珍阴道镜图像中的宫颈癌前病变进行分类。我们对这一提案的目标是 四方面：1）使用> 10，000个国家癌症研究所，改善和推广Pocket CARE的性能 (NCI)标准阴道镜检查图像; 2）生成合成图像以解决由于环境因素引起的域偏移 以及不同临床研究中心之间的人员变更; 3）将CARE算法嵌入到我们现有的软件中 为了使用袖珍阴道镜进行高质量的图像采集，以进行自动诊断，4）验证 在肯尼亚的基苏穆进行的一项临床研究中，Pocket CARE的前瞻性表现， 最终将被采纳。该提案的交付物将是一个经过充分验证的Pocket CARE软件 可根据特定地点的文化背景和基础设施扩展到不同的临床场景 以及Pocket CARE与其他公开可用算法和标准RI护理的比较有效性。

Abstract Cervical cancer is the second leading cause of death for women worldwide. Alarmingly, 85% of deaths occur in low and middle-income countries (LMICs), as they lack the health care infrastructure required for cytology-based screening, referral colposcopy diagnosis, and expert physicians, which have dramatically reduced the disease burden in high income countries (HICs). Highly sensitive human papillomavirus (HPV) testing has been effective at reducing the incidence and mortality from cervical cancer when directly coupled with treatment; however, a majority of women with HPV do not have cervical precancer, making HPV testing a poor triage test as overtreatment carries risks like hemorrhage and infertility. Colposcopy followed by biopsy, the preferred triage method in HICs, is untenable in most LMIC settings due to the cost of colposcopes and pathology facilities to process and interpret biopsy results. To make matters worse, women are lost to follow up in LMIC settings when a multi-visit model for cervical cancer screening is used. Visual Inspection with Acetic Acid (VIA), the World Health Organization recommended triage test following HPV testing, has widely varied sensitivity and specificity depending on the training level of the provider. In this proposal we are proposing a single visit model for precision diagnosis and treatment in LMICs for cervical cancer prevention. Two major technological tools are needed to implement this model: a low-cost method to perform imaging of the cervix and a machine learning algorithm to automate diagnosis in the absence of a provider. We have previously developed the Pocket Colposcope, which has shown high concordance with standard colposcopy at a fraction of the cost and validated it on thousands of women across nearly every continent. We are now in the process of developing a state-of-the-art convolutional neural network (CNN), called Colposcopy Automated Risk Evaluation (CARE), trained with Pocket colposcopy images to automate the diagnostic process. Our current prototype algorithm has been highly successful at classifying cervical pre-cancers from Pocket Colposcope images retrospectively. Our goals for this proposal are fourfold: 1) improve and generalize the performance of Pocket CARE using >10,000 National Cancer Institute (NCI) standard colposcopy images; 2) generate synthetic images to address domain shifts due to environmental and personnel changes between different clinical sites; 3) embed the CARE algorithm into our existing software to enable high quality image capture with the Pocket Colposcope for automated diagnosis 4) validate the performance of Pocket CARE prospectively with a clinical study in Kisumu, Kenya, a site where Pocket CARE would ultimately be adopted\. The deliverables for this proposal will be a fully validated Pocket CARE software ready for scale to different clinical scenarios based on location-specific cultural contexts and infrastructure and a comparative effectiveness of Pocket CARE to other publicly available algorithms and standard RI care.