Abstracts

RETURN TO ABSTRACT LISTING


E1317. Breast Ultrasound in the Age of Artificial Intelligence: Current State and Future Directions
Authors
  1. Jamie Oliver; NYU Grossman School of Medicine
  2. Beatriu Reig; Department of Radiology, NYU Grossman School of Medicine
  3. Yiming Gao; Department of Radiology, NYU Grossman School of Medicine
  4. Alana Lewin; Department of Radiology, NYU Grossman School of Medicine
  5. Linda Moy; Department of Radiology, NYU Grossman School of Medicine
  6. Laura Heacock; Department of Radiology, NYU Grossman School of Medicine
Background
Breast imaging was an early adopter of computer-aided detection systems and more recently AI for screening mammography. Recently, there has been a substantial increase in interest to develop AI tools for breast ultrasound (US). This is likely attributable in part to increased utilization of US as a supplemental screening modality as well as the development of automated breast ultrasound (ABUS) which has increased the number of images radiologists are tasked with reviewing. Recent efforts to apply deep-learning to breast US have produced promising results. In the future, radiologists may incorporate these algorithms into clinical practice, which will augment our ability to care for patients. For this reason, we aim to familiarize breast imagers with current applications of AI in breast US, discuss clinical implications, and future directions.

Educational Goals / Teaching Points
Understand what tasks AI currently do and do not do well, in the clinical context of breast US. Discuss the relative challenges of applying AI to breast US compared to digital mammography and digital breast tomosynthesis. Review current breast US AI research, and consider the clinical implications of this work.

Key Anatomic/Physiologic Issues and Imaging Findings/Techniques
Exhibit content: Define AI terminology (machine learning, deep learning, deep convolutional neural networks, etc.), highlighting the differences between terms. Review current limitations of screening breast US. Discuss what tasks deep learning models do well in breast US (quantification, feature extraction, classification, data mining). Contrast this with what AI tools currently do not do well (lesion detection and localization compared to mammography, integration of complex information, relative performance in the diagnostic vs. the screening setting). Review current areas of research utilizing AI in breast US: use of supervised and semi-supervised neural networks to improve breast imager performance, ability of AI tools to decrease the number of US guided benign biopsies, capacity of AI to evaluate breast lesions and lymph nodes, and use of AI lesion detection in ABUS. Discuss future areas of AI research in breast US including real time lesion detection with CAD and the potential value of AI and radiomics in precision medicine, such as the feasibility of AI tools to classify breast cancer by molecular subtype and predict likelihood of regional and distant metastasis.

Conclusion
The potential of AI in breast US is immense and offers breast imagers the opportunity to re-imagine the roles of breast US. Understanding the state and promise of this field is essential.