Ensemble Learning for Breast Cancer Detection in Mammography

Abnormal cell division in the mammary gland leads to the formation of either malignant or benign tumors, which is the cause of breast cancer (BC) [1]. Breast cancer is one of the most dangerous cancers worldwide. The World Health Organization reported 2.3 million new cases and 666,000 related deaths in 2022. These figures are projected to reach 3.3 million cases and 1.2 million deaths by 2050 [2]. Figure 1 illustrates a normal breast image and a suspicious breast from the INbreast database.

1.png

Figure 1. Normal breast image and a suspicious breast

Early diagnosis is critical for improving survival rates, but it remains complex and time-consuming [3]. Emerging technologies such as Deep Learning (DL) offer more efficient, accurate methods for diagnosing breast cancer, particularly when integrated into Computer-Aided Detection (CAD) systems, improving diagnostic accuracy and reducing false results [4, 5]. However, DL adoption is still limited as healthcare professionals require proof of its effectiveness [6, 7]. Deep learning and Deep Convolutional Neural Networks rely on analyzing patient data to identify patterns and make predictions, requiring large datasets and accurate parameters for better results [8]. The research also explores the potential of deep learning algorithms to enhance mammography-based breast cancer classification and detection, despite digital mammography being the most widely used method [9].

Convolutional Neural Networks (CNNs), inspired by the human visual cortex, have emerged as powerful tools for medical image analysis [10]. Many studies have applied CNNs to breast cancer diagnosis in mammography [11]. Ensemble strategies remain relatively underexplored, despite their potential to deliver more robust and reliable results by integrating complementary features from pre-trained architectures such as VGG16, ResNet50, and MobileNetV2. In this study, ensemble learning is combined with transfer learning and data augmentation, and the framework is evaluated on three benchmark datasets. This integrated approach enables a more comprehensive evaluation of CAD systems for mammographic classification, providing valuable insights into their clinical applicability. The ultimate objective is to develop a breast cancer classification model that can act as a reliable “second opinion” for radiologists, thereby reducing the risk of diagnostic errors.

This paper is structured as follows: Section 2 provides an overview of recent work in this field. In Section 3, the method for classifying breast anomalies into normal and suspicious is described. Section 4 presents and discusses the results of the experiments carried out. Finally, section 5 concludes and highlights directions for future work.

3.1 Datasets

In this study, we employ three open-access mammography datasets that are suitable for developing, training, and testing breast cancer models based on deep learning. The Mammographic Image Analysis Society (MIAS) dataset [19], although dated, remains ideal for small-scale studies and foundational research. INbreast [20] is appropriate for advanced deep learning techniques due to its high-resolution digital images, although it has not been updated since 2017. The Curated Breast Imaging Subset of DDSM (CBIS-DDSM) [21] excels in large-scale studies, offering a diverse and extensive collection of cases. Table 1 provides a detailed overview of the specific characteristics of each dataset used in this analysis.

Table 1. The specific characteristics of each datase

3.2 Data augmentation

Data augmentation techniques generate new samples of the training dataset through random transformations on the available data. This approach has multiple benefits, including accelerating the convergence process and preventing overfitting. For small datasets, the simplest approach is to perform basic transformations such as translation, zooming, flipping, mirroring, and rotation [22].

3.3 Transfer learning

Transfer learning enables the use of small datasets, such as medical images, by eliminating the need for costly training of deep models from scratch [23].

3.4 Ensemble learning

Ensemble learning is a robust strategy that integrates predictions from multiple models to achieve superior outcomes. By harnessing the strengths of various models, this approach can significantly improve classification accuracy compared to individual models [24].

3.5 CNN architecture

VGGNet [25] overcomes dataset limitations, allowing deep model training with flexible input sizes. As an extension of AlexNet. VGG16 include 13 convolutional layers and 5 pooling layers within five convolutional blocks. Each block employs 3 × 3 filters, followed by max-pooling to halve input size and double filters, optimizing with fewer parameters. Batch normalization accelerates convergence and minimizes error, preventing overfitting, while dropout reduces error rates further. The final block has three fully connected layers (4096, 4096, and 1000 units). For this study’s two-class output, the last layer (FC8) is modified to two units. Despite demanding significant computational resources, managing around 140 million parameters, this architecture has been quickly adopted and applied in healthcare for developing simple and effective predictive models [13, 26, 27]. Figure 2 illustrates the VGG16 architecture.

