Introduction
Predicting long-term breast cancer risk is a cornerstone of "precision screening," a clinical strategy where high-risk individuals receive tailored surveillance, such as more frequent exams or supplemental MRI imaging. While deep learning models have demonstrated significant potential in this domain, most existing state-of-the-art tools, including the widely recognized MIRAI model, rely on single-time-point mammograms.
In clinical practice, however, radiologists rarely interpret a mammogram in isolation. Instead, they perform a longitudinal comparison, evaluating current images against "priors" to detect subtle temporal changes in breast density or parenchymal patterns. This study introduces PRIME+ (PRIor Mammogram Enabled risk prediction), a framework designed to bridge the gap between AI-driven risk assessment and established clinical workflows by explicitly modeling the evolution of breast tissue over time.
The PRIME+ Framework
PRIME+ extends the MIRAI architecture by incorporating a sophisticated temporal fusion mechanism. The model's architecture is designed to capture the dynamic changes that occur between screening rounds.
Methodology
- Feature Extraction: The system utilizes a shared ResNet-34 backbone to extract high-dimensional feature representations from both the current mammogram ($x_{curr}$) and a prior mammogram ($x_{prior}$).
- Transformer Decoder: To fuse these features, the authors employ a transformer decoder rather than simple concatenation. In this setup, current features serve as the "Query" ($Q$), while prior features act as the "Key" ($K$) and "Value" ($V$). This cross-attention mechanism allows the model to selectively attend to regions where tissue characteristics have shifted between the two time points.
- Survival Analysis: Rather than a binary classification, PRIME+ predicts a cumulative hazard function. This enables the estimation of cancer risk at specific future intervals, typically 1, 2, 3, and 4 years.
Data and Evaluation
The model was developed and validated using a large, multi-institutional dataset comprising 16,113 exams (64,452 images) from 9,113 patients in the United States, collected between 2010 and 2021. The dataset included a diverse mix of biopsy-proven cancer, biopsy-proven benign, and normal exams, all of which contained at least one prior mammogram for comparison. Images were sourced from both Hologic (72.3%) and Siemens (27.7%) hardware.
Results and Analysis
PRIME+ demonstrated a statistically significant performance boost over single-image baselines across nearly all evaluated metrics. The overall C-index—a measure of the model's ability to rank patients by risk—improved from 0.68 to 0.73.
Key Performance Metrics
As shown in the ablation study results, the inclusion of prior images and the specific use of the transformer decoder were critical to the model's success. The improvement was most notable in mid-term (3-year) risk prediction, where the Area Under the Curve (AUC) rose from 0.70 to 0.77.
Robustness and Subgroup Analysis
To ensure the model was predicting future risk rather than merely detecting existing cancers missed by human readers, the authors conducted a sub-analysis excluding any cancer cases diagnosed within 180 days of the mammogram. PRIME+ maintained its performance edge in this scenario, confirming its predictive utility.
Furthermore, the model proved exceptionally effective for women experiencing breast density changes between exams. In this subgroup, PRIME+ achieved a C-index of 0.75, significantly outperforming the baseline's 0.63. This suggests that the transformer decoder successfully captures "density dynamics" that traditional models ignore.
Discussion and Limitations
The study highlights that the change in breast tissue is a powerful signal for risk assessment. By mimicking the longitudinal comparison performed by radiologists, PRIME+ provides a more nuanced and accurate risk profile than single-image models.
However, some limitations remain. The 4-year AUC results did not reach statistical significance, likely due to the smaller sample size of cases with long-term follow-up. Additionally, while the dataset is multi-institutional, it remains an internal cohort. Future research should focus on validating PRIME+ on diverse, public longitudinal datasets to confirm its generalizability across different populations and clinical settings.
Summary updated