Table 3: Summary of representative studies using statistical model building based methods for mammographic tissue segmentation. denotes correlation coefficient. Note that () largely identical studies are excluded in the list and () in case of multiple results, only the best reported results are listed.

Study Year Number of density categoriesModalities Number of viewsNumber of imagesSegmentation evaluation Risk/density estimation accuracy

Texture statistical variation
Miller and Astley [57]1992Fatty and dense Digitised SFM MLO40 Visually assessed; 80% overlapping areas matchedN/A
Suckling et al. [58]1995Fatty and fibroglandular Digitised SFM MLO30 (15 pairs) Visually assessed; 69% ± 12% overlapping areas matchedN/A
Heine and Velthuizen [59]2000Fatty and dense Digitised SFM MLO and CC50 Visually assessed N/A
Heine et al. [60]2008Dense and fatty Digitised SFM MLO and CC369 cases and 712 controlsVisually assessed Pearson (CC) = 0.70 (automatic-Cumulus) Spearman's (CC) = 0.49 (automatic-BI-RADS)
Petroudi and Brady [61]2006Dense, fatty, and breast edge Digitised SFM MLO and CC32 Visually assessed (88% very satisfactory); ~94.4% overlapping areas matchedN/A
Gong et al. [62]2006Dense, dense with structures, fatty, and fatty breast edgeDigitised SFM MLO and CC43 Visually assessed; ~87.9% overlapping areas matched based on 15 images~87.9% (Wolfe)
Oliver et al. [63]2010Dense and fatty Digitised SFM and FFDMMLO and CC322 (MIAS) and 250 (the Trueta DB)Visually assessed; accuracy 0.916 ± 0.038, area overlap 0.900 ± 0.122, and Dice coefficient 0.943 ± 0.077 (60 MLO from the Trueta DB)N/A

Texture descriptors
Zwiggelaar and Denton [64]20064 densities Digitised SFM MLO60 Visually assessed 72% (Wolfe)
Adel et al. [65]2007Fibroglandular, fatty, and uncompressed and fattyDigitised SFM MLO and CC50 (mini-MIAS) Visually assessed (68% good); 60% agreement with manual segmentationN/A
Zwiggelaar [66]20104 densities Digitised SFM MLO322 (MIAS) Visually assessed 64% (BI-RADS)