Study: Diagnostic AI system validated for breast MRI
Machine learning systems capable of differentiating malignant from benign breast lesions on dynamic contrast-enhanced MRI (DCE-MRI) is expected to be a valuable aid by reducing variability in diagnostic classification and increasing interpretation speed. Researchers described training and testing a system in nearly 2,000 Chinese women with lesions and masses on breast MRI in the September 18, 2019, Cancer Imaging.
The system was initially developed on breast MRI images of American women. An experienced breast imager manually identifies the location of a lesion, and the computer system conducts three-dimensional (3D) segmentation of the tumor. It then extracts radiomic features in 6 categories: size, shape, morphology, enhancement texture, kinetics, and enhancement-variance kinetics.
The objective of the research teams from the National Clinical Research Center for Cancer’s Department of Breast Imaging at Tianjin Medical University Cancer Institute and Hospital, and the University of Chicago Department of Radiology was to evaluate the potential of quantitative MRI radiomics and machine learning to distinguish between malignant and benign breast lesions on an independent, consecutive clinical dataset from a single institution. The 1,979-patient dataset is believed to be the largest database of its type.
Subjects ranged from 16 to 77 years old, with median age in the 40’s. Approximately two-thirds of both the benign and malignant lesions were masses. Most malignant and benign lesions were invasive ductal carcinomas and fibroadenomas, respectively.
The authors used a single machine learning model for both masses and non-mass enhancements to mimic clinical practice. Breast MRI scans performed in 2015 and 2016 were used as the training dataset, which totaled 1,455 lesions in 1,444 subjects. These included 946 mass lesions, 76% of which were malignant, and 509 non-mass lesions, 70% of which were malignant.
Once trained, the system was applied to a data set of 535 cases with 363 mass lesions, with 81% malignant, and 172 non-mass lesions, with 74% malignant. The system achieved a 99.5% sensitivity and yielded an area under the curve (AUC) value of 0.88 for mass lesions and 0.90 for non-mass lesions.
The researchers reported that actual clinical decisions made by the hospital’s radiologists had a slightly lower positive predictive value (PPV) than the machine learning system on both types of lesions---78.7% compared to 80.3%, respectively. The system suggested 11 fewer unnecessary benign biopsies, 9.6% of the total. However, it failed to recommend biopsy of 2 cancers that had been reported at BI-RADS 5.
The researchers concluded that the results reiterate the potential for breast DCE-MRI and machine learning analysis to provide clinically useful information for patient management on international datasets.
REFERENCE
- Ji Y, Li H, Edwards AV, Papaioannou J, et al. Independent validation of machine learning in diagnostic breast cancer on magnetic resonance imaging within a single institution. Cancer Imaging. Published online September 18, 2019;19(1):64.