Digital mammography: For Processing and For Presentation data
Digital mammography generates two types of images.
The “For Processing” (Raw) images are acquired from the imaging sensor when a mammogram is taken and the pixel value is related directly to x-ray attenuation.
The “For Presentation” (Processed) images are generated using each manufacturer’s own proprietary processing algorithms,
which essentially enhance the features of the raw mammogram (e.g.
contrast) for easier viewing by radiologists but lose any quantitative data associated with the raw image.
In this paper,
we demonstrate how viewing “For Presentation” mammograms over-time is far from ideal and how viewing the quantitative information inherent in the “For Processing” image is far superior for detecting change over time.
Consistency of automated density measurements across different x-ray vendors using Volpara™
Volpara™ is an automated,
objective volumetric breast density assessment system that is also FDA cleared to report a BI-RADS density score.
A by-product of the Volpara processing is the generation of a “density map” in which each pixel value is set to represent the breast density between that pixel and the x-ray source.
So,
for example,
a higher pixel value represents a higher breast density.
The idea is that by viewing those “quantitative” images over time as a movie,
the eye will pick up the changes in densities that are occurring.
Of course,
it is by no means certain that a woman will be imaged on the same x-ray machine each time she visits a clinic.
Thus,
our first task was to assess the ability of Volpara™ to produce consistent density measurements,
using mammograms taken from the same woman,
imaged on different imaging systems.
Using datasets of mammographic images from US sites,
we obtained:
(A) 105 cases where a woman had craniocaudal (CC) and mediolateral oblique (MLO) mammograms taken of both breasts,
on either a GE or Hologic x-ray system.
The women in this dataset ranged in age from 40-93 years and the median age was 61 years.
The sequential mammograms were taken one (102/103 women) or two years (3/103 women) apart.
(B) 18 cases where women were imaged sequentially on either Fuji CR or Fuji DR systems were also obtained with CC and MLO views of each breast taken.
The “For Processing” images for all cases were run through Volpara™ to obtain average volumetric breast densities for each timepoint.
A Pearson correlation co-efficient was then obtained for each dataset,
to assess the agreement between mammograms from the same woman on different vendors’ x-ray systems.
Transformation of Volpara™ density maps into temporal movies
The temporal analysis process is to make density maps,
from mammograms taken over multiple years,
more comparable when appended into a movie for visual presentation.
Density maps were generated,
by Volpara™,
for the left and right CC mammograms from 12 women,
screened every year over a 4-8 year period,
on a mix of GE,
Hologic and Siemens x-ray units.
Temporal movies were made of the density maps (as described below) and,
for comparison,
of the “For Presentation” images.
In order to avoid introducing distracting deformation to the density maps,
we set up image pairs by mapping each previous density map to the latest one.
The mapping process for each pair includes two steps,
which starts with a feature based image registration followed by a pixel based image registration.
For the pre-processing feature based registration,
a robust method was developed to identify corresponding landmarks and boundaries on the density map pair to be registered (for example,
nipples,
axilla or rib points,
breast contours etc.).
Pixel based registration was then used to improve the texture matching inside the breasts.
Texture mapping was constrained within a limit distance to avoid false deformation.
After the two-step image registrations,
the corresponding dense tissues in the breasts were very close to each other,
facilitating the visualization of any density changes over time.