Development and Validation of an Artificial Intelligence-Based Clinical Decision Support Tool for Videofluoroscopic Swallowing Studies

摘要 吞咽困难（吞咽功能障碍）在各种医疗条件和患病率中非常普遍 随着年龄的增长而增加。如果诊断错误或不及时治疗，吞咽困难可能导致严重的健康问题。 这些疾病的后果包括营养不良、脱水和肺炎。最常用的程序， 诊断吞咽困难的最佳方法是视频透视吞咽（VFS）研究。一项VFS研究利用钡提供了 对比增强荧光透视检查程序，允许可视化与以下相关的解剖结构和生理结构： 吞咽以及吞咽生物力学损伤的识别。当前VFS分析方法 临床上使用的主要是定性的，并受到可靠性问题的影响。定量方法 支持VFS临床解释确实存在，但主要是在研究环境中发现的，因为时间- 需要逐帧分析的消耗性质。本申请的总体目标是开发 并验证基于人工神经网络的软件，该软件将分割和跟踪临床重要信息 在吞咽视频内逐帧地观察吞咽结构。分割和跟踪将 在采集后自动发生，无需输入或视频编辑。逐帧自动分割 将允许通过算法确定定量度量。为了实现这一 目的，提出了两个具体目标：1）开发和验证基于人工智能的自动分割算法 在VFS研究中准确分割吞咽解剖结构和团注流， 中风和混合病因患者，以及2）应用自动分割算法来导出各种 VFS研究中的临床相关指标，并与手动得出的参考值进行比较。完成 第一个目标是建立预处理技术，以提高图像质量，减少图像伪影 使用一种新型的3D打印拟人头颈模型。使用强大的现有VFS数据集 图像，然后我们将开发一个掩模R-卷积神经网络自动分割的各种 VFS研究的临床相关特征，并将根据手动导出的 细分对于第二个目标，自动分割算法将被应用于导出重要的 从VFS图像中提取吞咽测量和相关度量。算法的输出将是 根据由经验丰富的评估者从VFS图像手动得出的测量和指标进行验证， 建立可靠性。通过该项目开发的工具将减少人类解释的主观性 VFS图像的准确性和可靠性，将提高吞咽困难诊断和治疗的一致性和可靠性。

ABSTRACT Dysphagia (swallowing dysfunction) is highly prevalent in a variety of medical conditions and prevalence increases with advancing age. If incorrectly diagnosed or left untreated, dysphagia can lead to serious health consequences, including malnutrition, dehydration, and pneumonia. The most commonly used procedure to diagnose dysphagia is the videofluoroscopic swallow (VFS) study. A VFS study utilizes barium to provide a contrast enhanced fluoroscopic procedure that allows for visualization of anatomy and physiology relevant to swallowing as well as identification of swallowing biomechanical impairments. Current VFS analysis methods used clinically are primarily qualitative in nature and subject to issues with reliability. Quantitative methods to support VFS clinical interpretation do exist but are primarily found in the research environment due to the time- consuming nature of frame by frame analysis required. The overall objective of this application is to develop and validate an artificial neural network-based software that will segment and track clinically important swallowing structures on a frame-by-frame basis within swallowing videos. Segmentation and tracking will automatically occur post acquisition with no needed input or video editing. Frame by frame auto-segmentation of regions of interest will allow for quantitative metrics to be determined algorithmically. To accomplish this objective, two specific aims are proposed: 1) to develop and validate an AI based auto-segmentation algorithm that accurately segments swallowing anatomy and bolus flow in VFS studies from a retrospective cohort of stroke and mixed etiology patients and 2) to apply the auto-segmentation algorithm to derive a variety of clinically relevant metrics in VFS studies and compare to manually derived reference values. To accomplish the first aim, pre-processing techniques will be established to improve image quality and reduce image artifacts using a novel 3D printed anthropomorphic head & neck phantom. Using a robust existing dataset of VFS images, we will then develop a Mask R-Convolutional Neural Network for automatic segmentation of a variety of clinically relevant features on VFS studies and will validate the auto-segmentation against manually derived segmentation. For the second aim, the auto-segmentation algorithm will be applied to derive important swallowing measures and associated metrics from the VFS images. The output of the algorithm will be validated against measures and metrics manually derived from the VFS images by experienced raters with established reliability. Tools developed through this project will reduce the subjectivity of human interpretation of VFS images, which will improve consistency and reliability of dysphagia diagnosis and treatment.