VolparaDensity works by estimating the volume of fibroglandular tissue and the volume of the breast and then working out the ratio to get the volumetric breast density – the percentage of the breast that is fibroglandular tissue.
One of the advantages with a volumetric approach to measuring breast density is that because it is a measure of a physical property of the breast, you can validate it by comparing results between imaging modalities, to that end we welcome several independent studies which have appeared over the last few months, including:
“Wavelia Breast Imaging: The Optical Breast Contour Detection Subsystem”, which appeared in Applied Sciences 2020. In this study, Cano et al., amongst other things, showed that breast volume estimates using microwaves were highly correlated with breast volume measurements from Volpara, with a Pearson’s correlation coefficient of 0.97.
“Realistic compressed breast phantoms for medical physics applications” by Garcia et al., which was presented at the International Workshop on Breast Imaging 2020 in Leuven, studied breast phantoms imaged with breast CT and Volpara and showed a remarkable correlation between the two:
A further interesting paper from South Korea appeared in September: “Fully automated measurement of volumetric breast density adapted for BIRADS 5th Edition: a comparison with visual assessment” – Youk et al, Acta Radiological, September 2020 looked at 4,000 studies in South Korea and compared Volpara to radiologists reads of BIRADS 5th Edition density. This study re-affirmed that Volpara correlates but does not match perfectly with South Korean radiologists judgement of breast density. The biggest difference was between “C” and “D” breasts, where South Korean radiologists tended to go for “C” compared to Volpara. However, it also shows that across the crucial “B/C” boundary that defines a woman as fatty or dense, Volpara and the South Korean radiologists appear to agree well, with Volpara rating 79% of women as dense, the South Korean radiologists in this study rating 75% as dense, compare that to a third-party automated density system also used in the study which rated just 54% of women as dense.
Of course, there have been plenty of deep learning papers looking at learning ABCD, or indeed cancer risk directly from images, but they all lack independent validation across different datasets read by different radiologists – given the subjectivity in reading density, we should expect markedly different results.