Improving Perception in Digital Breast Tomography

Digital Breast Tomosynthesis (DBT) is a breast cancer screening methodology in which radiologists search for cancer in 3D volumes of virtual slices through the breast. DBT performs better than classic, 2D mammography but it takes more time. Our goal is to compare different methods that could reduce the time required while maintaining or improving performance. Beyond the specific goal of improving DBT, we will uncover general principles of attention and perception that can be applied whenever the volume of images threatens to overwhelm the ability of observers to consume those images. We are particularly interested in conditions of “low target prevalence” (low percentage of positive cases). In tasks like breast cancer screening, with many images and very few clinically significant targets, interventions that are effective when tested at high prevalence in the lab may fail in the field when prevalence is much lower. There are three projects: Project 1: Self-Triage by 2D Full-field digital mammography or synthetic images: In screening for breast cancer, there will be some cases that a reader could safely declare ‘normal’ on the basis of the 2D image, alone. Is it reasonable to develop a protocol where some DBT images would be acquired but would not be examined? This “self-triage” would require a very conservative triage criterion in order to avoid triage of any positive cases, but if proven to be safe, self-triage could save significant time and might reduce false negative errors. Project 2: DBT as “hybrid search”: Breast cancer screening involves search for more than one type of target (masses and calcifications, at minimum). Research shows instances where such ‘hybrid’ search for multiple targets leads to elevated errors. We will test the hypothesis that readers are more efficient and/or more accurate if they perform separate searches for each target type. We will measure eye movements to compare search for each target type alone to search for both types together. Experiments with non-experts will investigate basic principles of search in 3D volumes of image data. Project 3: AI Targeted ‘drilling’: Eye tracking has identified two modes of search in 3D stacks of images: “drilling”, where readers move rapidly through depth (Z) while the eyes stay relatively stable in the XY plane, and “scanning”, where readers search widely in XY while moving slowly in Z. We will use eye tracking to evaluate a CAD system developed by iCAD that marks specific locations in the 2D XY image and invites readers to drill in these specific areas. Does that improve CAD performance? Summary: This program of research will produce recommendations for increasing the efficiency of breast cancer screening. Moreover, each study will produce basic science that will be generalizable to other settings and will deepen our understanding of visual search through 3D volumes of image data.