Breast MRI Background Parenchymal Enhancement Categorization Using Deep Learning: Outperforming the Radiologist

Abstract

Background: Background parenchymal enhancement (BPE) is assessed on breast MRI reports as mandated by the Breast Imaging Reporting and Data System (BI-RADS) but is prone to inter and intrareader variation. Semiautomated and fully automated BPE assessment tools have been developed but none has surpassed radiologist BPE designations.

Purpose: To develop a deep learning model for automated BPE classification and to compare its performance with current standard-of-care radiology report BPE designations.

Study type: Retrospective.

Population: Consecutive high-risk patients (i.e. >20% lifetime risk of breast cancer) who underwent contrast-enhanced screening breast MRI from October 2013 to January 2019. The study included 5224 breast MRIs, divided into 3998 training, 444 validation, and 782 testing exams. On radiology reports, 1286 exams were categorized as high BPE (i.e., marked or moderate) and 3938 as low BPE (i.e., mild or minimal).

Field strength/sequence: A 1.5 T or 3 T system; one precontrast and three postcontrast phases of fat-saturated T1-weighted dynamic contrast-enhanced imaging.

Assessment: Breast MRIs were used to develop two deep learning models (Slab artificial intelligence (AI); maximum intensity projection [MIP] AI) for BPE categorization using radiology report BPE labels. Models were tested on a heldout test sets using radiology report BPE and three-reader averaged consensus as the reference standards.

Statistical tests: Model performance was assessed using receiver operating characteristic curve analysis. Associations between high BPE and BI-RADS assessments were evaluated using McNemar's chi-square test (α* = 0.025).

Results: The Slab AI model significantly outperformed the MIP AI model across the full test set (area under the curve of 0.84 vs. 0.79) using the radiology report reference standard. Using three-reader consensus BPE labels reference standard, our AI model significantly outperformed radiology report BPE labels. Finally, the AI model was significantly more likely than the radiologist to assign "high BPE" to suspicious breast MRIs and significantly less likely than the radiologist to assign "high BPE" to negative breast MRIs.

Data conclusion: Fully automated BPE assessments for breast MRIs could be more accurate than BPE assessments from radiology reports.

Level of evidence: 4 TECHNICAL EFFICACY STAGE: 3.

Keywords: artificial intelligence; background parenchymal enhancement; breast MRI; cancer risk assessment; deep learning.

Figure 1:

AI model architecture schematic. A VGG-19 architecture was trained to classify images into four BPE categories, which were then pooled into “high BPE” and “low BPE” categories. In the MIP AI model, axial MIPs generated from the first subtraction phase were used as the model input. In the Slab AI model, axial slices from the first subtraction phase were pooled into three maximum intensity slabs, which were each used as an independent model input. CONV = convolutional layer, FC = fully connected layer.

Figure 2:

Patient selection flow diagram.

Figure 3:

AI model test set results. Receiver operating characteristic curves and areas under the curve for the Slab AI and MIP AI models, over the full test set (a), over the BI-RADS 4/5 subgroup (b) and over the reader study subgroup (c). In (c) results are displayed using both Reference Standard #1 (Radiology Report BPE labels) and Reference Standard #2 (Consensus Reading BPE labels).

Figure 4:

Case examples illustrating (a) the Slab AI model and Radiology Report (Reference Standard #1) both classifying as High BPE, (b) the Slab AI Model classifying as Low BPE and the Radiology Report classifying as High BPE, (c) the Slab AI Model classifying as High BPE and the Radiology Report classifying as Low BPE, and (d) both the Slab AI Model and Radiology Report classifying as Low BPE.

Figure 5:

Trends in AI model BPE designations. (a) In the Consensus Reading BI-RADS 1 Subgroup, the Slab AI Model was less likely than the Radiology Report to classify cases as High BPE (p = 0.004). In the BI-RADS 4/5 Subgroup, the trend was reversed (p < 0.0001). (b) Percentages of Breast MRI Exams with High BPE, according to the Slab AI Model versus the Radiology Report.