Deep Learning in Medical Imaging: Evaluation and Study Design

• Activity Number: 487 - Neural Networks, Deep Learning, and RKHS
• Type: Contributed
• Date/Time: Wednesday, August 1, 2018 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Robyn Ball* and David Larson and Pranav Rajpurkar and Matthew Chen and Jeremy Irvin and Jaden Yang and Matthew P Lungren
• Companies: Stanford University and Stanford University and Stanford University and Nines AI and Stanford University and Stanford University and Stanford University
• Keywords: deep learning; machine learning; radiology; medical imaging

Abstract

Deep learning methods are increasingly applied to medical imaging and are quickly achieving state-of-the-art performance in many tasks. These powerful techniques are being implemented into the clinical workflow to diagnose diseases on images, aid in image prioritization, manage imaging quality assurance, and reduce radiation and contrast dose. However, significant methodological challenges remain and have raised important questions regarding best practices for evaluation and study design including 1) sampling strategies to label a training set when positive cases are rare, 2) comparing model and radiologist performance in the absence of a hard ground truth, and 3) determining the appropriate size of the test set, a priori, based on the comparison metrics. We developed some practical approaches to address these challenges, which we illustrate in several medical imaging tasks.