by John R. Fischer
, Senior Reporter | December 18, 2020
Risk assessment models for breast cancer development incorporate only small amounts of patient data and only one type of information from screening mammograms themselves.
Researchers at Massachusetts General Hospital told viewers at the 2020 annual Radiological Society of North America Meeting that they are hoping to change this with their own deep learning image-only model.
"There are subtle but informative cues on a mammogram that may not be discernible by the human eye or simple volume-of-density measurements, and are not included in current risk assessment models. Deep learning leverages these cues to improve risk assessment techniques. Our model discovers these patterns directly from the data, rather than manually identifying discriminative image patterns," author Dr. Leslie Lamb, breast radiologist at Massachusetts General Hospital, told HCB News.
Data from patients incorporated in current models includes family history, prior breast biopsies and hormonal and reproductive history. Breast cancer is the only type of information accounted for from mammograms.
Lamb and her colleagues developed their model on a population of women with a personal history of breast cancer, implants or prior biopsies, and incorporated 245,753 consecutive 2D digital bilateral mammograms from 80,818 patients screened between 2009 and 2016. The model takes into consideration non-image risk factors, such as age, to further refine the accuracy, and imputes this information if it is not provided.
Training consisted of 210,819 of the mammograms for 56,831 patients, while testing used 25,644 exams from 7,021 patients. The model was validated with 9,290 exams from 3,961 patients.
When compared against Tyrer-Cuzick version 8, a commercially available risk assessment model, the MGH model showed greater accuracy in predicting the risk of developing breast cancer within five years of the index mammogram, with a predictive rate of 0.71 compared to 0.61.
The model has been externally validated in Sweden and Taiwan. Additional studies are planned for minority populations, especially larger African-American populations, who face greater delays in starting and undergoing treatment
, and are more likely to die from breast cancer than white people.
"We will also focus on expanding the utilization across vendors and for unilateral mammograms," said Lamb.