Table 1.
Representative studies in AI-enabled breast density evaluation from mammographic images
| Study | Model development dataset | Model design | Model performance | ||||
|---|---|---|---|---|---|---|---|
| Image format | # images (# women) | Vendors (# sites) | Model architecture | Output density measure | Density maps | ||
| Roth et al. [35] | FFDM (Processed) | 109,849 images (N/R) | N/R (7 sites) | DenseNet-121 | BI-RADS density | No | Four-class K = 0.62–0.77 |
| Dontchos et al. [25] | FFDM (Processed) | N/R (2174 women) | Hologic (1 site) | ResNet-18 | BI-RADS density (13 interpreting radiologists) | No |
Dense versus non-dense Acc: 94.9% (academic radiologists) 90.7% (community radiologists) |
| Matthews et al. [26] | FFDM (Processed) and SM |
FFDM: 750,752 images (57,492 women) SM: 78,445 images (11,399 women) |
Hologic (2 sites) | ResNet-34 | BI-RADS density (11 interpreting radiologists) | No |
Four-class K = 0.72 for FFDM, Site 1 Four-class K = 0.72 for SM, Site 1 Four-class K = 0.79 for SM, Site 2 |
| Saffari et al. [27] | FFDM | 410 images (115 women) | Siemens (1 site) | cGAN, CNN | BI-RADS density | Yes | DSC = 98% in dense tissue segmentation |
| Deng et al. [28] | FFDM | 18,157 images (women) | Hologic (1 site) | SE-Attention CNN | BI-RADS density | No | Acc = 92.17% |
| Perez Benito et al. [29] | FFDM (Processed) | 6680 images (1785 women) | Fujifilm, Hologic, Siemens, GE, IMS (11 sites) | ECNN | BI-RADS density (2 interpreting radiologists) | Yes | DSC = 0.77 |
| Chang et al. [30] | FFDM (Raw) | 108,230 images (21,759 women) | GE, Kodak, Fischer (33 sites) | ResNet-50 | BI-RADS density (92 interpreting radiologists) | No | Four-class K = 0.67 |
| Ciritsis et al. [31] | FFDM | 20,578 images (5221 women) | N/R (1 site) | CNN | BI-RADS density (consensus of 2 interpreting radiologists) | No |
AUC = 0.98 for MLO views AUC = 0.97 for CC views |
| Lehman et al. [32] | FFDM (Processed) | 58,894 images (39,272 women) | Hologic (1 site) | ResNet-18* | BI-RADS density (12 interpreting radiologists) | No | Four-class K = 0.67 |
| Mohamed et al. [33] | FFDM (Processed) | 22,000 images (1427 women) | Hologic (1 site) | CNN AlexNet | BI-RADS density | No | AUC = 0.94 |
| Mohamed et al. [34] | FFDM (Processed) | 15,415 images (963 women) | Hologic (1 site) | CNN AlexNet | BI-RADS density | No |
AUC = 0.95 for MLO views AUC = 0.88 for CC views |
| Haji Maghsoudi et al. [38] | FFDM (Raw) | 15,661 images (4437 women) | Hologic (2 Sites) | U-net* | APD% | Yes |
DSC = 92.5% in breast segmentation APDdiff = 4.2–4.9% |
| Li et al. [37] | FFDM (Raw) | 661 images (444 women) | GE (1 site) | CNN | APD% | Yes | DSC = 76% in dense tissue segmentation |
| Kallenberg et al. [36] | FFDM (Raw) | N/R (493 women) | Hologic (1 site) | CSAE | APD% | Yes | DSC = 63% in dense tissue segmentation |
The table describes the development image dataset used in each study, including format of mammographic images, sample size, and vendors, as well as methodological details for the AI model (output breast density measure, model architecture and availability of spatial density maps) and the model performance in breast density evaluation
FFDM: full-field digital mammography, SM 2D synthetic mammographic image acquired with digital breast tomosynthesis, APD% area percent density, MLO medio-lateral oblique, CC cranio-caudal, cGAN conditional generative adversarial network, CNN convolutional neural network, ECNN entirely convolutional neural network, CSAE convolutional sparse auto encoder, DSC dice score, APDdiff difference in APD%, K Cohen kappa coefficient, AUC area under the ROC curve, Acc accuracy
*Indicates publicly available AI model. N/R not explicitly reported in the paper