Predicting 5-Year Breast Cancer Risk from Longitudinal Digital Breast Tomosynthesis: A Single-center Retrospective Study

This single-center retrospective study demonstrates that a deep learning model utilizing longitudinal digital breast tomosynthesis (DBT) exams significantly outperforms both single-time-point DBT models and established clinical or FFDM-based risk models in predicting 5-year breast cancer risk, thereby enabling more dynamic and personalized screening strategies.

The Gist

The "Crystal Ball" for Breast Cancer: A New Way to Look at the Future

Imagine you are trying to predict the weather for the next five years. For a long time, meteorologists only looked at a single snapshot of the sky taken today (a 2D photo) to guess if a storm was coming. They also asked you questions like, "Does your family get storms?" or "How cloudy is your neighborhood?"

This new study introduces a smarter, more dynamic crystal ball for breast cancer risk. Instead of just one photo, it looks at a movie of your breast health over time, using a special 3D camera technology called Digital Breast Tomosynthesis (DBT).

Here is the story of how this new AI model works, broken down into simple parts:

1. The Old Way vs. The New Way

• The Old Way (The Snapshot): Traditional models (like the famous Tyrer-Cuzick calculator) rely mostly on your age, family history, and how "dense" your breast tissue is (like comparing a sponge to a sponge cake). They also use older 2D X-rays (FFDM), which are like looking at a flat shadow of a 3D object. It's helpful, but it misses a lot of the details hidden inside the shadow.
• The New Way (The 3D Movie): This study used a new AI model called DRP. Think of DBT as a 3D CT scan of the breast. It slices through the tissue, removing the "shadows" that hide problems.
  • The Secret Sauce: The DRP model doesn't just look at your latest 3D scan. It looks at all your past scans from the last few years. It's like a detective who doesn't just look at a crime scene today, but reviews the security footage from the last five years to see if anything suspicious started moving or changing.

2. How the AI "Thinks"

Imagine the AI is a very smart librarian.

• The Single-Photo Librarian: If you show it just one photo, it can guess your risk, but it's like guessing a book's plot by reading only the last page.
• The Longitudinal Librarian (DRP): This librarian reads your entire history. It sees how your "library" (your breast tissue) has changed over time. Did the texture get rougher? Did the patterns shift? By watching these changes, it can spot subtle warning signs that a single photo would miss.

3. The Big Results: Who Won the Race?

The researchers tested this new AI against two other "competitors":

1. The Classic Doctor (Tyrer-Cuzick): The standard risk calculator based on your life story.
2. The 2D AI (Mirai): A smart computer that looks at old-fashioned 2D X-rays.

The Scoreboard:

• The New DRP Model (3D + History) won with a score of 0.720.
• The 2D AI (Mirai) scored 0.687.
• The Classic Doctor (Tyrer-Cuzick) scored 0.567.

In plain English: The new model was significantly better at predicting who would get cancer in the next 5 years than the old methods. It was like upgrading from a black-and-white TV to a high-definition 3D movie.

4. The "Density" Surprise

For years, doctors have been worried about women with "dense" breasts (like a thick sponge) because it's hard to see tumors, and they are at higher risk.

• The Old Logic: "You have dense breasts? You are high risk. You need extra scans."
• The New Logic: The AI looked closely and said, "Wait a minute. Not all dense breasts are the same."
  • It found that 37% of women with very dense breasts were actually low risk (safe). This means they might not need expensive, scary extra MRIs.
  • Conversely, it found that 15% of women with "fatty" (low density) breasts were actually high risk. These women might have been overlooked by old methods but now get the extra care they need.

Analogy: Imagine a security guard checking bags. The old rule was, "If the bag is heavy (dense), search it thoroughly." The new AI says, "Let's look inside the bag. Some heavy bags are just full of books (safe), and some light bags are hiding a bomb (dangerous)."

5. Why This Matters for You

This study suggests a future where breast cancer screening is personalized.

• No more "One Size Fits All": Instead of every woman getting the same schedule, the AI can say, "Based on your unique 3D history, you are safe for another year," or "Your tissue is changing in a specific way; let's check you sooner."
• Less Fear, More Precision: It helps avoid unnecessary panic and extra tests for women who are actually safe, while catching the real risks earlier in women who might have been missed.

The Bottom Line

This paper is a breakthrough because it proves that looking at the past (longitudinal data) through a 3D lens (DBT) gives us a much clearer picture of the future. It's not just about taking a picture; it's about watching the story unfold.

Note: This is a research study (a preprint), meaning it's a very promising new tool that needs to be tested in real-world clinics before it becomes the standard for everyone. But the results are incredibly exciting!

Technical Summary

1. Problem Statement

Breast cancer risk prediction is critical for personalized screening strategies, yet current standard-of-care models (e.g., Tyrer-Cuzick, Gail model) rely heavily on demographic and clinical factors, often lacking granular imaging data. While recent deep learning models have improved risk prediction using Full-Field Digital Mammography (FFDM), the field is shifting toward Digital Breast Tomosynthesis (DBT), which provides quasi-3D volumetric data and reduces tissue overlap. However, existing AI risk models largely utilize single-timepoint 2D projections or synthetic 2D images derived from DBT, failing to leverage the full volumetric structural information or the longitudinal temporal changes in breast tissue that may signal future cancer development. There is a lack of validated models that utilize native 3D DBT volumes over time for long-term (2–5 year) risk prediction.

