Abstract
The ability to accurately estimate risk of developing breast cancer would be invaluable for clinical decision-making. One promising new approach is to integrate image-based risk models based on deep neural networks. However, one must take care when using such models, as selection of training data influences the patterns the network will learn to identify. With this in mind, we trained networks using three different criteria to select the positive training data: an inherent risk model trained on images with no visible signs of cancer, a cancer signs model trained on images containing cancer or early signs of cancer, and a conflated model trained on all images from patients with a cancer diagnosis. We find that these three models learn distinctive features that focus on different patterns, which translates to contrasts in performance. Short-term risk is best estimated by the cancer signs model, whilst long-term risk is best estimated by the inherent risk model. Carelessly training with all images conflates inherent risk with early cancer signs, and yields sub-optimal estimates in both regimes.
Introduction
In mammography, a critical distinction exists between risk prediction and cancer detection. Risk prediction aims to estimate the likelihood of developing cancer over a 5-to-10-year horizon, whereas cancer detection identifies existing tumors. Current state-of-the-art deep learning models often fail to distinguish between these two tasks. By training on all available images from patients who eventually develop cancer, models tend to focus on "loud" signals—existing tumors—rather than the subtle textures indicative of long-term predisposition. This paper explores the consequences of this "conflation" and proposes a decoupled training approach to improve both short-term detection and long-term risk assessment.
Methodology
The researchers utilized the Swedish CSAW dataset, comprising over 138,000 mammograms, to train three variants of a ResNet50 architecture:
- Inherent Risk Model: Trained on "clean" images (contralateral breasts or images taken >1 year before diagnosis).
- Cancer Signs Model: Trained on images taken within one year of diagnosis, likely containing early signs.
- Conflated Model: Trained on all images from cancer patients, following standard practice.
Performance was evaluated using Area Under the Curve (AUC) and compared against a clinical baseline of mammographic density.
Key Results and Performance
The study revealed that the choice of training data fundamentally alters the model's predictive capabilities over time.
As shown in Figure 1, the Inherent Risk Model (orange) maintains a stable AUC of approximately 0.61–0.62, even for predictions five or more years into the future. This suggests it captures a biological baseline of risk. In contrast, the Cancer Signs Model (red) and the Conflated Model (blue) show a sharp increase in performance as the time to diagnosis decreases, peaking at AUCs of 0.71–0.72 within the final year. Notably, the Conflated model—despite having the most training data—is sub-optimal in both regimes compared to the specialized models.
Decoupling the Signals
The models do not just differ in performance; they identify different sets of high-risk patients.
Venn diagram analysis (Figure 6) highlights the "conflation" trap. For exams taken within 31 days of diagnosis (c), the Conflated and Cancer Signs models show massive overlap (69 exams), while the Inherent Risk model identifies a largely distinct group (41 unique exams). Conversely, for long-term predictions (>2 years), the models diverge significantly, with each identifying unique populations. This confirms that the Inherent Risk model captures features that are ignored by models focused on immediate detection.
Discussion and Clinical Implications
The "lazy" nature of neural networks means they will prioritize the easiest features to identify. Because a tumor is a more prominent feature than long-term risk textures, standard "conflated" training leads to models that are effectively tumor detectors masquerading as risk predictors.
Clinically, this is dangerous. A conflated model might flag a patient as "high long-term risk" when they actually have an undiagnosed tumor requiring immediate intervention. By decoupling these signals, we can provide more precise clinical guidance: identifying those who need immediate follow-up versus those who require long-term preventative monitoring.
Conclusion
This research demonstrates that more data is not always better; the curation of "positive" samples is paramount. To accurately predict long-term risk, models must be forced to ignore the present signs of disease. Decoupling inherent risk from early cancer signs is a necessary step toward reliable AI-assisted clinical decision-making in oncology.
Summary updated