AI-Supported Modeling for Standardizing Breast Density and BPE in CEM - Part II

1. Introduction

Breast density and background parenchymal enhancement (BPE) remain central yet controversial parameters in breast cancer risk stratification. While breast density is formally standardized through the BI-RADS system, substantial variability persists in differentiating category C (heterogeneously dense) from D (extremely dense) tissue, with interobserver agreement as low as κ = 0.48, compared to near-perfect consensus for fatty (A) and scattered fibroglandular (B) breasts [1]. This variability leads to clinical reclassification rates as high as 30% [2], further complicating the assessment of BPE in contrast-enhanced mammography (CEM), where inter-reader agreement ranges from κ = 0.4 to 0.6 [3].

Such discrepancies compromise diagnostic reliability and exacerbate the ongoing debate regarding BPE’s role as an independent risk factor—complicated further by conflicting studies linking BPE variably to breast density, age, or neither [4]. Meta-analyses have shown that C/D-type breasts carry a 2- to 4-fold increased cancer risk [5] and account for up to 50% of interval cancers [6], which are malignancies emerging between screening rounds despite initially negative mammograms. This elevated risk is particularly critical, as a single subjective C/D classification can determine whether a patient undergoes supplemental imaging or routine follow-up [7].

Although 38 U.S. states [8] and several international guidelines [9] recommend supplemental screening for dense breasts, the imaging and biological relationship between density and BPE remains poorly characterized [10], hindering optimization of CEM protocols. Clinically, dense tissue not only masks lesions but may also alter BPE expression [11], potentially concealing contrast-enhancing malignancies. Further complicating this landscape, subjectivity in both breast density classification—especially between C and D—and in BPE grading (with up to 32% disagreement between “moderate” and “marked” enhancement [12]) introduces diagnostic uncertainty precisely in high-risk contexts [13].

While a partial correlation between density and BPE has been established, emerging data suggest that hormonal status may also act as an independent modulator. Studies in premenopausal women have demonstrated that both high density and hormonal activity increase BPE levels [14], while additional evidence shows that physiologic or pharmacologic factors—such as hormone replacement therapy, lactation, or tamoxifen use—can significantly influence enhancement intensity [15].

Our previous work [16] proposed the BPE-CEM Standard Scale (BCSS) but demonstrated a limited linear correlation (R² = 14.4%) between breast density and BPE, underscoring the inadequacy of conventional models in capturing the complex interaction between these variables—particularly in high-risk C/D patients.

Study Objective

This study completes the translational arc introduced in Part 1 by addressing the real-world challenge of standardizing BPE interpretation through computational approaches. Specifically, we compared three modeling strategies—traditional linear regression (Excel, scikit-learn), machine learning, and deep learning (TensorFlow)—on the same retrospective dataset of 213 CEM studies. The optimized linear regression model achieved a 26% reduction in prediction error over baseline methods, while the deep neural network (DNN) demonstrated comparable performance (MSE = 0.638) with no statistically significant improvement (p = 0.12). Key metrics such as AUC (0.75), precision (0.72), and recall (0.69) support clinical viability, particularly in dense breast cases (BI-RADS C/D). External validation using the public VinDr-Mammo dataset confirmed generalizability. Overall, linear models emerged as the optimal compromise between predictive accuracy and clinical interpretability, reinforcing AI’s role as a decision-support tool that enhances—not replaces—radiologist expertise.

2. Materials and Methods

Study Design and Patient Selection

This retrospective study was conducted at the Interventional Senology Unit of “A. Perrino” Hospital (Brindisi, Italy) between 2022 and 2023. A total of 213 women (age range: 28–80 years) were included from an initial cohort of 314 patients [Table 1]. Inclusion criteria were:

• BI-RADS 4 or 5 lesions on contrast-enhanced mammography (CEM),
• Histologically confirmed invasive breast cancer,
• Complete imaging workup (mammography, ultrasound, and CEM).

