Researchers have developed a deep learning (DL) model that assesses a woman’s five-year cancer risk with a single breast MR image, sharing their findings in the American Journal of Roentgenology.
“Although a large number of risk assessment models exist for the general screening population, they are not readily applicable to high-risk patients,” wrote Tally Portnoi, department of electrical engineering and computer science at Massachusetts Institute of Technology in Cambridge, and colleagues. “Most of the existing models are well calibrated at the population level, but they are not sufficiently discriminative at the individual level. This limitation becomes even more pronounced for high-risk cohorts, for whom commonly used risk factors are not sufficiently predictive and for whom it is less clear which additional patient features drive outcomes.”
The authors aimed to develop an image-based DL model that “learns useful features directly from the data” rather than building in specific biomarkers in advance. This, they explained, “enables the model to capture subtle patterns that may not be discernible to the human eye.”
Portnoi and colleagues gathered more than 1,600 consecutive breast MR images from breast cancer screening examinations performed on more than 1,100 high-risk women from January 2011 to June 2013 at a single institution. Women were excluded from the study if a five-year screening follow-up examination was not available or if they had cancer other than primary breast cancer develop in the breast after the examination.
The team ultimately completed two models: a risk factor logistic regression (RF-LR) model based on traditional risk factors and the image-based DL model at the center of their study, which was trained to predict if a patient would develop breast cancer within five years of the screening examination. The authors then compared those models with a breast cancer risk evaluation tool that already existed, the Tyrer-Cusick (TC) model.
Overall, the RF-LR model achieved an area under the ROC curve (AUC) of 0.558 ± 0.108. The DL model had an AUC of 0.638 ± 0.094. The TC model, meanwhile, had an AUC of 0.493 ± 0.092.
“In the present study, this DL model delivered more accurate predictions than traditional risk factor models, such as the TC model,” the authors wrote.
Portnoi and colleagues observed that the AUC value of their model increased, and variability decreased as the size of the training dataset increased. In addition, they added, focusing on “the full spectrum of 3D data” instead of 2D images shows potential to “further improve risk prediction.”