E1352. Machine Learning for Diagnostic Ultrasound of Triple Negative Breast Cancers: Vascular Features
Authors
Aneka Khilnani;
The George Washington University School of Medicine and Health Sciences
Murwarit Rahimi;
The George Washington University School of Medicine and Health Sciences
Kathleen Johnson;
The George Washington University School of Medicine and Health Sciences
Anna Hu;
The George Washington University School of Medicine and Health Sciences
Ramin Javan;
The George Washington University School of Medicine and Health Sciences
Background
Triple negative (TN) breast cancers tend to be high-grade and are associated with a worse prognosis compared to non-triple negative (NTN) breast cancers. The absence of ER, PR, or HER2 receptors renders these cancers less amenable to hormonal therapies, altering treatment, so the ability to differentiate these using imaging is imperative. There is increasing recognition that angiogenesis is associated with malignant neoplastic changes in breast lesions, therefore doppler ultrasound may play an important role in the diagnosis of TN breast cancers. Machine learning advancements in digital image processing have been proposed to increase diagnostic accuracy.
Educational Goals / Teaching Points
The goal of this study was to evaluate the current literature on the potential of machine learning with quantitative doppler ultrasound image features for the diagnosis of TN breast cancer. We aim to assess whether machine learning may optimize the relative weights of different features to maximize breast lesion discrimination between TN and NTN subtypes.
Key Anatomic/Physiologic Issues and Imaging Findings/Techniques
Vascular fractional areas can be characterized by three key features on color doppler: fractional area of flow in the lesion, mean blood flow velocity, and blood flow index. In TN, there is a lower magnitude or complete lack of vascularity, that may be explained by the HER2 gene’s contribution to angiogenesis, such that the absence of HER2 expression leads to a decrease in vascularity. When using machine learning to predict TN vs NTN breast cancers the addition of color doppler features compared to quantitative grayscale imaging features alone (including margin sharpness, margin echogenicity difference, angular variance in margin, depth-to-width ratio, axis ration, tortuosity, circularity, radius variation, and elliptically normalized skeleton), increased the diagnostic performance by 3.4%.
Conclusion
Machine learning methods that analyze doppler ultrasound images of breast masses can improve differentiation between TN and NTN compared to standard differentiation by visual assessment alone.