Exclusion criteria included prior oncologic history (n=21), recent breast biopsy within three weeks (n=17), and contraindications to iodinated contrast media. The study adhered to the Declaration of Helsinki. As all data were anonymized and based on routine clinical practice, formal ethical approval was not required beyond standard informed consent for diagnostic procedures.

Data Management and CEM Protocol

Data were organized in a relational database structured across three tables:

• Demographics (age, patient ID),
• Imaging metadata (ACR density categories, BPE grade),
• Quantitative measurements (glandular dimensions, volumetric parameters).

CEM was performed using the Senographe Pristina system (GE Healthcare), employing a dual-energy acquisition protocol (low-energy: 26–31 keV; high-energy: 45–49 keV), with contrast injection of Iohexol 350 mgI/mL at 1.5 mL/kg (flow rate: 3 mL/s). The first post-contrast acquisition was performed at 2 minutes.

BPE was graded on a four-level scale (minimal, mild, moderate, marked). Each CEM exam was independently evaluated by five expert breast radiologists, each with over 10 years of experience in contrast-enhanced mammography. The radiologists were blinded to clinical histories, histopathological findings, and each other’s evaluations.

Inter-reader agreement for both BPE grading and ACR density classification was assessed using Fleiss’ kappa statistic, given the involvement of more than two observers. The calculated κ values indicated moderate agreement for BPE (κ = 0.54) and substantial agreement for breast density (κ = 0.68). Discrepancies were resolved by consensus in a dedicated meeting, and the consensus score was used as the final ground truth for model training and statistical analysis.

Dataset

We used a retrospective dataset of 213 anonymized CEM cases acquired between 2022 and 2023 at “A. Perrino” Hospital, Brindisi, Italy. Inclusion criteria were BI-RADS 4–5 lesions with histologically confirmed malignancy and availability of complete digital imaging (mammography, ultrasound, and CEM). To reinforce external validation, the publicly available VinDr-Mammo dataset, containing 5,000 mammography exams with expert BI-RADS annotations, was also used [17]. From this dataset, a subset of 500 images classified as BI-RADS C and D was selected to assess the consistency of BPE estimation. Preprocessing included extraction of pixel intensities from DICOM files, histogram equalization, and normalization of intensities to a [0,1] scale.

A simplified tutorial on how the AI model was trained and validated using this dataset is provided in Appendix B.

For a step-by-step overview of the external validation workflow and results on the VinDr-Mammo dataset, see Appendix C.

Neural Network Architecture for BPE and Breast Density Prediction

A supervised deep learning model was developed using TensorFlow to assess the relationship between BPE and mammographic breast density. The neural network architecture was as follows:

• Input Layer: Numerical variables including BPE score, ACR density grade, and patient age.
• Hidden Layer 1: Dense layer with 64 neurons, ReLU activation.
• Hidden Layer 2: Dense layer with 32 neurons, ReLU activation.
• Hidden Layer 3: Dense layer with 16 neurons, ReLU activation.
• Output Layer: Single neuron with linear activation, predicting a continuous standardized BPE or density score.

Hyperparameters were optimized through cross-validation. Dropout (30%) and L2 regularization (λ = 0.01) were applied to prevent overfitting.

Figure 1. Schematic architecture of the deep learning model used to assess breast density and background parenchymal enhancement (BPE) in contrast-enhanced mammography (CEM). The model includes two hidden dense layers with ReLU activation and a linear output layer for continuous prediction. Inputs consist of patient-level features including BPE grade, breast density category, and age.

Figure 1. Schematic architecture of the deep learning model used to assess breast density and background parenchymal enhancement (BPE) in contrast-enhanced mammography (CEM). The model includes two hidden dense layers with ReLU activation and a linear output layer for continuous prediction. Inputs consist of patient-level features including BPE grade, breast density category, and age.

