A new deep learning model leveraging data from screening mammograms can predict a patient’s breast cancer risk with significant accuracy, according to a new study published in Radiology. The authors noted that their technique outperformed the popular Tyrer-Cuzick breast cancer risk model.
Researchers from Massachusetts General Hospital (MGH) in Boston and the Massachusetts Institute of Technology (MIT) in Cambridge collaborated on the risk model, which combines deep learning and risk factors associated with the patient to make its assessments. Data from more than 88,000 consecutive screening mammograms was used to train, validate and test its performance.
“There’s a very large amount of information in a full-resolution mammogram that breast cancer risk models have not been able to use until recently,” lead author Adam Yala, PhD candidate MIT in Cambridge, Massachusetts, said in a prepared statement. “Using deep learning, we can learn to leverage that information directly from the data and create models that are significantly more accurate across diverse populations.”
Yala and colleagues found that their model had an area under the ROC curve (AUC) of 0.71 for predicting breast cancer in both white and black patients. The Tyrer-Cusick model, on the other hand, had a lower AUC for white (0.62) and black (0.45) patients.
“Unlike traditional models, our deep learning model performs equally well across diverse races, ages and family histories,” co-author Regina Barzilay, PhD, a professor at MIT, said in the same statement. “Until now, African-American women were at a distinct disadvantage in having accurate risk assessment of future breast cancer. Our AI model has changed that.”
The team also found that patients at a high risk of breast cancer (according to their model) and nondense breasts had 3.9 times the cancer incidence of patients with dense breast tissue and a low risk of cancer.
This new technique is already in use at MGH, assessing each patient’s risk of breast cancer as they undergo screening mammograms. The authors did disclose that MIT and MGH have filed patents on the prediction model.