Computer-Aided Triage of Body CT Scans with Deep Learning

利用深度学习对身体 CT 扫描进行计算机辅助分类

• 批准号: 10585553
• 负责人: JOSEPH Y LO
• 金额: $ 58.68万
• 依托单位: DUKE UNIVERSITY
• 依托单位国家: 美国
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-08-16 至 2027-07-31
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10585553
• 关键词: 3-Dimensional; Abdomen; Active Learning; Address; Algorithms; Anatomy; Archives; Artificial Intelligence; Attention; Automated Annotation; Case Study; Chest; Classification; Clinical; Clinical Decision Support Systems; Complex; Computer Assisted; Computer-Assisted Diagnosis; Consumption; Data; Data Set; Databases; Detection; Development; Diagnostic; Disease; Expert Systems; Face; Foundations; Future; Goals; Health system; Heterogeneity; Image; Institution; Label; Lung; Manuals; Medical Imaging; Modeling; Natural Language Processing; Nodule; Organ; Outcome; Patients; Pelvis; Performance; Radiology Specialty; Reporting; Research Personnel; Resolution; Scanning; Shapes; Site; Structure; Supervision; System; Testing; Therapeutic Equivalency; Time; Training; Triage; Variant; X-Ray Computed Tomography; artificial intelligence algorithm; body system; chest computed tomography; cone-beam computed tomography; cost; deep learning; deep learning model; federated learning; imaging modality; improved; innovation; large datasets; learning strategy; radiologist; tool

PROJECT SUMMARY / ABSTRACT Computed tomography (CT) imaging for the body can result in thousands of images spanning many organs and myriad possible diseases. With growing patient load as well as increasing resolution and complexity of scans, the task of CT interpretation has become daunting. To improve radiologist performance, many artificial intelligence (AI) algorithms have been produced, but most are limited by their very narrow application to a specific disease in a specific organ or have been trained on limited data due to the high cost and complexity of manual annotation. As a result, there is an unmet need because existing AI solutions have not significantly improved the workflow or performance of radiologists. To meet these needs, we propose to develop a computer-aided diagnosis triage tool for CT of the chest, abdomen, and pelvis (CAP) that would focus radiologists’ attention on regions with a high likelihood of actionable disease while minimizing search efforts in regions of low likelihood. Our hypothesis is that a triage tool will improve radiologist workflow while simultaneously maintaining or improving performance. Our long-term goal is to create a clinical decision support system that will address bottlenecks of radiologist workflow and performance. As key steps toward demonstrating feasibility for that goal, we propose the following three specific aims: 1. Create framework for the assembly, deidentification, annotation, and sharing of over a million chest, abdomen, pelvis (CAP) CT cases from two major health systems. 2. Develop a triage system trained using multi-site CT datasets through collaborative/federated learning. 3. Pilot use of the triage system at multiple sites to allow radiologists to perform efficiently and equivalently for clinical tasks of assessing actionable disease in CAP CT. Key innovations will include the use of weak supervision to label a massive number of cases from two health systems. Labeling will be based on rule-based expert systems as well as natural language processing. Image classification will be based on deep learning models capable of processing an entire 3D CT volume and trained with federated learning to leverage the rich heterogeneity of data from the two health systems. The expected outcome of this project will be evidence to support a new clinical workflow for radiologist interpretation, which is the foundation for all medical imaging. For this project, we will maximize impact by addressing CAP CT because of the large patient load and complex anatomy/disease, and by producing one of the largest medical imaging datasets that can be shared for future research including grand challenges. In addition, by leveraging existing data in patient archives and radiology reports, our approach has the potential to be applicable to other body sites or imaging modalities in the future.

项目摘要 /摘要 人体的计算机断层扫描（CT）成像可能会导致数千个跨越许多器官的图像 和无数可能的疾病。随着患者负荷的增长以及分辨率和复杂性的增加 扫描，CT解释的任务变得艰巨。为了提高放射科医生的表现，许多艺术 智能（AI）算法已经产生 特定器官中的特定疾病或由于高成本和复杂性而对有限数据进行了培训 手动注释。结果，有未满足的需求，因为现有的AI解决方案尚未显着 改善了放射迷的工作流程或表现。 为了满足这些需求，我们建议为胸部CT开发计算机辅助的诊断分流工具， 腹部和骨盆（CAP）将放射科医生的注意力集中在很有可能的地区 可操作的疾病，同时最大程度地减少了可能性很小的地区的搜索工作。 我们的假设是，分诊工具将改善放射科医生的工作流程，同时维护或 提高性能。我们的长期目标是建立一个临床决策支持系统，以解决 放射科医生工作流程和性能的瓶颈。作为证明可行性的关键步骤 目标，我们提出以下三个具体目标： 1。为集会，去识别，注释和共享一百万箱子创建框架， 腹部，骨盆（CAP）CT病例来自两个主要的卫生系统。 2。通过协作/联合学习开发使用多站点CT数据集训练的分类系统。 3。在多个地点使用分类系统的试点使用，以使放射科医生能够有效而同样地执行 评估CAP CT中可行疾病的临床任务。 关键创新将包括使用弱监督来标记两个健康的大量病例 系统。标签将基于基于规则的专家系统以及自然语言处理。图像 分类将基于能够处理整个3D CT量并训练的深度学习模型 通过联合学习来利用两个卫生系统的数据的丰富异质性。 该项目的预期结果将是支持放射科医生的新临床工作流程的证据 解释，这是所有医学成像的基础。对于这个项目，我们将通过 由于患者负荷和复杂的解剖学/疾病，解决CAP CT，并通过产生一种 最大的医学成像数据集，可以共享未来的研究，包括巨大的挑战。在 补充，通过利用患者档案和放射学报告中的现有数据，我们的方法有可能 将来适用于其他身体部位或成像方式。