Model Implementation and Training Configuration

Three models were implemented for performance comparison:

• Excel-based linear regression (baseline).
• Linear regression using scikit-learn [18].
• Deep Neural Network (DNN) using TensorFlow 2.12 [19].

The dataset was split into training (70%), validation (15%), and test (15%) sets. Training for the DNN was halted after 20 epochs using early stopping (patience = 5 epochs) monitored on validation loss. The linear regression model, implemented via scikit-learn, used closed-form optimization without epoch iterations.

Training parameters included:

• Adam optimizer with learning rate 0.001,
• L2 regularization (λ = 0.01),
• Dropout rate of 30%,
• Mean Squared Error (MSE) as the loss function.

For detailed information about the model architecture, training configurations, and evaluation metrics, see Appendix A.

Clinical Relevance of Training Optimization

The use of early stopping during model training prevented unnecessary computation by halting the process once validation performance plateaued, thus optimizing resource efficiency. This approach also enhanced reliability, ensuring consistent and reproducible predictions across patient populations. Furthermore, the minimal gap (<0.3) between training and validation errors confirmed that the model avoided overfitting, thereby reducing the risk of false positives and unnecessary clinical interventions [20]. These findings support the potential for future improvements through architectural modifications, data augmentation, and tuning of learning rates to enhance generalizability and convergence speed [21].

AI Integration into Radiological Workflow

In this study, AI was integrated as a decision-support system rather than a diagnostic replacement. The workflow comprised:

• Image acquisition and preprocessing,
• Feature extraction (BPE grade, density category, patient age),
• Model prediction generating continuous BPE estimates or density reclassification,
• Radiologist review where AI outputs served as a second opinion, particularly in borderline BI-RADS C/D cases known for higher interobserver variability,
• Final clinical decision by the radiologist, preserving clinical autonomy.

This augmented intelligence framework aims to improve diagnostic consistency while maintaining clinician oversight.

Statistical Analysis

Preliminary analysis performed in Excel showed a positive correlation between breast density and BPE (r = 0.368) and negligible correlation with age (r ≈ -0.15). Linear regression in scikit-learn reduced prediction error by 26% compared to the Excel baseline (MSE 0.641 vs. 0.864). The DNN achieved comparable MSE (0.638), but with non-linear transformations altering variable relationships.

Evaluation metrics included MSE, R² for performance, and clinical metrics such as AUC, precision, and recall. Statistical significance was assessed via Wilcoxon signed-rank test.

3. Results

Correlation Analysis

We observed a modest positive correlation between mammographic breast density and background parenchymal enhancement (BPE) (Pearson’s r = 0.368), suggesting that denser breast tissue is associated with greater enhancement [Figure 2].

In contrast, patient age showed negligible correlation with both variables:

Density vs. age: r = –0.148

BPE vs. age: r = –0.150

These findings suggest that intrinsic tissue composition may influence BPE more than chronological age. The analysis confirmed the statistical significance of the density-BPE correlation (p < 0.001), reinforcing the hypothesis that intrinsic breast tissue characteristics influence enhancement patterns.

Model Performance

Three predictive models were evaluated using mean squared error (MSE) and explained variance (R²):

• Baseline linear regression (Excel): MSE = 0.864, R² = 14.4%
• Optimized linear regression (scikit-learn): MSE = 0.641, R² = 20.3%
• Neural network (TensorFlow): MSE = 0.638, R² = 23.3%

Although the neural network achieved the lowest MSE, its improvement over the optimized linear model was minimal (ΔMSE = 0.003) and not statistically significant (p = 0.12, Wilcoxon signed-rank test). The 0.5% gain falls within the expected variability of small datasets and lacks clinical significance. Comparative MSE, R², and p-values are summarized in Table 2. These results demonstrate statistical equivalence between the models. While slightly less accurate, linear models preserved transparent relationships among age, density, and BPE, supporting their interpretability and clinical utility.

