Machine learning-based analysis of average glandular dose in mammography: The role of breast density.
Authors
Affiliations (6)
Affiliations (6)
- Instituto Nacional de Astrofísica, Óptica y Electrónica, Luis Enrique Erro #1, Sta María Tonanzintla, 72840 San Andrés Cholula, Puebla, Mexico; Instituto Nacional de Investigaciones Nucleares, Carretera La Marquesa S/N, Ocoyoacac, 52750, Estado de México, Mexico.
- Facultad de Medicina, Universidad Autónoma del Estado de México, Paseo Tollocan S/N, Toluca, Estado de México 50130, Mexico.
- Hospital de Ginecología y Obstetricia, Instituto Materno Infantil del Estado de México, Puerto de Palos S/N, Isidro Fabela Primera Secc, 50170 Toluca de Lerdo, Mexico.
- Instituto Nacional de Investigaciones Nucleares, Carretera La Marquesa S/N, Ocoyoacac, 52750, Estado de México, Mexico.
- Instituto Nacional de Astrofísica, Óptica y Electrónica, Luis Enrique Erro #1, Sta María Tonanzintla, 72840 San Andrés Cholula, Puebla, Mexico.
- Instituto Nacional de Astrofísica, Óptica y Electrónica, Luis Enrique Erro #1, Sta María Tonanzintla, 72840 San Andrés Cholula, Puebla, Mexico. Electronic address: [email protected].
Abstract
Average glandular dose (D<sub>g</sub>) is the primary metric for assessing radiation risk in screening mammography. Although D<sub>g</sub> analysis is commonly interpreted according to compressed breast thickness (CBT), breast density (BD) may significantly influence D<sub>g</sub>. This study evaluated the influence of BD on D<sub>g</sub> variability using the TG282-based dosimetry model and a machine learning approach. The cross-sectional study included 1233 full-field digital mammography images from three mammography units. D<sub>g</sub> was estimated using a TG282-based model. Images were stratified by CBT, view (CC and MLO), and BD (ACR BI-RADS). Non-parametric tests and Spearman's correlation assessed group differences and associations. A Random Forest (RF) regression model was implemented to explore non-linear relationships between D<sub>g</sub> and age, CBT, compression force, and BD. Model performance was assessed using R², MAE, and MAPE. Feature importance analysis identified the contribution of each predictor. D<sub>g,75</sub> benchmarks were defined as the 75th percentile of D<sub>g</sub> distributions stratified by BD and CBT. D<sub>g</sub> distributions were non-normal, and D<sub>g</sub> was higher in MLO than CC views (p < 0.05). Significant differences were observed across CBT and BD categories. D<sub>g</sub> correlated positively with mAs and exposure time, while BD and compression force showed moderate associations. Within the four-predictor RF model (R² = 0.56, MAPE = 14.38 %), compression force (45.72 %) and BD (23.93 %) had the highest relative feature importance. D<sub>g,75</sub>benchmarks increased with BD within CBT groups. Incorporating BD improves the characterisation of D<sub>g</sub> variability in mammography. The RF model characterised multivariable relationships between BD and compression-related factors, supporting more anatomically informed dose optimisation strategies. BD-stratified D<sub>g,75</sub>benchmarks may improve identification of atypical dose values and support targeted dose monitoring in mammography.