A hallmark of successful breast cancer treatment is knowing if the cancer has spread to nearby lymph nodes. But traditional methods, like biopsy, imaging, surgery, and clinical exams, may lack precision, sometimes leading to unnecessary surgery or overexposure to radiation.

NCI-funded researchers are using advances in machine learning to develop a more precise way of predicting cancer’s spread from breast tissue to adjacent lymph nodes.

By blending breast magnetic resonance imaging (MRI) data with clinical and pathologic information, the researchers developed a four-dimensional convolutional neural network capable of predicting whether cancer will spread to lymph nodes, regardless of tumor size or stage.

According to corresponding author, Dr. Dogan Polat, of the University of Texas Southwestern Medical Center, their aim was to create a safe and time-efficient tool that operates in the same way a radiologist approaches his or her care—tailored to each patient.

Dr. Polat said, “In our retrospective study, our model was able to avoid unnecessary biopsies (in more than half of the cases). And we were able to correctly predict (87% of the time) which patients’ cancers would spread to adjacent lymph nodes.”

“Using our model, we would have reduced surgeries in the study’s breast cancer patients by at least 50% while still correctly diagnosing cancer in 95% of these cases,” added Dr. Basak Dogan, clinical principal investigator and study lead. “Our findings translate into better care—reducing harmful side-effects and anxiety associated with surgery and radiation.”

Dr. Albert Montillo, co-author of the study and machine learning development lead, added, “We had our best results with machine learning models that not only analyzed the MRI imaging data but also learned to intelligently fuse both clinical and demographic information. We’re currently extending this approach to predict long-term survival.”

The researchers are refining their model. They’re introducing more varied data and further optimizing the process for analyzing images to ensure the tool will be applicable to all patients and easy to incorporate in a busy clinical setting.