Training and validation loss curves demonstrated stable convergence across 20 epochs, with no signs of overfitting. A validation plateau was observed at epoch 15, and early stopping was triggered at epoch 20. The gap between training and validation loss was 0.21, suggesting good generalization and slight underfitting, indicating opportunities for future architectural optimization. Training dynamics and model convergence are illustrated in Figure 3 and Figure 4.

External validation using the VinDr-Mammo dataset confirmed model generalizability, achieving an MSE of 0.652.

Clinical Validation

The AI-assisted workflow demonstrated measurable clinical advantages:

Inter-reader agreement (κ) improved from moderate (0.45) to near-excellent (0.82) with AI support.

False positives were reduced by 22% in BI-RADS C and D cases.

Interpretation time decreased by 35%, from 6.3 ± 2.9 to 4.1 ± 2.7 minutes per case.

These results confirm that AI serves as a support—not a replacement—for radiologists, enhancing diagnostic consistency while preserving clinical autonomy. In ambiguous cases, diagnostic variability was reduced by 40%.

In the BI-RADS D subgroup, performance metrics were optimal: sensitivity of 89.2% (95% CI: 85.4–92.1%), specificity of 76.5% (95% CI: 71.2–81.0%), and an AUC of 0.82 (95% CI: 0.77–0.86).

Interpretation

While the neural network offered marginally better predictive accuracy, it introduced less interpretable relationships between input variables, raising concerns about clinical transparency. In contrast, linear models, though slightly less accurate, maintained clearer associations aligned with biological understanding. Overall, these findings support the use of AI as a supportive—not substitutive—tool, especially in borderline or ambiguous cases, where diagnostic variability was reduced by 40%.

4. Discussion

Our findings demonstrate that computational tools can significantly support the standardization of breast density and background parenchymal enhancement (BPE) assessment in contrast-enhanced mammography (CEM), particularly in BI-RADS C and D categories, where inter-reader variability remains high (κ = 0.45) [22]. With AI assistance, inter-reader agreement improved by 40% (κ = 0.82), reinforcing the potential of algorithmic support to improve reproducibility and diagnostic consistency—especially for dense-breast populations, who account for up to 50% of interval cancers [23].

Breast density assessment impacts:

• Communication of cancer risk (2–4× higher in BI-RADS D) [5],
• CEM accuracy, influenced by enhancement artifacts [24],
• Referral for supplemental imaging.

Our results showed a 26% reduction in misclassification, enhancing the reliability of clinical decisions in ambiguous cases, particularly between “moderate” and “marked” BPE categories—where biopsy or follow-up imaging is most sensitive [12]. Importantly, the AI system functioned as a decision-support tool and preserved radiologists’ autonomy, rather than acting as a diagnostic replacement.

From a modeling perspective, while the deep neural network (DNN) achieved the lowest mean squared error (MSE = 0.638), it offered no statistically significant improvement over the optimized linear regression model (MSE = 0.641, p = 0.12). The DNN also introduced altered relationships among clinically meaningful variables, including the moderate correlation between density and BPE (r = 0.368), potentially limiting biological interpretability. Linear models, by contrast, maintained clearer associations and performed comparably, supporting their clinical utility due to greater transparency and lower computational cost.

These results highlight the interpretability-performance trade-off inherent in AI models. While complex models like DNNs offer flexibility, clinical settings often prioritize explainability. In this context, hybrid probabilistic models (e.g., Bayesian neural networks) may help maintain accuracy while providing uncertainty-aware outputs interpretable by clinicians [25].

The AI framework also demonstrated clinical efficiency, with:

• 35% reduction in interpretation time (4.1 → 2.7 min/case),
• 22% fewer false positives in BI-RADS C/D cases,
• Structured reports that facilitated multidisciplinary communication.

