An AI model trained to detect abnormalities on breast MR images accurately depicted tumor locations and outperformed benchmark models when tested in three different groups, according to a study published today in Radiology, a journal of the Radiological Society of North America (RSNA).

“AI-assisted MRI could potentially detect cancers that humans wouldn’t find otherwise,” said the study’s lead investigator Felipe Oviedo, Ph.D., a senior research analyst at Microsoft’s AI for Good Lab.

Screening mammography is considered the standard of care in breast cancer screening. However, mammography is less effective in patients with dense breasts. Breast density is an independent risk factor for breast cancer and can mask a tumor. Physicians may order breast MRI to supplement screening mammography for women who have dense breasts and those at high risk for cancer.

“MRI is more sensitive than mammography,” Dr. Oviedo said. “But it’s also more expensive and has a higher false-positive rate.”

To enhance the accuracy and efficiency of screening breast MRI, Dr. Oviedo’s research team closely collaborated with clinical investigators in the Department of Radiology at the University of Washington to develop an explainable AI anomaly detection model. Anomaly detection models distinguish between normal and abnormal data, flagging the anomalies, or abnormalities, for further investigation.

“Previously developed models were trained on data of which 50% were cancer cases and 50% were normal cases, which is a very unrealistic distribution,” Dr. Oviedo said. “Those models haven’t been rigorously evaluated in low-prevalence cancer or screening populations (where 2% of all cases or less are cancer), and they also lack interpretability, both of which are essential for clinical adoption.”

To address these limitations, the researchers trained their model using data from nearly 10,000 consecutive contrast-enhanced breast MRI exams performed at the University of Washington between 2005 and 2022. Patients were predominately white (greater than 80%), and 42.9% had heterogeneously dense breasts, while 11.6% had extremely dense breasts.

“Unlike traditional binary classification models, our anomaly detection model learned a robust representation of benign cases to better identify abnormal malignancies, even if they are underrepresented in the training data,” Dr. Oviedo said. “Since malignancies can occur in multiple ways and are scarce in similar datasets, the type of anomaly detection model proposed in the study is a promising solution.”

In addition to providing an estimated anomaly score, the detection model produces a spatially resolved heatmap for an MR image. This heatmap highlights in color the regions in the image that the model believes to be abnormal. The abnormal regions identified by the model matched areas of biopsy-proven malignancy annotated by a radiologist, largely surpassing the performance of benchmark models.

The model was tested on internal and external datasets. The internal dataset consisted of MRI exams performed on 171 women (mean age 48.8) for screening (71.9%; 31 cancers confirmed on subsequent biopsy) or pre-operative evaluation for a known cancer (28.1%; 50 cancers confirmed by biopsy). The external, publicly available, multicenter dataset included pre-treatment breast MRI exams of 221 women with invasive breast cancer.

The anomaly detection model accurately depicted tumor location and outperformed benchmark models in grouped cross-validation, internal and external test datasets, and in both balanced (high prevalence of cancer) and imbalanced (low cancer prevalence) detection tasks.

If integrated into radiology workflows, Dr. Oviedo said the anomaly detection model could potentially exclude normal scans for triage purposes and improve reading efficiency.

“Our model provides an understandable, pixel-level explanation of what’s abnormal in a breast,” he said. “These anomaly heatmaps could highlight areas of potential concern, allowing radiologists to focus on those exams that are more likely to be cancer.”

Before clinical application, he said the model needs to be evaluated in larger datasets and prospective studies to assess its potential for enhancing radiologists’ workflow.