Abstract
In a mammogram, the lighter (less dense) areas have a causal effect on the brightest (highly dense) areas, including through a causal pathway via textural features. These causal relationships help us gain insight into the relative aetiological importance of different mammographic features in breast cancer. For example, our findings are consistent with the brightest areas being more aetiologically important than lighter areas for screen-detected breast cancer; conversely, light areas being more aetiologically important for interval breast cancer. Additionally, specific textural features capture aetiologically independent breast cancer risk information from dense areas. These findings highlight the utility of ICE FALCON and family data in decomposing the associations between intercorrelated disease biomarkers into distinct biological pathways.
Introduction
Mammographic density—the amount of white, radiopaque tissue on an X-ray—is a well-established risk factor for breast cancer. However, "density" is not a monolithic trait. It can be measured at various brightness thresholds (e.g., Cumulus, Altocumulus, and Cirrocumulus) and through complex textural patterns (e.g., the Cirrus score).
Because these measures are highly correlated, it has historically been unclear whether they represent independent biological pathways or if one feature causally influences the development of another. Furthermore, these associations are often clouded by "familial confounding"—shared genetic or environmental factors that influence all these traits simultaneously. This study seeks to untangle these correlations to determine the actual causal flow between mammographic features and their relative importance in cancer etiology.
Methods
The researchers employed a novel approach called ICE FALCON (Inference about Causation from Examination of FAmiliaL CONfounding). This method is analogous to Mendelian Randomization but uses a co-twin’s exposure as a proxy instrumental variable. By analyzing monozygotic (identical) twins, the model can effectively "subtract" the effects of shared genetics and environment to isolate direct causal links.
Feature Extraction
Mammograms were decomposed into three non-overlapping spatial zones based on brightness thresholds:
- Light areas: Low-density regions.
- Bright areas: Medium-density regions.
- Brightest areas: High-density regions (Cirrocumulus).
In addition to brightness thresholds, the study used Cirrus, an agnostic deep-learning algorithm that generates a risk score based on 20 textural features, independent of simple pixel brightness counting.
Dataset
The study utilized data from the Australian Mammographic Density Twins and Sisters Study (AMDTSS), comprising 371 monozygotic twin pairs (742 individuals). Participants were women aged 40–70 with no prior breast cancer diagnosis at the time of mammography. The analysis was performed on digitized craniocaudal-view film mammograms.
Results
The analysis revealed a clear directional hierarchy in mammographic development. Lighter (less dense) areas were found to causally influence the formation of bright areas, which in turn cause the development of the "brightest" high-density regions.
Key Causal Insights
- Causal vs. Confounded: The researchers estimated that 28–92% of the associations between different mammogram risk scores are actually causal, rather than just being products of shared genes.
- Direct Effects: Bright areas have a significant causal effect on the brightest areas, accounting for 85% of their association. Light areas account for 55% of the association with the brightest areas.
- Texture as a Mediator: Textural features (Cirrus) act as a causal bridge, accounting for 34% of the association with the brightest areas. This suggests that light and bright areas influence the breast's texture, which then contributes to the development of high-density tissue.
Aetiological Specialization
The findings explain why different features predict different types of cancer. The "brightest" areas appear to be the primary drivers for screen-detected cancers. Conversely, "light" areas are more important for interval cancers (cancers that appear between regular screenings), likely because they represent the background tissue where tumors can be masked.
Discussion and Limitations
While the study establishes a robust causal framework, several limitations remain. The research was conducted on digitized film mammograms; while similar results are expected on modern digital systems, replication is currently underway. Additionally, while 371 twin pairs is substantial for this type of analysis, subgroup power may be limited. Finally, while the paper establishes that these causal relationships exist, the specific biological mechanisms (e.g., hormonal or cellular changes) that drive these transitions remain a subject for future research.
Conclusion
By using ICE FALCON to decompose intercorrelated biomarkers, this study provides a biological roadmap for how different mammographic features contribute to breast cancer risk. These findings highlight that specific textural features capture risk information independent of dense areas, offering a more nuanced understanding of breast cancer etiology.
Summary updated