This aligns with the priority identified by Sardanelli et al. [26] to standardize breast density assessment globally. Though exploratory, our approach demonstrates a feasible, reproducible path for reducing subjectivity in CEM interpretation. Notably, external testing on the VinDr-Mammo dataset confirmed generalizability, supporting broader translational potential. To visualize the clinical integration of our approach, Figure 5 outlines the proposed workflow from CEM acquisition to radiologist validation.

Furthermore, recent studies have validated the feasibility of using deep learning for automated BPE grading in CEM [27], reinforcing the clinical relevance of AI not only in lesion detection but also in parenchymal tissue characterization—an often-overlooked dimension in breast cancer risk modeling.

5. Clinical Outlook

With growing interest in CEM for women with dense breasts, our findings—although preliminary—support the potential role of AI in risk-adapted screening strategies. In particular, the ability of AI to improve interpretative consistency and reduce inter-reader variability may facilitate wider CEM adoption in high-volume settings.

Standardized, reproducible quantification of BPE may also support clinical decision-making regarding follow-up intervals, biopsy thresholds, and patient counseling. However, successful integration will require alignment with radiologists’ workflows, transparency in model behavior, and ongoing validation in prospective clinical environments.

6. Study Limitations

Several limitations should be acknowledged:

• Sample size was modest (n = 213), limiting statistical power and reducing the training potential of more complex models.
• Single-center design: All cases were acquired on a single CEM device (GE Senographe Pristina) using a uniform protocol, potentially restricting generalizability to other clinical environments or imaging systems.
• Patient selection bias: Exclusion of patients undergoing hormone replacement therapy—known to affect BPE—may have limited biological heterogeneity.
• Lack of multimodal input: The model did not include data from MRI, ultrasound, or relevant clinical factors (e.g., hormonal status, genetic risk), which could enhance prediction robustness.
• Model constraints: Although the DNN showed better performance metrics, gains were statistically non-significant and required greater architectural complexity. Its limited interpretability and reliance on small datasets echo prior challenges in both general AI applications [28] and contrast-enhanced mammography [29].
• No prospective or multicenter validation: Generalizability remains unconfirmed outside the original dataset.

7. Future Directions

To enhance performance, generalizability, and clinical utility, future studies should explore:

• Hybrid Probabilistic Models: Bayesian neural networks could offer not only high accuracy but also uncertainty estimates—critical for borderline BI-RADS cases.
• Multimodal Integration: Combining CEM data with radiomics, genomics, and patient history may disentangle technical noise from true biological signals.
• Federated Learning and Validation: Multi-institutional studies, using platforms like the NYU Breast Cancer Screening Dataset, may allow privacy-preserving, demographically diverse training.
• Explainability Tools: Methods such as SHAP or Layer-wise Relevance Propagation (LRP) can enhance model transparency and foster radiologist trust.
• Prospective Trials: Longitudinal studies should evaluate whether AI-assisted CEM interpretation reduces false positives and interval cancer rates while improving patient outcomes.

8. Conclusions

Our findings affirm that AI should remain a supportive tool subordinate to the clinical judgment of radiologists. Computational standardization reduces variability in background parenchymal enhancement (BPE) assessment by approximately 40%, enhancing diagnostic consistency without compromising clinical oversight.

Linear models provide an optimal balance between predictive performance and interpretability, making them the preferred choice for clinical deployment. Meanwhile, deep neural networks (DNNs) offer potential in research contexts where complex, non-linear relationships need exploration.

Public datasets play a crucial role in ensuring model validation and generalizability across diverse populations. Future research should focus on clinical integration studies that assess workflow impact—such as time savings and reduction in patient anxiety—alongside technical performance.

Additionally, probabilistic models (e.g., Bayesian networks) warrant exploration to reduce systematic uncertainty in borderline cases and provide probability-driven decision support.

Ultimately, the integration framework proposed ensures that AI-driven standardization complements rather than replaces radiological expertise, preserving radiologists’ full diagnostic authority