The robustness of automated breast density assessment to imaging errors.
Over the last two weeks we heard about the potential limitations of radiologist-based breast density assessment. One study has indicated substantial disagreement between radiologists in the context of a screening programme; another cast light on the fact that an automated measure produced better consistency than radiologists did across serial mammograms. A promising solution to these issues is the use of an objective algorithmic method to measure density. One such method—Volpara—has been extensively described in peer-reviewed literature and has been found to have excellent correlation with MRI (often considered the “gold standard” of measuring breast density). It is also predictive of clinically relevant end-points, such as the risk of developing or missing breast cancer. Lau and colleagues at the University of Malaya further tested the sensitivity and robustness of Volpara to errors in a number of imaging physics parameters, with the aim of further assessing the clinical utility of the software.
Such testing was done by artificially introducing errors to a number of imaging parameters. Values that are output by the mammography machine (compressed breast thickness, CBT; X-ray tube voltage, kVp; filter thickness; tube current-exposure time product, mAs) were manipulated in the DICOM header to introduce ±10% variation. “Errors” of ±10% were also inserted into image correction parameters (detector gain or detector offset) and image noise was implanted via addition of Gaussian noise to the original image. A total of 3 317 mammograms were thus manipulated and compared to the ground truth (the original images, prior to the introduction of errors) in terms of fibroglandular volume (FGV), breast volume (BV) and VBD.
The results worked excellently in Volpara’s favour, with mAs, detector gain and filter thickness having no significant effect on the final measurements. While detector offset and image noise caused minor variation, these did not alter the VBD by more than 0.1%. The greatest discrepancy was found to be caused by errors in the CBT, which led to up to a 1.2% mis-estimation of the VBD. This change in the VBD was mediated by an erroneous estimation of BV. Errors in kVp were also found to introduce a variation in VBD of up to 0.8%, because variation in kVp had a positive relationship with the estimated FGV. In both scenarios, the errors were amplified in large breasts and in dense breasts. However, when one compared the VDG of the images with introduced error to the original unmodified images, there was almost perfect agreement between them (weighed kappa of 0.8 to 1.0), suggesting that the ultimate effect of these errors on clinical outcome would be small.
This study highlights Volpara’s utility in the current clinical setting and its suitability for “routine clinical use in quantifying breast density even if there are minor errors in some of the imaging physics parameters”. The authors state that “Volpara was robust to expected clinical variations”—an encouraging endorsement that automated assessment is the optimal method for measuring breast density and thus for helping in the early detection of breast cancer.
