Development of Artificial Intelligence-Based Approaches for Computer-Aided Management of Colorectal Polyps

背景和目的：结直肠癌（Colorectal cancer，CRC）是世界上第二大癌症死亡原因， 美国，每年有近15万例新病例和5万例死亡。结肠镜检查伴息肉切除术 仍然是CRC筛查和监测的金标准，因为在手术期间切除肿瘤性息肉， 结肠镜检查改变疾病的结果，并通知后续管理。标准实施规程 仍然倾向于切除所有可见的息肉进行组织病理学评估，尽管估计， 几乎一半的息肉是非肿瘤性的。研究表明，能够可靠地预测息肉 真实的实时内窥镜下病理学检查可以显著提高 结肠镜检查结直肠癌进行结肠镜检查的数量正在增加，在VA中， 在五年内翻了一番。该需求不包括约30%的 筛选患者因此，结肠镜检查可以从各个层面的效率提高中受益匪浅。在 鉴于此，在过去的十年中，内窥镜技术的发展呈爆炸式增长， 更准确地诊断和治疗结直肠肿瘤。人工智能（AI）的最新进展， 特别是在深度学习领域，以及它们在内窥镜成像中的应用，已经显示出了希望。 用于自动化内窥镜息肉病理预测，克服基于操作者的息肉病理 评估因素，如观察者之间的差异，技能和经验。这样的能力最终会 打开了广泛采用节省成本的“切除-丢弃”和“留下”模式的大门， 小息肉，如美国胃肠内窥镜保存学会所提出的， 纳入有价值的内窥镜创新指南。更重要的是，AI- 基于定量图像解释的临床实践，包括VA，有可能 增加早期癌症检测，从而降低患者发病率和死亡率。为此，主要目标 这项研究的目的是利用人工智能，特别是深度学习模型，开发一个准确的， 强大的计算机辅助诊断（CADx）平台，可实现纯粹的内窥镜光学评估， 粘膜病变，特别是结肠直肠息肉。同时，使用AI模型评估结肠 将研究已知的告知结肠镜检查质量的粘膜和管腔特征。 方法：本研究将以三个目标为指导。目的1预测息肉的稳健分类模型 病理学将得到发展。来自现有数据集和临床记录的标记图像和临床数据， 将用于设计和验证深度学习模型。设计将包括两个步骤：概述 包含息肉的图像中的区域，以及随后分析息肉区域以提供病理 预测.借用增强现实的方面，病理预测沿着估计的 息肉边界，将以直观和临床友好的方式呈现给内窥镜医生，作为伪 颜色叠加，增强模型输出预测的透明度和可解释性。这 因此，即时视觉反馈可以在结肠镜检查期间通知临床决策。目标2将侧重于使用 与结直肠肿瘤相关的临床危险因素结合内镜成像数据， 增强息肉病理学的预测。目标是研究公认的CRC临床 从患者的电子健康记录中获得的风险因素和生物标志物， 预测深度学习模型。最后，在目标3中，基于深度学习的检测、分割和 目标1中开发的分类框架将适用于肠道准备评分， 盲肠标志和直肠反曲，识别结肠憩室，并描绘内镜纹身 标志.这些特征与高质量结肠镜检查的执行相关， 识别可以改善和促进内窥镜检查结果的记录和报告生成。

Background and Objectives: Colorectal cancer (CRC) is the second leading cause of cancer death in the United States, with nearly 150,000 new cases and 50,000 deaths annually. Colonoscopy with polypectomy remains the gold standard for CRC screening and surveillance since removal of neoplastic polyps during colonoscopy modifies disease outcomes and informs subsequent management. Standard practice continues to favor removal of all visualized polyps for histopathological assessment, despite estimates that nearly half of the polyps are non-neoplastic. Studies have shown that the capability to reliably predict polyp pathology endoscopically in real time could result in substantial improvement in the cost-effectiveness of colonoscopy for CRC. The number of colonoscopies performed is increasing, and in the VA more than doubled in a five-year span. This demand does not include subsequent procedures required in ~30% of screened patients. Thus, colonoscopy can benefit greatly from efficiency improvements at every level. In light of this, the past decade has seen an explosion in advances in endoscopic technologies toward diagnosing and treating colorectal neoplasia more precisely. Recent advances in artificial intelligence (AI), specifically in the field of deep learning, and their application to endoscopic imaging, have shown promise for automating endoscopic polyp pathology predictions, overcoming operator-based polyp pathology assessment factors such as interobserver variability, skill, and experience. Such capability would finally open the door to widespread adoption of cost-saving resect-and-discard and leave-behind paradigms for diminutive polyps, as proposed by the American Society for Gastrointestinal Endoscopy Preservation and Incorporation of Valuable Endoscopic Innovations guidelines. More importantly, the incorporation of AI- based quantitative image interpretation into clinical practice, including in the VA, has the potential to increase early cancer detection thus reducing patient morbidity and mortality. To this end, the main goal of the proposed study is to leverage AI, specifically deep learning models, to develop an accurate and robust computer aided diagnosis (CADx) platform to enable the purely endoscopic, optical assessment of mucosal pathologies, specifically colorectal polyps. In parallel, the use of AI models to assess colonic mucosal and luminal features known to inform colonoscopy quality will be investigated. Methods: The study will be guided by three aims. In Aim 1 robust classification models for predicting polyp pathology will be developed. Labeled images and clinical data, from existing datasets and clinical records, will be used to design and validate deep learning models. The design will consist of two steps: outlining regions in an image containing a polyp, and subsequent analysis of the polyp region to provide a pathology prediction. Borrowing from aspects of augmented reality, the pathology prediction along with the estimated polyp boundary, will be presented to endoscopists in an intuitive and clinically friendly manner as a pseudo- color overlay, enhancing the transparency and interpretability of the models output predictions. This immediate visual feedback can thus inform clinical decisions during colonoscopy. Aim 2 will focus on using clinical risk factors associated with colorectal neoplasia in combination with endoscopic imaging data to enhance predictions of polyp pathology. The goal is to investigate incorporation of recognized CRC clinical risk factors and biomarkers, obtained from patients’ electronic health records, in our polyp pathology prediction deep learning models. Finally, in Aim 3 the deep learning-based detection, segmentation, and classification frameworks developed in Aim 1 will be adapted for scoring bowel preparation, recognizing cecal landmarks and rectal retroflexion, identifying colonic diverticula, and delineating endoscopic tattoo markings. These features are associated with performing of high-quality colonoscopy, for which automated identification could improve and facilitate documentation of endoscopic findings and report generation.