TR&D Project 3: Virtual Readers

点击翻译按钮获取中文摘要

ABSTRACT – TRD3: Virtual Readers The Center proposes virtual imaging trials (VITs), a new paradigm to evaluate rapidly advancing imaging technologies, including computed tomography (CT). VITs offer a computational alternative to the evaluation of these technologies through clinical trials, which are slow, expensive, and often lack ground truth, while exposing subjects to ionizing radiation. The Center will develop a VIT platform to emulate key elements of the imaging chain from virtual patients (TRD1) to virtual scanners (TRD2) to virtual readers (TRD3). The virtual reader, the focus of this TRD, are defined as image analysis tools that emulate and extend the clinical reading of images for specific tasks or needs such as lesion detection, classification, or measurement. Specifically, the virtual readers comprise three representative categories: observer models, radiomics, and machine learning. Virtual readers can efficiently and effectively analyze the vast amounts of data in imaging trials, be they clinical or simulated. To date, most virtual reader approaches have been limited by their narrow focus, uncertainty of ground truth (normal anatomy and disease), or lack of interoperability. As a result, these technologies have not yet been translated broadly. To address this unmet need, TRD3 will codify a suite of easy-to-use virtual reader tools to enable not only VITs but also a wide range of other medical image evaluation needs. This work will proceed in three Specific Aims: (1) implement an observer model and radiomics toolset for task- based assessment of CT images, (2) create deep learning resources for analysis and processing of CT images, and (3) integrate virtual reader utilities into a unified VIT platform and validate it against studies with real images and radiologists. While TRD3 focuses primarily on virtual readers, as the final technology development project of the Center, it will also validate Center resources as a whole. The deliverables of TRD3 include the following: (1) virtual reader tools that go beyond niche applications and generalize to different subjects, systems, and tasks; (2) performance assessment that is informed by controllable ground truth for both normal anatomy and disease; (3) “estimability index” to assess bias and precision of virtual reader metrics; (4) machine learning tools that perform disease detection and classification as well as data augmentation, all of which are crucial to VITs; (5) resources for medical imaging that transcend VITs with applications including clinical evaluation and education, and (6) benchmark databases and performance levels that facilitate a culture of open science where technology assessment becomes fair and reproducible. TRD3 will have a significant impact on clinical imaging science and practice by not only enabling effective ways of evaluating imaging technology but also spurring new developments in data science for medical imaging. The virtual reader resources combined with myriad clinical and simulated image data of the Center will provide the essential framework to enable VITs in CT imaging and beyond.