Read like a Radiologist: Cancer Detection using Multi-view Correspondence in Digital Breast Tomosynthesis

Abstract

Breast cancer is the second leading cause of cancer death among women, with approximately 43,780 related deaths annually. Effective screening programs are essential in reducing mortality rates by providing early diagnosis and treatment. Historically, mammography has been the most reliable screening method, significantly dropping the per capita mortality rate since its widespread adoption in the 1980s. However, the increasing workload of approximately 40 million mammography procedures conducted annually in the US poses a significant challenge to the healthcare system. This leads to reports of high rates of burnout among breast radiologists, which can decrease reading accuracy and patient care quality. Therefore, there is a need to improve the existing breast cancer screening workflow to address this challenge.CAD algorithms have been developed to reduce radiologists' workload and address burnout. However, existing single-view CAD systems often have limited cancer detection performance as well as clinical impact. To overcome this limitation, we collaborated with iCAD (Nashua, NA) to develop a novel Computer-Aided-Detection (CAD) framework for digital breast tomosynthesis (DBT) that mimics the multi-view mammography reading practice used by breast radiologists. As of May 2023, the algorithm is under the initial submission for FDA 510(k) approval. This dissertation introduces a multi-view DBT lesion detection framework consisting of four chapters. Chapter 1 highlights the challenges of breast cancer screening and the necessity for an enhanced DBT CAD algorithm. Chapter 2 presents the single-view detection pipeline, while chapter 3 proposes the ipsilateral refinement concept that improves cancer lesion detection performance. Chapter 4 outlines the temporal matching concept that enhances system-level performance by integrating lesion temporal growth information. Chapter 5 showcases a few additional studies that supported the development of the multi-view lesion detection algorithm.Our design uses cascaded task-specific models for each of our proposed modules, enabling intermediate reasoning of the multi-view reading steps. This approach allows the radiologist to inspect the output generated by the CAD system and verify the reasoning behind the system's decision, providing an additional layer of validation for complex high-stakes decisions.