Going Beyond Conventional Mammographic Density to Discover Novel Mammogram-Based Predictors of Breast Cancer Risk

John L. Hopper, Tuong L. Nguyen, Daniel F. Schmidt, Enes Makalic, Yun‐Mi Song, Joohon Sung, Gillian S. Dite, James G. Dowty, Shuai LiJournal of Clinical Medicine 2020 · DOI

TL;DR

This paper argues that conventional mammographic density is a poor predictor of breast cancer risk itself, primarily acting as a "mask" for tumors. Instead, the authors introduce novel mammogram-based risk predictors derived from image texture and brightness variations that show significantly stronger predictive power for both interval and screen-detected breast cancers, even outperforming genetic risk scores.

What Problem Does This Paper Address?

For decades, mammographic density—the proportion of white, dense tissue on a mammogram—has been recognized as a risk factor for breast cancer. However, the precise role of density and how to best leverage mammographic information for risk prediction has remained an open question. While efforts have focused on accurately measuring density, this paper posits that the goal should be risk prediction, not just density measurement. The authors highlight that conventional density measures may be confounded by age and body mass index, and their predictive power might be overestimated due to their role in masking tumors rather than directly indicating risk. They aim to identify and validate new mammogram-based risk predictors that are more directly linked to a woman's inherent risk of developing breast cancer.

Methods

The core of the paper's methodology lies in developing and evaluating new mammogram-based risk measures beyond conventional density. They introduce three new measures: Cumulus, Altocumulus, and Cirrocumulus, which are derived from different brightness thresholds of mammographic images, moving beyond just the overall dense area. They also developed Cirrus, a measure based on textural features identified through machine learning. A key methodological contribution is the OPERA (odds ratio per adjusted standard deviation) metric, which provides a standardized way to compare the risk discrimination ability of different factors, accounting for adjustments made for covariates like age and BMI. This metric is crucial for interpreting and comparing risk estimates across various factors. The authors then analyze how these new measures perform individually and in combination with conventional density, using statistical models to assess their predictive power for interval and screen-detected cancers.

Datasets

The research utilizes data from multiple studies, including:

  • Australian and Korean women's mammograms.
  • A prospective cohort study involving Australian women.
  • A sample of American-Hawaiian women.
  • Data from twin pairs for causal inference studies.

The paper does not introduce a new benchmark dataset but rather uses existing datasets to validate their novel measures and compare them against established risk factors.

Key Insights and Ideas

  • Density as a Mask, Not a Cause: Conventional mammographic density is primarily implicated in "masking" tumors, making them harder to detect on mammograms, rather than being a direct causal factor for breast cancer risk itself.
  • Brightness and Texture Hold the Key: New measures derived from higher brightness thresholds (Altocumulus, Cirrocumulus) and textural patterns (Cirrus) within mammograms are significantly better predictors of breast cancer risk than conventional density.
  • OPERA: A Standardized Risk Metric: The OPERA metric is essential for accurately comparing the predictive power of different risk factors, especially when they are measured on different scales or adjusted for covariates.
  • New Measures Outperform Genetics: The novel mammogram-based risk predictors demonstrate superior risk discrimination compared to conventional density and even polygenic risk scores, which are derived from extensive genetic studies.
  • Causal Pathways: The research suggests that while conventional density's role is largely masking, the new measures may have a more direct causal effect on both interval and screen-detected breast cancers.

Key Results

  • New mammogram-based measures (Cirrus, Altocumulus, Cirrocumulus) achieved OPERA values of 1.6 or higher, significantly outperforming conventional mammographic density (OPERA ~1.4).
  • In multivariable analyses, conventional density often attenuated to the null when combined with the new measures, indicating its predictive power was largely explained by these novel factors.
  • The Cirrus measure, derived from textural analysis, showed an OPERA of 1.90 in one analysis, demonstrating strong predictive capability.
  • The paper presents a path diagram suggesting that conventional mammographic density primarily contributes to interval cancer risk through masking, while new mammogram-based risk measures have a more direct causal effect on both interval and screen-detected breast cancers.

Limitations and Future Work

  • Measurement Challenge: The authors acknowledge the challenge of comparing different measures in an open and transparent manner, advocating for standardized datasets and methodologies.
  • Biological Mechanisms: While the new measures show strong predictive power, the specific biological mechanisms linking them to breast cancer risk require further investigation.
  • Causality: While the path diagram suggests causal effects, further research using methods like ICE FALCON is needed to confirm these causal relationships.
  • Automation and Validation: The authors emphasize the need for automated implementation of these new measures and validation across diverse populations, including different racial and ethnic groups, to address disparities in breast cancer outcomes.
  • Integration with Other Risk Factors: Future work should integrate these new mammogram-based predictors with other established risk factors like family history and polygenic risk scores for a more comprehensive risk assessment.

Summary updated

Abstract

This commentary is about predicting a woman's breast cancer risk from her mammogram, building on the work of Wolfe, Boyd and Yaffe on mammographic density. We summarise our efforts at finding new mammogram-based risk predictors, and how they combine with the conventional mammographic density, in predicting risk for interval cancers and screen-detected breast cancers across different ages at diagnosis and for both Caucasian and Asian women. Using the OPERA (odds ratio per adjusted standard deviation) concept, in which the risk gradient is measured on an appropriate scale that takes into account other factors adjusted for by design or analysis, we show that our new mammogram-based measures are the strongest of all currently known breast cancer risk factors in terms of risk discrimination on a population-basis. We summarise our findings graphically using a path diagram in which conventional mammographic density predicts interval cancer due to its role in masking, while the new mammogram-based risk measures could have a causal effect on both interval and screen-detected breast cancer. We discuss attempts by others to pursue this line of investigation, the measurement challenge that allows different measures to be compared in an open and transparent manner on the same datasets, as well as the biological and public health consequences.
BibTeX
@article{hopper2020going,
  author = {John L. Hopper and Tuong L. Nguyen and Daniel F. Schmidt and Enes Makalic and Yun‐Mi Song and Joohon Sung and Gillian S. Dite and James G. Dowty and Shuai Li},
  title = {{Going Beyond Conventional Mammographic Density to Discover Novel Mammogram-Based Predictors of Breast Cancer Risk}},
  journal = {Journal of Clinical Medicine},
  year = {2020},
  doi = {10.3390/jcm9030627},
}
Figure