Abstracts

RETURN TO ABSTRACT LISTING


1576. Behind NME Lines: A Machine Learning Approach to Reducing MRI-Guided Biopsies of Benign Nonmass Enhancement
Authors * Denotes Presenting Author
  1. Iram Dubin *; David Geffen School of Medicine at UCLA
  2. Brian Dubin; Newport Harbor Radiology Associates
  3. Melissa Joines; David Geffen School of Medicine at UCLA
  4. Kyung Sung; David Geffen School of Medicine at UCLA
  5. Haoxin Zheng; David Geffen School of Medicine at UCLA
  6. Cheryce Fischer; David Geffen School of Medicine at UCLA
Objective:
Nonmass enhancement (NME) seen at breast MRI presents with a varied appearance and considerable overlap between benign, high-risk, and malignant lesions. Although imaging features such as distribution, internal enhancement, and kinetics are used to characterize NME, biopsy is often required for definitive diagnosis with benign biopsy rates greater than 50% in certain large studies. The purpose of this study was to develop a machine learning algorithm to help reduce the biopsy rate of benign NME.

Materials and Methods:
100 lesions defined as NME that underwent MRI-guided biopsy at our institution between 2013-2019 with complete clinical, imaging, and pathological data were included in the study. Lesions were categorized as benign, high-risk, or malignant based on pathology results. Random Forest was used for the machine learning model, and results from five-fold cross-validation were averaged over ten repetitions to predict the likelihood of high-risk or malignant versus benign lesions.

Results:
41 benign (41%), 29 high-risk (29%), and 30 malignant (30%) NME lesions were biopsied. Feature selection using multivariate logistic regression was employed to identify six variables predictive of high-risk and malignant NME versus benign NME. These variables included personal history of breast cancer, amount of fibroglandular tissue, background parenchymal enhancement, NME distribution, NME internal enhancement, and NME delayed phase kinetics. The receiver operating curve (ROC) analysis yielded an area under the curve (AUC) of 0.81. Using an ideal cutoff value with sensitivity and specificity of 71% and 82%, respectively, the rate of benign biopsy results would decrease by 56%.

Conclusion:
A machine learning algorithm that models six predictive variables of high-risk and malignant NME may halve the rate of benign biopsies.