Table 2.
Representative studies in AI-enabled direct breast risk assessment from mammographic images
| Study | Image format | Time from exam to breast cancer diagnosis | # images (# women) | Vendors (# sites) | Model architecture | Model performance |
|---|---|---|---|---|---|---|
| Long-term risk assessment | ||||||
| Yala et al. [46] | FFDM (processed) | 1–5 years | 295,002 images (91,520 women) | Hologic (3 sites) | ResNet-18* |
AUC = 0.84, 1-year risk AUC = 0.76, 5-year risk |
| Dembrower et al. [43] | FFDM (processed) | 3.6 ± 2.2 years | 150,502 images (1188 cases; 10,563 controls) | Hologic (N/R) | Inception-ResNet* |
OR = 1.55 ORadj = 1.56 AUC = 0.65 |
| Arefan et al. [45] | FFDM (processed) | 1–4 years | 452 images (113 cases; 113 controls) | Hologic (1 site) | GoogleLeNet |
AUC = 0.68, CC AUC = 0.60, MLO |
| Yala et al. [44] | FFDM (processed) | 1–5 years | 88,994 images (1821 cases; 38,284 controls) | Hologic (1 site) | ResNet-18* |
AUC = 0.68 for image only DL AUC = 0.70 for hybrid DL + risk factors |
| Ha et al. [47] | FFDM (processed) | 2–5.3 years | N/R (210 cases; 527 controls) | GE (1 site) | CNN |
OR = 4.42 Acc = 72% |
| Short-term risk assessment | ||||||
| Lotter et al. [48] |
FFDM (processed) DBT (MSP) |
1–2 years | N/R (> 1000 cases; 62 K controls) | GE, Hologic (7 databases/sites) | RetinaNet* | AUC = 0.75–0.76 |
| Eriksson et al. [49] | FFDM (processed) | 3 months–2 years | N/R (974 cases, 9376 controls) | GE, Philips, Sectra, Hologic, Siemens (4 sites) | CNN** |
HR = 7.9 AUC = 0.73 |
| McKinney et al. [50] | FFDM (processed) | 0 months–3.25 years | N/R (> 105 k women) | Hologic, GE, Siemens (4 sites) |
RetinaNet MobileNetV2 ResNet-v2-50 ResNet-v1-50 |
AUC = 0.76–0.89 |
The table describes the development image dataset used in each study, including format of mammographic images, time window from mammographic exam to breast cancer diagnosis, sample size, and vendors, as well as model architecture and performance in breast cancer risk assessment
FFDM full-field digital mammography, CNN convolutional neural network, AUC area under the ROC curve, Acc accuracy, OR odds ratio, HR hazard ratio
*Indicates publicly available AI model. **Indicates commercial model. N/R not explicitly reported in the paper