Called "three-compartment mammography," the technique uses dual-energy imaging to distinguish water, lipid, and protein signatures in a mammography image, wrote the team led by Dr. Heather Greenwood. The study included 233 women who underwent conventional exams plus three-compartment mammography; lesions found with the technique were categorized as either BI-RADS 4 or 5. Half of the lesions were used to train the model, and half were used to test it.
Of the 117 lesions used to validate the model, 98 were benign and 19 were malignant. The researchers used the area under the curve (AUC) to estimate the model's accuracy, and they compared these results with interpretations of the exams by a radiologist who gave a percent probability of malignancy for each lesion using the BI-RADS system.
The AUC for the three-compartment mammography model was 0.69, leading the group to conclude that it shows promise but more work needs to be done.
"[Three-compartment mammography] with machine learning has the potential to aid radiologists in reducing false-positive breast biopsies," the group wrote. "Model optimization is still needed and ... future work will require prospective testing of this model in the clinical setting."
