RETURN TO ABSTRACT LISTING


Background
Machine learning (ML), a subfield of artificial intelligence (AI) in which a program is designed to complete a task while continually improving its performance, can be a powerful tool for clinical radiology.

Educational Goals / Teaching Points
In breast imaging, ML models have demonstrated utility in mammography, MRI, and ultrasonography. Despite recent advances in ML research in breast radiology, there remain significant barriers to widespread clinical implementation.

Key Anatomic/Physiologic Issues and Imaging Findings/Techniques
In traditional ML, features of interest are first defined and extracted from input images (i.e., orthogonal images of mass seen on ultrasound) and then are used as input for ML algorithm (i.e., support vector machine [SVM]) to classify input into diagnostic categories (i.e., benign or malignant). In deep learning (DL), however, higher-order features are extracted automatically without human supervision and are used to classify input into diagnostic categories. DL can accurately estimate mammographic density and 5-year risk of breast cancer. Applying these models to two patients imaged at our institution demonstrates accurate breast density classification as BI-RADS category B (scattered fibroglandular tissue) and BI-RADS category D (extremely dense); DL estimates of 5-year breast cancer risk were 1.9% and 2.4%, respectively. Comparison to the Gail clinical risk model gives 5-year breast cancer risk estimates of 1.1% and 1.4%, respectively. Both patients are Black women without a personal or family history of breast cancer. Koios DS uses an ensemble of ML algorithms to generate a probability of malignancy from an ultrasound ROI. A diagnostic mammogram of a patient at our institution demonstrated a circumscribed, round, mass in the upper inner quadrant. Ultrasound showed a round isoechoic mass correlating with the mammographic mass, which Koios DS interpreted as suspicious. Pathology demonstrated invasive ductal carcinoma with mucinous features. Patients with locally advanced breast cancer are often treated with neoadjuvant therapy (NAT) to reduce tumor size prior to surgery. MRI-based ML models using radiomic features have shown promise in predicting NAT response. These studies often compute hundreds of radiomic features for each subject and use feature selection methods to identify a small number of features for inclusion in the final ML model.

Conclusion
Although there remain significant challenges to widespread clinical implementation, ML models in breast imaging have demonstrated promising results in multiple modalities.