A Review on Automatic Mammographic Density and Parenchymal Segmentation

Table 6

Summary of representative studies using 2D projection based volumetric methods for mammographic tissue segmentation. denotes correlation coefficient. Note that in case of multiple results, only the best reported results are listed.

Study	Year	Number of density categories	Modalities	Number of views	Number of images	Segmentation evaluation	Risk/density estimation accuracy
Prior calibration
Heine et al. [79]	2011	Dense and fatty	FFDM (raw)	CC	106 cases and 106 controls	Visually assessed; linear = 0.78 (automatic-Cumulus)	Risk estimates associated with the lowest to highest quartiles, odds ratios: 1.0, 3.4, 3.6, and 5.6
In-image reference phantom based calibration
Alonzo-Proulx et al. [80]	2012	Dense and fatty	FFDM	CC	55087	Pearson (left, right volumetric density-breast volume) = 0.92, 0.91	N/A
Ourselin et al. [81]	2014	Dense and fatty	FFDM	CC	480	Percent breast fibroglandular volume, = 0.8 (automatic-predicted)	N/A
Physical image formation model
Hartman et al. [82]	2008	Dense and fatty	FFDM and MRI	MLO and CC	550 (275 pairs) and 88 MRI	Breast density volumes, Pearson (left-right breast, Quantra-MRI) = 0.923, 0.937	N/A
Highnam et al. [83]	2010	Dense and fatty	FFDM and MRI	MLO and CC	2217 and MRI from 26 younger women	Breast density volumes, Pearson (left-right breast, CC-MLO view, Volpara-MRI) = 0.923, 0.915, 0.94	N/A
Gubern-Mérida et al. [84]	2014	Dense and fatty	FFDM and MRI	MLO and CC	680 and 168 MRI	Pearson (volumetric breast density Volpara-MRI, fibroglandular tissue volume Volpara-MRI) = 0.91, 0.84	Density grade, = 0.40 (Volpara-BI-RADS)