Purpose: Machine learning (ML) applications in breast MRI have shown promise for both breast cancer detection and assessment of cancer risk. Common ML approaches for binary classification of medical imaging utilize cross entropy loss. However, a limited number of highly diverse positive (malignant) cases in these datasets decreases the discriminative power of resulting models and reduces clinical utility. To exploit the abundance of negative data, we propose using supervised contrastive learning, an alternative ML approach that allows better separation of negative and positive classes. We tested the performance of these two ML approaches for the distinct tasks of breast cancer detection and prediction of future breast cancer.
Materials and Methods: In this IRB-approved study, we utilized data from 10,185 breast MRIs in 5,248 women (2005-2014). Breasts with BI-RADS 6 lesions were excluded. Subtracted maximum intensity projections (MIPs) from first post-contrast images were used in ML modeling for (a) cancer detection and (b) risk of future breast cancer. For cancer detection, breast outcomes were categorized ‘Malignant’ if cancer was diagnosed <12 months after MRI and ‘Benign’ (negative) for all others. For risk assessment, patients with bilateral ‘Benign’ results on their earliest screening MRI were evaluated and categorized as ‘High Risk’ if they had a future breast cancer diagnosis (>12 months) or ‘Low Risk’ if not. Deep learning models were trained using a ResNet-50 architecture with either cross entropy or supervised contrastive loss functions. Models were trained to predict Malignant/Benign outcomes and High-Risk/Low-Risk using a random 70-30% train-test split; performance was assessed by area under the receiver operating characteristic (ROC-AUC) curve.
Results: Final analyses included 16816 breasts for cancer detection (3.5% Malignant, 96.5% Benign) and 1163 women for risk assessment (5.9% High-Risk; 94.1% Low-Risk; mean follow-up=8.2years). The conventional cross-entropy approach achieved ROC-AUC of 0.819 for cancer detection and 0.609 for risk assessment. For both applications, improved predictive performance was achieved using the supervised contrastive models (Table 1). Explainable machine learning techniques are also being developed to automatically identify the most suspicious areas for clinical follow-up (Figure 1).
Conclusion: Deep learning models for MRI that make use of negative data (via supervised contrastive learning) improve predictive power for cancer detection and risk assessment.
Acknowledgments: This study was supported by NIH/NCI CCSG P30 CA015704, NIH/NCI R37 CA240403, Microsoft.
Clinical Relevance Statement: Deep learning models may be useful for cancer detection and risk assessment, providing radiologists with improved identification of highly suspicious cancer cases and aiding in refinement of risk-based screening strategies.
