A new approach to optimizing and evaluating computer-aided detection schemes

描述（由申请人提供）：本研究的长期目标是提高计算机辅助诊断系统的临床影响。在这个项目中，一个经过校准的 数据集将被开发，将用于优化计算机辅助检测（CADe）算法。目前，CADe算法针对图像中的癌症检测进行了优化（所谓的独立性能，在不考虑放射科医生用户的情况下进行测量）。我们提出的方法将最大化放射科医师在阅读筛查乳房X线照片中的表现（即，CADe将针对放射科医师的临床受益进行优化，而不是独立性能）将测试两个假设：假设1：使用经校准数据集优化的CADe方案的放射科医师将具有比使用当前方法优化的CADe方案时更高的性能;以及假设2：放射科医师使用任意CADe方案的改进性能，如在观察者研究中测量的，可以使用校准的数据集以足够的精度预测。这将通过以下具体目标来实现：1.基于一组放射科医师阅读而没有CADe和一组放射科医师分析个体CADe标记来开发校准数据库; 2.通过观察者研究对校准的数据集进行建模;以及3.开发一种优化CADe的新方法，以最大限度地提高放射科医生的绩效。校准数据集应提高CADe的临床有效性。目前的临床研究表明，通过使用CADe，放射科医生可以将癌症检测的灵敏度提高10%。然而，放射科医生忽略了高达70%的正确CADe标记的癌症。我们认为，通过优化CADe系统，以最大限度地提高放射科医生的受益，而不是在不考虑对放射科医生的影响的情况下最大限度地提高CADe性能，将导致放射科医生在灵敏度方面获得更大的收益。

DESCRIPTION (provided by applicant): The long-term objective of this research is to improve the clinical impact of computer-aided diagnosis systems. Specifically in this project, a calibrated data set will be developed that will be used to optimize a computer- aided detection (CADe) algorithm. Currently CADe algorithms are optimized for detecting cancers in images (so-called stand-alone performance, which is measured without considering the radiologist user). Our proposed method will maximize radiologists' performance in reading screening mammograms (i.e., CADe will be optimized for clinical benefit to the radiologist, not stand-alone performance) Two hypotheses will be tested: Hypothesis 1: Radiologists using a CADe scheme optimized using the calibrated dataset will have higher performance than when using a CADe scheme optimized using current methods; and Hypothesis 2: The improved performance of radiologists using an arbitrary CADe scheme, as measured in an observer study, can be predicted with sufficient accuracy using a calibrated dataset. This will be accomplished through the following specific aims: 1. Develop the calibrated database based on a group of radiologists reading without CADe and a group of radiologists analyzing individual CADe marks; 2. Validate the calibrated dataset through an observer study; and 3. Develop a novel method for optimizing CADe to maximize radiologists' performance. The calibrated dataset should improve the clinical effectiveness of CADe. Current clinical studies show that by using CADe, radiologists can increase their sensitivity for cancer detection by 10%. However, radiologists ignore up to 70% of correct CADe marked cancers. We believe that by optimizing CADe systems to maximize the benefit to the radiologist, as oppose to maximizing CADe performance without considering the effect on radiologists, will lead to larger gains in sensitivity by radiologists.