4.1 Summary and key findings

This study highlights the potential of utilizing a 2.5D deep learning model to process digital breast tomosynthesis (DBT) data for early breast cancer diagnosis. By leveraging contiguous slices to capture the three-dimensional structure of the breast, the 2.5D model offers significant advantages in tumor feature extraction over traditional two-dimensional (2D) models while reducing computational burden [8]. To enhance performance and interpretability, we used external machine learning classifiers, such as support vector machines and random forests, to analyze the high-dimensional feature vectors extracted by the deep learning model. This combined approach leverages the strengths of different algorithms, resulting in more precise lesion classification.

Our deep learning model exhibited high accuracy and AUC values across different breast densities, including ACR C and D. In the test set, the logistic regression, LightGBM, and multilayer perceptron models achieved accuracies of 72.2%, 75.0%, and 79.2%; AUC values of 0.826, 0.756, and 0.859; sensitivities of 63.8%, 70.2%, and 80.9%; and specificities of 88.0%, 84.0%, and 76.0%, respectively. Additionally, the model demonstrated significant potential in identifying seemingly benign cancers, aiding early diagnosis and treatment. These results confirm the robustness of the model under various conditions.

In women with higher breast density, breast tissue appears whiter on imaging, potentially obscuring underlying lesions and reducing detection accuracy [9]. ACR C and D represent high-density breast tissues, which are more challenging to detect in clinical practice. Our AI model maintained high accuracy and AUC values even in these high-density tissues by fully utilizing the three-dimensional information of DBT images, better identifying and segmenting fine structures within the breast. This robustness stems from the model’s ability to leverage spatial information during training, thus reducing the impact of high-density tissue on detection performance.

Furthermore, the AI model demonstrated significant potential in identifying seemingly benign cancers [10]. In clinical practice, BI-RADS 3 and 4 lesions are typically considered uncertain and require further imaging follow-up or biopsy for confirmation [11]. Traditional imaging methods have a certain rate of misdiagnosis in identifying these lesions. However, our AI model, through deep learning and feature extraction from DBT images, can more accurately identify these seemingly benign lesions. This not only aids in early diagnosis and timely treatment but also reduces unnecessary follow-ups and biopsies, lowering patient anxiety and medical costs.

4.2 Technical advantages

One of the major technical innovations of this study is the adoption of the 2.5D deep learning model, which improves feature extraction accuracy by incorporating information from multiple adjacent slices and effectively utilizing the spatial information of DBT images. This approach achieves a good balance between computational efficiency and detection accuracy [12].

First, the 2.5D deep learning model exhibits significant advantages in feature extraction. While traditional 2D models have advantages in computation speed, they often extract features from a single slice, lacking utilization of depth information. In contrast, 3D models can comprehensively utilize the spatial information of images, but their high computational complexity results in longer training and inference times and greater resource consumption. The 2.5D model, by jointly learning from adjacent slices, captures three-dimensional features to some extent while avoiding the high computational cost of full 3D modeling. This approach not only enhances the model's feature extraction capabilities but also significantly improves detection accuracy while maintaining computational efficiency [13].

Specifically, the 2.5D model inputs multiple adjacent DBT slices into the pre-trained 2D ResNet-50 model, allowing the model to capture more contextual information during the learning process. This method leverages the three-dimensional characteristics of DBT images, enabling the model to better identify and segment fine structures within the breast when processing high-density breast tissue, thereby reducing the impact of high-density tissue on detection performance [14]. For instance, in this study, the 2.5D model outperformed traditional 2D models in high-density breast tissues such as ACR C and D, showcasing its advantages in processing complex imaging data.

Additionally, the computational efficiency of the 2.5D deep learning model in practical applications is significantly enhanced. Compared to full 3D models, the 2.5D model only needs to process a limited number of adjacent slices, greatly reducing computational load. This allows the 2.5D model to complete the training and inference process in a shorter time while maintaining high detection accuracy, thus improving the model’s feasibility for clinical applications [15]. For example, in this study, the 2.5D model reduced training time by approximately 30% compared to traditional 3D models, while improving detection accuracy by about 15%.

