Combining AI with Radiologists: exploring the possibilities in implementation of Computer-Aided Detection

R. Kluge

Master thesis 2020.

Since the introduction of deep learning, a new era in radiology has started: the transformation of Computer-Aided Detection (CAD) tools that approaches radiologist level performance. Compared to traditional CAD, modern deep-learning CAD is applied to various problems in radiology. Less is known about where to implement such tools in the clinical workflow and how this impacts the workflow of radiologists. As an example use case, we evaluated three (two commercial) CAD systems for pulmonary nodule detection on chest radiographs. We considered three broad strategies: CAD as first reader, CAD as second reader, and CAD concurrently with readers. Even though standalone CAD performs worse than readers, a performance increase of 10% sensitivity or 7% increase in specificity can be achieved depending on the implementation scenario. During CAD as second reader, the possibility of a reader to score an image as 'uncertain' allowed an increase in sensitivity from 69% to 72% using CAD with no further drawbacks. For most other scenarios, the trade-off between specificity - sensitivity versus reading time was observed. Apart from measuring performance, change in workflow and risk were also considered as metrics to capture qualitative results. By comparing scenarios against these metrics, we show the effects of various implementation strategies for CAD on one dataset.