RSNA 2019

Abstract Archives of the RSNA, 2019


SSE01-05

Application of Machine Learning in the Calculation of Breast Density Using Transmission Ultrasound: A Comparison with Automated Mammographic Assessment

Monday, Dec. 2 3:40PM - 3:50PM Room: E450A



Participants
Bilal Malik, PhD, Novato, CA (Presenter) Employee, QT Ultrasound Labs
Sanghyeb Lee, PhD, Novato, CA (Abstract Co-Author) Employee, QT Ultrasound
James Wiskin, PhD, Salt Lake City, UT (Abstract Co-Author) Employee, QT Ultrasound Labs
Rajni Natesan, MD, MBA, Houston, TX (Abstract Co-Author) Officer, QT Ultrasound Labs

PURPOSE

Increased mammographic density has been found to be an important input into breast cancer risk models. Current breast density assessments rely upon 2D projections or a 3D model consisting of 2D reconstructed images, which may not fully capture the topologically complex nature of the breast. In this study, we (1) describe and compare threshold- and clustering-based algorithms that use transmission ultrasound (TU) for the calculation of breast density, and (2) compare Quantitative Breast Density (QBD) with automated mammographic density calculations.

METHOD AND MATERIALS

Retrospective data was used from all women screened at a single breast center between April 2017 and November 2018 for a total of 309 breast scans. Within a 3-month interval, each subject received both a digital screening mammogram with tomosynthesis and TU of the breast. Mammographic breast density values were provided by VolparaDensity 3.1 (Volpara Health Technologies). QBD algorithms (1) segment breast tissue from water using attenuation, and (2) segment fibroglandular tissue by both thresholding based on the speed of sound, and clustering into fibroglandular tissue and fat. The ratio of fibroglandular tissue to total breast volume is calculated as QBD. QBD values were correlated with mammographic breast density scores and BI-RADS breast composition categories using Spearman's correlation coefficient (r), where p<0.05 was considered significant. We discuss the variability of QBD as affected by iterative image reconstruction schemes.

RESULTS

We found strong correlations between automated breast density values from TU and mammography (Spearman r=0.93, 95% CI:0.91-0.94, p<0.01), and between QBD and BI-RADS breast composition categories (Spearman r=0.88, 95% CI: 0.86-0.91, p<0.01). The machine learning-based QBD was less sensitive to variability (by 65%) than the threshold-based QBD.

CONCLUSION

We provide evidence that QBD calculations derived from TU are strongly correlated with automated mammographic breast density assessments. Further, machine learning-based QBD calculation is more robust and repeatable than threshold-based methods.

CLINICAL RELEVANCE/APPLICATION

The presence of dense breast tissue is an independent risk factor for breast cancer. An accurate calculation of breast density is critical for risk stratification in screening for breast cancer.

Printed on: 03/01/22