Sequential Reading Effects in Digital Breast Tomosynthesis

点击翻译按钮获取中文摘要

Project Summary/Abstract Digital breast tomosynthesis (DBT) has emerged as a viable imaging modality for breast cancer screening. But despite its improvements over standard mammography, the breast imaging community is well aware of its limitations. Roughly 40 million women in the United States have a screening mammogram each year, and the combination of a high-volume exam and less than ideal performance have made screening mammography an active area of research for many years. Prior studies have shown that reading breast screening images in batches improves performance, in terms of lower overall recall rates without significant change in the cancer detection rate. But no mechanism for this improvement has been demonstrated. A recently published clinical trial in the UK with over 1 million screening mammograms read by 360 practicing clinicians finds that screening performance actually improves as a reader progresses through a batch of mammograms. We hypothesize that, rather than vigilance, the operational mechanism affecting performance in batch reading is adaptation, which is consistent with improved performance over the course of a batch. We propose to characterize sequential reading effects in DBT images, to see if there is evidence to support our hypothesis that readers visually adapt to the statistical structure of the images as they read them. A few studies have shown adaptation effects in digital mammograms, but there have been no studies to date evaluating sequential effects in DBT. If adaptation is functioning in batch reading of DBT images, then there are a variety of was to take advantage of it to improve screening performance. We propose a thorough evaluation of sequential effects in DBT images that includes basic visual assessments of adaptation, statistical analysis of retrospective clinical data, and prospective analysis of performance in clinical reader studies. These will be analyzed to determine if image readers are consistent with the adaptation hypothesis, improved performance with depth in a batch, and modeled to see if there are additional factors that influence adaptation. Our project consists of three specific aims that cover these topics. Aim 1 implements a battery of visual assessments that define and characterize visual adaptation to DBT images. Aim 2 proposes mining clinical DBT reading data at the University of Pittsburgh Medical Center to evaluate retrospective performance detecting cancer (true-positive rates) and recalling patients (false-positive rates). This data will be modeled with factors that include adaptive sequential effects. Aim 3 will evaluate screening performance in DBT images in “laboratory” studies using clinical readers and cases that evaluate the effects of different batch length and the ordering of cases within a batch.