Mammographic breast density is a recognised risk factor for breast cancer, but it fails to capture the complexity and variability of tissue structure observable in imaging. Advances in radiomics now allow detailed analysis of mammographic texture beyond global density measures. A large study explored whether parenchymal phenotypes, identified using radiomic features extracted from full-field digital mammography (FFDM), could independently predict breast cancer risk and the likelihood of false-negative findings or interval cancers. The investigation, inclusive of racially diverse populations, aimed to determine the value of these phenotypes in improving early detection and risk stratification.
Radiomic Feature Analysis and Phenotype Identification
The study adopted a two-stage approach. In the first stage, researchers analysed digital mammograms from approximately 30,000 women across three screening centres. Radiomic features were extracted using an automated pipeline, with over 1000 initial features narrowed to 390 reproducible ones. These features were derived from multiple texture classes, such as co-occurrence patterns and spectral characteristics. Using hierarchical clustering and principal component analysis (PCA), six distinct clusters and six principal components (PCs) were identified, reflecting the underlying heterogeneity of breast tissue. These phenotypes were defined without knowledge of cancer outcomes, ensuring an unbiased classification based solely on image-derived data.
Must Read: Predicting Breast Cancer Using Sequential Mammogram Analysis
The clusters and PCs were assessed for their association with various clinical and demographic factors. For instance, PC1 was notably higher in phenotypes associated with greater cancer risk. This method allowed the identification of structural patterns that might not be evident through conventional density readings. Moreover, the approach proved robust across validation subsets, confirming the reproducibility of the radiomic phenotypes.
Association with Breast Cancer and False-Negative Outcomes
In the second stage, phenotypes were applied to a nested case-control cohort of 1055 women with invasive breast cancer and 2764 matched controls. Associations were tested against cancer incidence, false-negative mammography results and symptomatic interval cancers. PCs—particularly PC1 and PC2—were positively associated with invasive cancer, while PC3 showed an inverse relationship. These findings held across both Black and White women, suggesting broad applicability of the model regardless of race. Notably, inclusion of PCs improved the area under the receiver operating characteristic curve (AUC) for cancer prediction, especially in Black women, where the AUC rose to 0.68 compared with 0.63 without PCs.
The study also assessed whether the phenotypes could predict missed cancers. Clusters showed limited relevance for false-negative or interval cancers, but PCs significantly improved discrimination. For instance, incorporating PCs raised the AUC for false-negative predictions from 0.66 to 0.73, and for symptomatic interval cancers from 0.68 to 0.77. These improvements suggest that certain radiomic patterns may correspond with imaging features that either conceal tumours or indicate aggressive growth between screenings. Importantly, these associations were observed independently of breast density classifications, highlighting the added predictive value of texture analysis.
Implications for Risk Prediction and Equity in Screening
The integration of radiomic phenotypes into risk models offers a promising enhancement over traditional metrics. In particular, the ability to detect phenotypes predictive of both long-term cancer risk and short-term missed diagnoses positions this approach as a tool for refining screening protocols. For women identified as high-risk, more frequent or supplementary screening could be justified, while those at lower risk might avoid unnecessary procedures. Additionally, the consistent associations across racial groups suggest that these models could contribute to reducing disparities in breast cancer detection.
Radiomic phenotypes may also complement machine learning and artificial intelligence systems in breast cancer diagnostics. While previous systems have focused primarily on lesion detection, the addition of nuanced texture patterns may enhance the interpretability and predictive strength of AI models. The study’s approach, which involved harmonising radiomic data across multiple institutions and imaging systems, paves the way for broader implementation in clinical settings.
The study confirmed that parenchymal phenotypes derived from radiomic analysis of mammograms are independently associated with invasive breast cancer and false-negative findings. These phenotypes capture the structural heterogeneity of breast tissue more comprehensively than density alone and offer valuable insights for improving cancer risk prediction. With demonstrated applicability across diverse populations and imaging centres, this method holds promise for enhancing personalised screening strategies and addressing inequities in breast cancer outcomes. Future research should further explore its integration with AI tools and genomic data to maximise its predictive potential.
Source: Radiology
Image Credit: iStock
