Figure 1.
Example slices from 4 different datasets (correspondent to 4 subjects).
Each dataset belongs to a different ACR group.
Figure 2.
Intensity profile along the line, marked red in the example slice.
The intensities of the same tissue (for example, the fatty part) in the upper slice part are about 700, and in the lower breast part they are close to 100.
Figure 3.
Another intensity inhomogeneity example.
The intensities of the breast tissue close to the breast air boundaries (marked by the red rectangle) have higher values than the intensities of the breast tissue located in the middle of the breast.
Figure 4.
Steps of manual breast segmentation.
In each slice, the user first marks the breast boundary and then detects the parenchymal regions.
Figure 5.
An overview of the automated breast density evaluation framework.
The approach consists of three main steps: segmentation and bias field correction, breast tissue delineation, and fibroglandular tissue extraction. The results of the bias field correction (Step 1) are used both for breast tissue (Step 2) and parenchyma (Step 3) extraction. In Step 3, the results of Steps 1 and 2 are utilized for parenchyma extraction. The data flow is schematically explained by the lines, connecting the pipeline steps.
Figure 6.
Intermediate steps of the postprocessing for breast tissue extraction.
Left: a segmentation result with the breast tissue connected to the pectoral muscle. Right: the segmentation is corrected and the minor holes are closed as well. The lower part of the breast image is excluded from further postprocessing.
Figure 7.
Final steps for breast tissue extraction.
Left: the concavities of the fibroglandular tissue are closed with morphological operations. Right: the approximate location of the sternum bone is computed and the cut is done along its lower boundary.
Figure 8.
Results for slices, shown in Figure 1.
Left column: segmented images; middle column: corrected images; right column: histograms of the corrected images. The breast tissue becomes homogeneous after the correction. The histograms show the clearly separated intensity classes.
Figure 9.
An example of breast tissue (upper row) and fibroglandular tissue (lower row) segmentation results.
The results are overlaid in 2D and 3D with the original data.
Figure 10.
Bland-Altman with the regression line of differences on average (left) and linear regression (right) plots for the Breast Volume (BV) values of 37 datasets.
The agreement is slightly higher for bigger breasts.
Figure 11.
Bland-Altman with the regression line of differences on average (left) and linear regression (right) plots for the Parenchyma Volume (PV) values of 37 datasets.
No exact influence of the breast density on the agreement results is observed.
Table 1.
Dice's Similarity Coefficients, Delineation Sensitivity, and Specificity for Breast Volumes (BV).
Table 2.
Dice's Similarity Coefficients, Sensitivity, and Specificity for Parenchyma Volumes (PV).
Table 3.
Breast and Parenchyma Mean Bias SD.
Table 4.
Coefficients of the regression lines of the differences.
Figure 12.
Manual (purple) and automatic (red) parenchyma segmentation results for four datasets.
Whereas the results are very similar, one can observe that the user selects usually more than the algorithm.