Abstract
Early detection of breast cancer is critical for improving patient outcomes. While traditional risk assessment tools and recent AI models for 2D mammography exist, Digital Breast Tomosynthesis (DBT)—a near-3D imaging modality—has lacked long-term risk forecasting models. This study introduces a deep learning framework that leverages a pre-trained DINOv2 vision transformer and a cumulative hazard layer to predict 5-year breast cancer risk directly from DBT scans. Trained on over 160,000 exams, the model achieves a 0.80 AUROC for 5-year forecasting, demonstrating the presence of latent predictive biomarkers in DBT imaging.
Introduction
Digital Breast Tomosynthesis (DBT) has become a clinical standard since its FDA approval in 2011, offering superior visualization of small tumors by reducing tissue overlap compared to standard 2D mammography. However, existing risk assessment tools like the Tyrer-Cuzick model rely primarily on clinical factors and achieve only modest performance. While AI has advanced risk prediction in 2D imaging, there is a significant gap in long-term (5-year) forecasting for DBT. This paper addresses this gap by proposing an end-to-end deep learning model to forecast risk directly from screening DBTs.
Methods
The proposed framework avoids the computational burden of training 3D Convolutional Neural Networks from scratch by utilizing a resource-efficient pipeline based on foundation models.
The pipeline consists of several key stages:
- Feature Extraction: Individual 2D slices from a DBT series are processed through a pre-trained DINOv2 (ViT-B/14) image encoder.
- Patch Embeddings: Instead of global image tokens, the model extracts patch embeddings to capture fine-grained, local tissue textures, which were found to be more predictive of cancer risk.
- Dimensionality Reduction: To manage the massive feature space of a four-view DBT study, summary statistics (mean, standard deviation, minimum, and maximum) are computed across all frames.
- Cumulative Hazard Layer: Aggregated features feed into a fully-connected layer with a softplus activation, outputting five monotonically increasing risk probabilities for Years 1 through 5.
- Masked Training: A sliding window approach with masked Binary Cross-Entropy loss handles variable follow-up data lengths.
Datasets
The study utilized a large-scale dataset from Mass General Brigham, comprising 161,753 screening DBT examinations from 50,590 patients (2012–2023). The cohort included 6,449 "pre-cancer" exams (taken prior to diagnosis) and 155,304 "healthy" exams. The data was partitioned into training (70%), validation (15%), and testing (15%) sets.
Results and Discussion
The model demonstrated strong predictive performance, achieving an AUROC of 0.80 for 5-year breast cancer prediction on the test set. Performance scaled with the time horizon, with yearly AUROCs of 0.71 (Year 2), 0.75 (Year 3), 0.78 (Year 4), and 0.80 (Year 5).
Kaplan-Meier survival analysis showed a clear and statistically significant separation in cancer-free survival probabilities between high-risk and low-risk groups across all predicted years. Key insights from the development process include:
- Local vs. Global Features: Patch-level details are more informative for risk than global breast structure, as global tokens led to rapid overfitting.
- Statistical Distributions: Capturing the full distribution of patch features (rather than just the mean) significantly improved performance.
- DBT as a Biomarker: The results confirm that DBT imaging contains rich, latent signals indicative of future malignancy.
Limitations and Future Directions
Despite its success, the model has limitations. DINOv2 is pre-trained on natural images; future work will explore self-supervised pre-training on DBT volumes to bridge the domain gap. Additionally, replacing simple statistical aggregation with attention-based Multiple Instance Learning (MIL) could better highlight biologically relevant regions. Finally, incorporating clinical context (e.g., age, hormonal history) and diversifying the dataset—which is currently 81-85% White—are essential steps toward a more robust and equitable clinical tool.
Summary updated