Breast density estimation is critical for assessing BC risk and improving mammographic screening accuracy. Traditional approaches rely on radiologists’ visual assessments or semi-automated software, which can be subjective and inconsistent Park et al. (9). Deep Learning (DL), particularly Convolutional Neural Networks (CNNs), offers a more objective and standardized approach Saffari et al. (17). By training CNNs on labeled mammographic datasets (BI-RADS categories), these models can accurately identify patterns and textures associated with different density levels, reducing variability and outperforming conventional methods Khara et al. (14)Saffari et al. (17).

Tumor morphology, including shape, size, and margins, plays a crucial role in characterizing breast cancer and assessing its aggressiveness Ma et al. (25) Gillies et al. (26). Traditional feature extraction methods rely on handcrafted techniques, which may overlook critical tumor characteristics SIN (2020). CNNs provide an automated and comprehensive approach by learning from annotated datasets to detect and segment tumor regions with high precision (27, 28). These models can uncover microtextural patterns linked to specific molecular subtypes, offering deeper insights into tumor behavior and treatment response Sutton et al. (29). Advanced techniques, such as multi-task learning and attention mechanisms, further enhance feature extraction by emphasizing the most relevant tumor regions. Recent advancements in radiomics and machine learning have enabled a more refined prediction of molecular subtypes, leveraging quantitative imaging biomarkers Yap et al. (30). CNNs efficiently capture spatial relationships and patterns, facilitating accurate classification of tumor morphology.

In this work, we implemented a one-dimensional convolutional neural network (1-D CNN) tailored to the challenge of BC relapse prediction. As shown in the Figure 2 , the model architecture initially emphasizes a streamlined process for converting features into high-dimensional representations to reflect better the spatial or sequential relationships within the underlying radiomics features Mathew et al. (31). The newly transformed features are then reshaped into a multi-channel pseudo-image format, effectively simulating a structured spatial domain. By expanding the feature representation to emulate structured spatial inputs, the network adapts to the morphological variability inherent in the dataset.

Further, deeper convolutional layers refine the learned features, providing robustness against variability in patient-specific tumor characteristics. Unlike traditional machine learning approaches such as Random Forest or XGBoost, which require extensive manual feature engineering, this deep learning framework learns these representations end-to-end, demonstrating adaptability to high-dimensional data. The final stages of the network incorporate a global pooling operation and fully connected layers with dropout for regularization, culminating in a sigmoid activation for binary relapse classification. This architecture integrates high-quality tumor features and leverages the inherent benefits of convolutional operations, such as spatial locality and feature abstraction.

The dataset used in this study was developed in collaboration with our partner hospital, following ethical data collection and patient privacy guidelines. This dataset comprises clinical and pathological records for 148 BC patients, resulting in a total of 270 mammographic samples from both craniocaudal (CC) and mediolateral oblique (MLO) views, providing a robust foundation for our analysis. Our dataset includes detailed pathological annotations encompassing the following features. Molecular subtypes are classified into four categories: Luminal A (0), Luminal B (1), HER2-positive (2), and Triple-Negative (3). Tumor shape is categorized as oval (0), round or irregular (1), asymmetric (2), and distorted (3). Tumor margins are labeled as circumscribed (0), obscured or microlobulated (1), ill-defined (2), and spiculated (3). Breast density is recorded using the BI-RADS classification system, including entirely fatty (0), scattered fibroglandular tissue (1), heterogeneously dense (2), and extremely dense (3). Relapse status is encoded as a binary outcome, where 0 denotes no recurrence and 1 indicates documented relapse during the follow-up period. This comprehensive information is valuable for analyzing breast cancer characteristics and assessing relapse likelihood based on each patient’s unique tumor anatomy.

Figure 3 provides a visual summary of the distribution of key breast cancer features in the dataset through six histograms. Breast density is predominantly classified as type 1 (scattered fibroglandular tissue), with smaller peaks at types 0 and 2, reflecting patterns common in screening populations. Tumor shape is most frequently type 1 (round or irregular), while types 0 (oval) and 3 (distorted) occur less often, and type 2 (asymmetric) is rare. Tumor margins are distributed primarily across non-circumscribed categories (types 1–3), with fewer circumscribed cases (type 0). Tumor sizes, measured in both CC and MLO views, follow a right-skewed distribution, with most lesions between 5 and 30 mm and a peak near 10 mm. Molecular subtypes are imbalanced, with Luminal A (0) and Luminal B (1) being the most common, while HER2-positive (2) and Triple-Negative (3) subtypes are less frequent but clinically more aggressive. Relapse cases are relatively few, introducing a moderate class imbalance that should be considered in downstream modeling.

To explore the relationship between tumor characteristics and relapse, we examined feature distributions across relapse groups using violin plots ( Figure 4 ). Patients who experienced relapse tended to have larger tumors, were more likely to present with aggressive molecular subtypes such as HER2-positive and Triple-Negative, and often exhibited irregular shapes and poorly defined margins—traits commonly associated with higher relapse risk. These patterns were further supported by Kendall’s rank correlation analysis ( Figure 5 ), which showed moderate positive correlations between relapse and tumor size (τ = 0.23), as well as relapse and molecular subtype (τ = 0.24). A stronger correlation between tumor shape and margin (τ = 0.37) suggests a consistent association between morphological irregularity and invasive features. Some other feature pairs had near-zero correlations, suggesting minimal linear associations and pointing to the independence of many BC characteristics.

