Early detection through screening mammography has decreased death rates from breast cancer. There are approximately 39 million mammogram procedures conducted each year in the US. However, there are still alarmingly high error rates in radiological interpretations, with missed cancer rates ranging from 10-18 percent and false positive rates as high as 67% over a 10-year period. To reduce error rates, digital breast tomosynthesis, a new 3D imaging technology intended to make cancers more visible to the radiologist, is rapidly being introduced throughout clinics in the US. The widespread adoption of new 3D technologies has dramatically increased the data volume that radiologists must scrutinize and has fundamentally altered how they search, relying on vision away from where they are fixating (peripheral vision), eye movements, and image (slice) scrolls to find potential disease. Yet, we are missing a theoretical and empirical understanding of how radiologists might best search through 3D clinical images to maximize target detection while maintaining reasonable reading times. Furthermore, what metrics to use to optimize image processing and acquisition parameters to maximize radiologists’ performance with the new 3D technologies? The last 30 years of the field of medical imaging have been shaped by task-based image quality metrics based on ideal and model observers that mimic human performance. And yet, these often omit human bottlenecks of peripheral visual processing and do not work with 3D search with clinical images or phantoms. In this context, the overall goal of the current proposal is to combine recent Deep Neural Network developments and biologically-plausible models of human-foveated vision to create a model that learns about anatomy and optimally (performance maximizing) programs eye movements and scrolls to find lesions in digital phantoms and clinical images (foveated search DNN, FS-DNN). If successful, the FS-DNN could be used to evaluate in what way a particular radiologist is not adequately examining the 3D clinical, estimate the accuracy costs of the scroll/search inefficiency, and determine how they might improve. The project will implement learning protocols based on FS-DNN/human comparisons and assess their impact on improving diagnostic accuracy. If successful, the FS-DNN could be a new powerful metric of image quality for 3D search that, unlike previous models, can be applied to phantoms and clinical images. A collaboration with the Food and Drug Administration (FDA) aims at integrating the developed FS-DNN model with the FDA Virtual Imaging Clinical Trial for Regulatory Evaluation (VICTRE) pipeline that is made available to academic researchers and industry technology developers. Together, these advances can potentially help reduce radiological errors with digital breast tomosynthesis and also help evaluate and optimize new technologies. Although the proposed work is developed for breast cancer and DBT, the approach, framework and concepts investigated are potentially applicable to other areas of 3D medical images in radiology.

3D imaging is rapidly replacing 2D planar images in radiology. We aim to create a model with radiologist-like vision that learns to best search and scroll through 3D volumes. The model can be used to identify inefficiencies in radiologists' search/scroll strategies and mitigate search errors, and as a tool to evaluate and optimize the quality of 3D medical imaging technology.