Another technical advantage is the flexibility and scalability of the 2.5D model. Based on a pre-trained 2D model, the 2.5D model can easily apply transfer learning techniques to utilize image feature learning results from other fields in DBT images. This approach can accelerate the training process and utilize a large amount of existing image data, further enhancing the model’s generalization capability and robustness [16]. In this study, the 2.5D model successfully applied features from other medical imaging data to breast cancer detection through transfer learning techniques, demonstrating excellent performance in different types of breast tissue and lesion types.

4.3 Integration of clinical and imaging data

Integrating clinical data with imaging data significantly improves diagnostic accuracy, providing a more comprehensive diagnostic perspective. This approach not only enhances the model's practicality but also supports the development of precise treatment strategies. Our study's results confirm this, with the integrated model achieving a diagnostic accuracy of up to 79.2%, supported by related research [17, 18].

4.4 Multi-model performance analysis and optimization strategies

This study evaluated various machine learning algorithms, highlighting the strengths of both deep learning and traditional models. Deep learning models like CNN and ResNet-50 demonstrated superior classification accuracy, with ResNet-50’s residual structure effectively addressing the vanishing gradient problem and enhancing feature learning. In contrast, traditional models such as XGBoost excelled in interpretability, offering valuable insights for clinical decision-making. Logistic Regression and SVM, while less effective, were limited by their inability to capture nonlinear relationships.

Model performance was optimized through integrated strategies, including feature extraction, data augmentation, and parameter tuning. Transfer learning, using pre-trained weights, reduced training time and improved generalizability, while techniques like random flipping and MixUp enhanced model robustness against edge cases.

While deep learning excels in accuracy, its computational demands may limit real-world applicability, especially in resource-constrained settings. Tree-based models like XGBoost, with lower resource requirements and strong interpretability, provide a practical alternative. future studies should balance accuracy, interpretability, and efficiency.

Advancing model performance will require adopting state-of-the-art architectures, such as Vision Transformers, and leveraging larger, more diverse datasets. Additionally, integrating domain knowledge in feature selection will enhance clinical relevance, ensuring model outputs effectively support patient management and treatment planning.

4.5 Clinical applications and potential improvements of model predictive performance

Our model demonstrates high sensitivity (87.2%) and moderate specificity (68.0%), highlighting its value in predicting the nature of breast lesions. However, these metrics present potential risks. For instance, the overtreatment rate for predicted malignant cases is approximately 16.33%, while the undertreatment rate is around 12.77%. These figures indicate a trade-off between avoiding unnecessary biopsies and minimizing missed diagnoses.

To mitigate these risks, model predictions should be integrated with clinical expertise and supplementary diagnostic tools. For high-risk patients, combining the model with multimodal imaging (e.g., MRI or ultrasound) or histological validation can reduce missed diagnoses. For low-risk patients, incorporating clinical assessments alongside model predictions can minimize unnecessary invasive procedures. Future studies should focus on optimizing feature selection methods, such as PCA or LASSO, and developing stratified screening strategies to improve diagnostic efficiency。

4.6 Study limitations

Despite including a relatively large patient sample, our study is still limited by sample size, especially in subgroups with high breast density and specific lesion types, potentially affecting the model's generalizability.

4.7 Technical challenges

The image preprocessing and segmentation techniques involved in this study require high-quality images and professional expertise, presenting certain technical difficulties. Additionally, the image quality of some patients may be affected by noise and artifacts, necessitating further optimization of these preprocessing steps.

Future research directions: (1) Expand Sample Size: Future research should include larger-scale multicenter datasets to improve the model's generalizability and reliability. Special attention should be given to balanced distribution across high breast density and different lesion types. (2) Optimize Model: Further optimize the model structure and algorithms, including incorporating more clinical features and using advanced deep learning techniques to enhance model performance. Research should also explore the feasibility and effectiveness of the model in real clinical applications, including assisting and improving the physician diagnostic process. (3) Technical Improvements: Improve image preprocessing and segmentation techniques to enhance image quality and segmentation accuracy, reducing the impact of noise and artifacts on model performance. Additionally, develop more convenient and automated processing workflows to lower technical barriers. By addressing these directions, future research can further enhance the capabilities and clinical applicability of AI-based breast cancer diagnostic models.