To predict BC relapse, we implemented a one-dimensional convolutional neural network (1D-CNN) model leveraging critical radiomic features extracted from digital mammograms. Additionally, a series of conventional machine learning models—including Random Forest, Support Vector Machines (SVM), Logistic Regression, and XGBoost—were systematically evaluated to establish comprehensive comparative insights and benchmark the performance of our CNN approach in Table 1 .

The results indicate that the 1D-CNN model achieves comparatively better predictive performance than the other models evaluated. Specifically, the CNN model achieved an average area under the receiver operating characteristic curve (AUC) of 0.9579, significantly surpassing the performance of Random Forest (AUC = 0.8986), XGBoost (AUC = 0.9116), CatBoost(AUC = 0.8672). Beyond AUC, critical evaluation metrics, including sensitivity, specificity, precision, recall, and F1-score (shown comprehensively in Table 1 ), consistently highlighted the superior robustness and reliability of the CNN approach compared to traditional machine learning methods.

We conducted both McNemar’s and DeLong’s statistical significance tests to evaluate pairwise differences in classification accuracy and AUC, respectively, to assess the performance advantage of our proposed 1D-CNN model rigorously. McNemar’s test revealed that the CNN model achieved statistically significant improvements in classification accuracy over all traditional machine learning models evaluated: Random Forest (statistic = 6.0, p = 0.0023), XGBoost (statistic = 8.0, p < 0.0001), and CatBoost (statistic = 4.0, p < 0.0001). Complementarily, DeLong’s test demonstrated that the CNN significantly outperformed CatBoost (z = -3.6436, p = 0.0008) and Random Forest (z = -1.5992, p = 0.0420) in terms of AUC. However, the difference between CNN and XGBoost was not statistically significant (z = -1.7748, p = 0.0608). Collectively, these results suggest a robust and statistically validated advantage of the CNN model in classification performance relative to conventional machine learning models.

The observed performance differences can be attributed primarily to the CNN architecture’s inherent capacity to capture complex spatial and sequential relationships within radiomic features. Although our model used only five fundamental radiomic features—tumor size, tumor shape, margin characteristics, molecular subtype, and breast density—the CNN architecture facilitated a highly effective conversion of these limited input features into sophisticated, high-dimensional representations. This capacity enabled CNN to better interpret subtle but critical patterns within the radiomic data, capturing interactions and non-linear relationships missed by conventional machine learning models.

We conducted comprehensive feature importance assessments and SHAP (SHapley Additive exPlanations) analysis to rigorously interpret and validate our predictive outcomes and further enhance clinical acceptability. As shown in Figure 6 , the feature importance ranking highlighted molecular subtypes, tumor shape, and tumor size as the most critical predictive features associated with breast cancer relapse. The high significance of molecular subtypes aligns closely with existing clinical literature, as aggressive subtypes such as HER2-positive and Triple-Negative breast cancers (TNBC) are known to have a substantially higher tendency for recurrence due to their invasive biological behavior. Additionally, metastatic status at the time of initial diagnosis is a well-established, independent prognostic factor, particularly in HER2-positive BC patients, who exhibit significantly higher risk of relapse and worse survival outcomes due to the aggressive nature and high proliferative capacity of HER2-enriched tumors. Tumor shape also emerged as an essential predictive feature, supported by clinical evidence indicating that tumors exhibiting irregular shapes or spiculated margins typically correlate with greater malignancy potential, higher invasiveness, and consequently increased likelihood of future relapse.

Figure 7 illustrates that SHAP analysis offered comprehensive insights into feature contributions, improving model interpretability. Notably, SHAP analysis demonstrated that molecular subtypes, particularly HER2-positive and Triple-Negative cases, had the strongest impact on relapse likelihood. Additionally, tumor shape irregularities and larger tumor size were positively associated with increased relapse risk, aligning with clinical evidence that these characteristics often indicate aggressive tumor behavior. These insights enrich clinical interpretability, providing a transparent understanding of essential predictive factors influencing relapse risk.

Our analysis focuses on confirmed mass cases, which represent approximately 80% of breast cancer cases. We excluded cases where the tumor was unclear or where high breast density obscured tumor visibility, as radiomic feature extraction in such instances demands additional imaging modalities like MRI or ultrasound for accurate tumor localization and classification. These modalities are crucial for identifying non-mass lesions and ensuring a more comprehensive relapse prediction framework.

To improve the generalizability of our relapse prediction framework, future work will incorporate data from multiple institutions that reflect variations in imaging protocols, patient demographics, and clinical settings. Integrating such heterogeneous data sources will facilitate broader applicability across real-world scenarios, including cases with high breast density and non-mass lesions. Moreover, the inclusion of multi-modal imaging, such as MRI and ultrasound, will support more accurate tumor characterization across a broader range of clinical presentations.