2. Methodology

Study Design and Data

• Cohort: A retrospective study involving 161,165 women (mean age 58.5) undergoing 313,531 DBT exams at a single academic institution (Institute A) between January 2016 and August 2020.
• Data Split: Patient-level random splitting into:
  • Training: 140,393 patients (267,485 exams).
  • Validation: 5,042 patients (11,456 exams).
  • Independent Test Set: 15,184 patients (34,580 exams).
  • Matched Case-Control Cohort: 432 patients (216 cases, 216 controls) for comparison with clinical models.
• Outcome Definition: 5-year risk prediction. A "positive" outcome was defined as a malignant pathology diagnosis (invasive cancer, DCIS, etc.) occurring within 2–5 years post-index exam. Exams with cancer diagnosed within 6 months were excluded to avoid prevalent cases.

Model Architecture: DRP (Deep Risk Predictor)
The authors developed a hierarchical Transformer-based deep learning framework:

1. Slice-Level Feature Extraction: Pre-trained networks extract features from individual DBT slices:
   • GMIC (Global Classifier): For global image context.
   • YOLOX & MogaNet (Object Detectors): For localized lesion features.
   • These extractors were frozen during risk model training.
2. Exam-Level Aggregation: A Transformer Encoder aggregates slice-level features into a single exam-level representation.
3. Longitudinal Modeling: Exam-level representations from the index exam and up to 10 prior exams are combined with clinical variables (age, breast density). A lightweight two-layer Transformer models the temporal dependencies across the longitudinal sequence.
4. Prediction Head: A survival prediction head estimates the conditional probability of cancer occurrence across discrete yearly intervals (2–5 years).
5. Calibration: Platt scaling was applied using the validation set to calibrate probabilities for the test set.

Comparators

• Single-Timepoint DRP: A simplified version using only the index exam (no longitudinal data).
• Mirai: A state-of-the-art FFDM-based deep learning risk model (using the same-day FFDM corresponding to the DBT exam).
• Tyrer-Cuzick (v8): A standard clinical risk model based on demographics and family history.

Evaluation Metrics

• Area Under the Receiver Operating Characteristic Curve (AUC) for 2–5 year horizons.
• Time-dependent Concordance Index (C-index).
• Integrated Brier Score (IBS) for calibration.
• Risk Stratification analysis based on NCCN/ASCO thresholds (Average <1.7%, Increased 1.7–2.9%, High ≥3.0%).

3. Key Contributions

• First Longitudinal 3D DBT Risk Model: Demonstrated that integrating native 3D DBT volumes over time significantly outperforms single-timepoint models and 2D-based models.
• Superiority Over FFDM and Clinical Models: Showed that volumetric and temporal information in DBT provides risk signals not captured by FFDM-based AI (Mirai) or clinical risk factors (Tyrer-Cuzick).
• Refinement of Density-Based Risk: Proved that AI can identify high-risk individuals within "fatty" breast categories and low-risk individuals within "extremely dense" categories, challenging the reliance on breast density alone for risk stratification.

4. Key Results

Performance Metrics (Independent Test Set, n=34,580)

• Longitudinal DRP vs. Single-Timepoint DRP: The longitudinal model achieved a 5-year AUC of 0.720 (95% CI: 0.703–0.738), significantly outperforming the single-timepoint model (AUC 0.706; p < 0.001).
• Longitudinal DRP vs. Mirai (FFDM-based): The DRP model achieved a 5-year AUC of 0.720, significantly higher than Mirai's 0.687 (p < 0.001).
• Longitudinal DRP vs. Tyrer-Cuzick (Case-Control, n=432): The DRP model achieved a 5-year AUC of 0.676, significantly outperforming Tyrer-Cuzick's 0.567 (p < 0.001).
• Calibration: The model showed excellent calibration with a mean absolute error of 0.005.

Risk Stratification Insights

• Reclassification of Dense Breasts: The model reclassified 37.6% (705/1,877) of women with extremely dense breasts (BI-RADS D) as "average risk," with a low observed 5-year incidence of 0.7%.
• Identification of High Risk in Fatty Breasts: The model identified 15.5% (404/2,605) of women with fatty breasts (BI-RADS A) as "high risk," with an observed 5-year incidence of 2.5%.
• Discordant Classifications: In cases where models disagreed, the DRP-only high-risk group showed a higher incidence (3.4%) than the Mirai-only high-risk group (3.0%), suggesting DRP captures unique risk factors.

Subgroup Analysis

• Performance was consistent across age, race, and family history, though slightly lower in Asian patients (AUC 0.674), likely due to underrepresentation in the training data.
• The model successfully stratified risk across all breast density categories, showing heterogeneity within density groups.

5. Significance and Clinical Impact

• Personalized Screening: The study supports a shift from static, density-based screening to dynamic, AI-driven risk assessment. By leveraging longitudinal 3D data, clinicians can better identify women who need intensified screening (e.g., MRI) or those who can safely reduce supplemental imaging.
• Beyond Density: The findings highlight that breast density is an imperfect proxy for risk. AI models can detect subtle parenchymal texture changes and temporal evolution that density scoring misses.
• Feasibility: The model utilizes data already present in routine screening workflows (longitudinal DBT exams), making it potentially implementable at the point of care without requiring additional patient data collection.
• Limitations: The study is single-center and retrospective, primarily using Hologic systems. External validation across diverse populations and vendors is required before clinical deployment.

In conclusion, this research establishes that longitudinal Digital Breast Tomosynthesis combined with deep learning offers a superior method for 5-year breast cancer risk prediction compared to current FFDM-based AI and clinical models, paving the way for more precise, individualized breast cancer prevention strategies.