AI models accurately detect breast cancer with MRI scans

The AI model trained to detect abnormalities in breast MR images accurately portrayed tumor location when tested in three different groups, accurately depicting benchmark models that were superior to the benchmark model. RadiologyJournal of the North American Radiology Society (RSNA).

“AI-assisted MRI could potentially detect cancers that humans cannot find that they are not,” said Dr. Felipe Oviedo, the research chief investigator, who is 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 not very effective in patients with dense breasts. Breast density is an independent risk factor for breast cancer and can hide tumors. Doctors can supplement breast MRI to supplement screening mammography for women with dense breasts and high risk of cancer.

MRI is more sensitive than mammography. However, it is more expensive and has a higher false positive rate. ”

Dr. Felipe Oviedo, Senior Research Analyst at Microsoft's AI for Good Lab

To improve the accuracy and efficiency of breast MRI screening, Dr. Oviedo's research team worked closely with clinical researchers in the Department of Radiology at the University of Washington to develop an explanatory AI anomaly detection model. Anomaly detection model distinguishes between normal and abnormal data and flags abnormal or abnormality for further investigation.

“The previously developed model was trained on data where 50% are cancer cases and 50% are regular cases, which is a very unrealistic distribution,” Dr. Oviedo said. “These models have not been rigorously evaluated in low-cost cancer or screening populations (2% of all cases are cancer), and neither of them have any interpretability essential for clinical recruitment.”

To address these limitations, researchers trained the models using data from nearly 10,000 consecutive contrast breast MRI trials conducted at the University of Washington between 2005 and 2022.

“Unlike traditional binary classification models, our anomaly detection models have learned a robust representation of benign cases to better identify abnormal malignant tumors, even if they were underestimated in training data,” Dr. Oviedo said. “The type of abnormality detection model proposed in this study is a promising solution, as malignant tumors can occur in multiple ways and are rare in similar data sets.”

In addition to providing estimated anomaly scores, the detection model generates spatially resolved heatmap of the MR images. This heatmap highlights in color areas of the image that the model believes to be abnormal. The abnormal areas identified by the model are consistent with areas of malignant tumors demonstrated on radiologist-annotated biopsies, significantly outperforming the benchmark model's performance.

This model was tested on internal and external datasets. The internal dataset consisted of an MRI study (71.9%; 31 cancers confirmed by subsequent biopsies) or preoperative assessment of known cancers (28.1%; 50 cancers confirmed by biopsies) performed in 171 women (mean age 48.8) for screening. An externally published multicenter dataset included pretreatment MRI examinations of 221 women with invasive breast cancer.

The anomaly detection model accurately portrayed grouped cross-validation, internal and external test data sets, and tumor locations and more outperformed benchmark models for balanced (high prevalence of cancer) and imbalance (low cancer prevalence) detection tasks.

When integrated into the radiology workflow, Dr oviedo said anomaly detection models could eliminate normal scans for triage purposes and improve reading efficiency.

“Our model provides an easy to understand, pixel-level explanation of breast abnormalities,” he said. “These anomalous heat maps can highlight areas of potential concern, allowing radiologists to focus on trials that are likely to develop cancer.”

Prior to clinical applications, he stated that the model should be evaluated with a larger dataset and prospective study to assess the potential to enhance the workflow of radiologists.

“Cancer detection in breast MRI screening with explainable AI abnormality detection.” Collaborated with Dr. Oviedo were Dr. Anum S. Kazerouni, Ph.D., Dr. Philip Rizzlesky, Yixi Xu, Ph.D., Michael Hirano, MS, Robert A. Vandermalen, Ph.D., Marius Croft, Ph.D., William B. Weeks, MD, Ph.D., MBA, Rahul Dodhia, Ph.D., Juan M. Lavista Ferres, Ph.D., Habib Rahbar, MD, and Savannah C. Partridge, Ph.D.