An NCI-funded study highlights a new, deep learning (DL) system for interpreting breast cancer images (i.e., dynamic contrast-enhanced magnetic resonance imaging, or DCE-MRI).

The DL system performed on par with radiologists when differentiating between cancerous and benign tumors. Moreover, the system showed it could help personalize the management of patients who test positive for cancer (i.e., patients falling within a BI-RADS 4 category).

As noted by Dr. Krzysztof J. Geras, co-corresponding author and team leader on this study, “DCE-MRI is highly sensitive, often leading to an over-diagnosis or false-positive results. Our model was able to refine those results, which could help in reducing unnecessary workups and biopsies.” Ultimately, he said, his team hopes this model will lead to a better strategy than simply calling for “biopsy-all” in cases with BI-RADS 4 findings.

The researchers trained the model using a New York University (NYU) Langone Health data set, which featured 21,537 DCE-MRI examinations from 13,463 patients. They validated the model using three international data sources—Poland’s Jagiellonian University Hospital, Duke University, and The Cancer Genome Atlas Breast Invasive Carcinoma data set.

These data sets varied in terms of the ratio between positive and negative samples, patient characteristics, and imaging parameters. This diversity, in turn, yielded varying validation results—from near-perfect performance on data sets of invasive-cancer cases (i.e., cancer that’s spread beyond the initial tumor), to less-strong results on data sets with a wide range of malignancy subtypes.

“Our model’s performance was especially remarkable given the many different scanners and image-acquisition protocols used in these data sets.” said co-corresponding author Dr. Jan Witowski. “This speaks to the value of using robust training data, like we had with our NYU Langone Health data, when training DL systems.”

What’s next? The researchers will be conducting further analyses to see if the benefits of using the DL system outweigh the costs. Other future work includes investigating how the system’s decision-making process holds up in a real-world hospital system, especially when it comes to subtle clues that might not be evident from the images alone (i.e., patient characteristics that influence risk).