The quantitative potential for breast tomosynthesis imaging

Abstract

Purpose: Due to its limited angular scan range, breast tomosynthesis has lower resolution in the depth direction, which may limit its accuracy in quantifying tissue density. This study assesses the quantitative potential of breast tomosynthesis using relatively simple reconstruction and image processing algorithms. This quantitation could allow improved characterization of lesions as well as image processing to present tomosynthesis images with the familiar appearance of mammography by preserving more low-frequency information.

Methods: All studies were based on a Siemens prototype MAMMOMAT Novation TOMO breast tomo system with a 45 degrees total angular span. This investigation was performed using both simulations and empirical measurements. Monte Carlo simulations were conducted using the breast tomosynthesis geometry and tissue-equivalent, uniform, voxelized phantoms with cuboid lesions of varying density embedded within. Empirical studies were then performed using tissue-equivalent plastic phantoms which were imaged on the actual prototype system. The material surrounding the lesions was set to either fat-equivalent or glandular-equivalent plastic. From the simulation experiments, the effects of scatter, lesion depth, and background material density were studied. The empirical experiments studied the effects of lesion depth, background material density, x-ray tube energy, and exposure level. Additionally, the proposed analysis methods were independently evaluated using a commercially available QA breast phantom (CIRS Model 11A). All image reconstruction was performed with a filtered backprojection algorithm. Reconstructed voxel values within each slice were corrected to reduce background nonuniformities.

Results: The resulting lesion voxel values varied linearly with known glandular fraction (correlation coefficient R2 > 0.90) under all simulated and empirical conditions, including for the independent tests with the QA phantom. Analysis of variance performed on the fit line parameters revealed statistically significant differences between the two different background materials and between 28 kVp and the remaining energies (26, 30, and 32 kVp) for the dense experimental phantom. How ever, no significant differences arose between different energies for the fatty phantom, nor for any of the many other combinations of parameters.

Conclusions: These strong linear relationships suggest that breast tomosynthesis image voxel values, after being corrected by our outlined methods, are highly positively correlated with true tissue density. This consistent linearity implies that breast tomosynthesis imaging indeed has potential to be quantitative.

Figure 1

To-scale schematic of voxelized phantom used in the Monte Carlo simulations. The phantom is 4.5 cm thick with three identical sets of 11 cuboid lesions each. One lesion set was placed at each of three depths within the phantom.

Figure 2

Schematic of phantom arrangements used to test for lesion depth dependence. The lesion layer was identical for all images, whereas the background slabs were both either fat-equivalent or glandular-equivalent plastic. To test energy and exposure dependence, the central arrangement was used for all images.

Figure 3

(Top) One slice through the cylindrical lesions from a reconstructed volume image with the known glandular fraction (in percent) of each plastic piece. For display purposes, the regions of interest were resorted from the actual random order. (Bottom) The same was done for the cuboidal lesions from a Monte Carlo simulation result. Note the change in grayness from left to right in both images (lesion brightness increases with density); however, it is difficult to distinguish between the simulated lesions because lesions are placed directly adjacent to each other in order of increasing density; dark outlines and labels were added later to facilitate lesion distinction.

Figure 4

Demonstration of local background correction method. Original lesion values (squares) were corrected by subtracting nearby background values (pluses) to produce corrected lesion values (x’s) plotted on the secondary y-axis that vary linearly with lesion glandularity. Least-squares best fit line for corrected (postsubtraction) voxel values is shown R²=0.989. Bars on corrected voxel values represent standard error.

Figure 5

Plot of image voxel values for simulations comparing both background material density and the effects of scattered radiation along with trend lines determined by least-squares analysis and standard error bars

Figure 6

Plot of voxel values for different lesion depths for the virtual phantoms with (left) fat background and (right) glandular background with bars showing standard error. Both plots share the same y-axis scale. Legend refers to lesion set location (“bottom” meaning closest to detector).

Figure 7

(Left) Plot showing lack of dependence of corrected voxel values on depth for the fat background phantom and (right) likewise for the glandular phantom. Legend refers to relative position of lesion layer to background slabs. Standard error bars not shown as they are smaller than marker symbols.

Figure 8

Plot illustrating the dependence of corrected voxel values on background material density for the empirical phantoms. Standard error bars are not shown, as they are smaller than marker symbols.

Figure 9

Plot demonstrating lack of dependence of voxel values on energy for phantom with fat-equivalent plastic as the background material. Standard error bars are not shown, as they are smaller than marker symbols.

Figure 10

Plot demonstrating the dependence on energy for glandular background phantom imaged at 28 kVp and independence of energy for same phantom imaged at 26, 30, or 32 kVp. Standard error bars are not shown, as they are smaller than marker symbols.

Figure 11

Plots showing lack of lack of dependence on exposure level for the corrected image voxel values within each background density (fat left, glandular right). Standard error bars are not shown, as they are smaller than marker symbols

Figure 12

Plot showing test of methods on CIRS Model 11A breast phantom 1 cm thick stepwedge and three of the seven hemispheric lesions with thicknesses noted on the plot legend.

Figure 13

Plots of z-dependence of voxel values in slices above and below the 10% and the 70% lesion for phantoms containing 11 lesions and likewise for all five stepwedge lesions in the CIRS Model 11A phantom. Slices known to include the lesion cross-sections are marked with black bar near the x-axis. Analysis was performed using simulation (top row) and empirical phantoms (bottom rows), and with fat (left column) versus glandular (right column) versus 50% (bottom center) surrounding background.