ResNet50, also known as a deep residual network, is a pre-trained model built on ImageNet, designed to address vanishing gradient issues by allowing the network to skip one or more layers. This structure comprises 50 layers arranged in residual blocks with shortcut connections, enabling layer skipping by adding the convolutional output to the block’s input tensor, thereby reducing computational complexity while supporting complex feature learning. The architecture begins with an initial convolutional layer, followed by batch normalization and two pooling layers, and includes 16 residual modules. These modules alternate between blocks with 4 convolutional layers and those with 3, each followed by batch normalization and ReLU activation for efficient connectivity. ResNet50 achieves a top-5 error rate of 7.8 on ImageNet [28]. This architecture is further validated by researchers and has demonstrated effectiveness in various predictive models. The results obtained reinforce its strong reputation in breast cancer classification tasks [13, 26, 27].

2.png

Figure 2. VGG16 architecture

MobileNetV2, an evolution of MobileNetV1 by Howard et al. [29], is designed for embedded systems, mobile devices, and resource-limited environments, minimizing computational and memory demands with minimal loss in accuracy. Its core innovation, the inverted residual with a linear bottleneck, processes low-dimensional inputs using depthwise convolution before compressing them, enhancing efficiency. This replaces standard convolution with depthwise convolution for lightweight filtering and pointwise (1 × 1) convolution for output merging, allowing MobileNetV2 to operate on individual channels, reducing complexity while preserving feature extraction.

MobileNetV2 contains 53 layers, fewer parameters, and an input size of 224 × 224. It leverages depthwise separable convolutions with two residual block types: stride 1 for regular processing and stride 2 for downsampling, striking a balance between performance and efficiency [30]. MobileNetV2 is widely applied in medical image analysis [14, 26]. Figure 3 illustrates the MobileNetV2 building blocks [30].

3.png

Figure 3. Structure of the building blocks of lightweight MobileNetV2 CNN model

4.png

Figure 4. Flowchart of the proposed ensemble pipeline

Our pipeline begins with preprocessing and data augmentation to increase dataset variability and reduce overfitting. Using transfer learning, we employed three convolutional neural networks (CNNs): VGG16, ResNet50 and MobileNetV2, each independently estimating the probability of an image belonging to the benign or malignant class. The choice of these networks was motivated by their complementary strengths. VGG16 provides a widely used baseline with a straightforward convolutional design; ResNet50 incorporates residual connections, enabling deeper and more generalizable feature extraction; and MobileNetV2 offers a lightweight architecture well suited for resource-constrained environments. To combine their outputs, a stacking strategy was applied in which the predictions from the three CNNs were used as input features for a meta-classifier. A Random Forest was chosen as the meta-learner due to its robustness and ability to capture nonlinear relationships. Its hyperparameters (number of estimators and tree depth) were optimized using RandomizedSearchCV with cross-validation, while the area under the ROC curve (AUC) served as the evaluation metric. Figure 4 presents an overview of the complete workflow.

3.6 Key metrics

Key metrics for model evaluation and comparison include accuracy, which measures the proportion of correct predictions (Eq. (1)). Precision, or positive predictive value, represents the percentage of predicted malignant cases that are correct (Eq. (2)). Sensitivity (Recall), or true positive rate, indicates the proportion of malignant cases detected (Eq. (3)). The F1 score, defined as the harmonic mean of Precision and Sensitivity, provides a balanced assessment (Eq. (4)). AUC, which represents the area under the ROC curve, serves as a global performance measure. Formal definitions are given below [31].

Accuracy $A c c=\frac{T P+T N}{T P+T N+F P+F N}$     (1)

Recall $S v=\frac{T P}{T P+F N}$     (2)

Specificity $S_p=\frac{T N}{T N+F P}$     (3)

Precision $P=\frac{T P}{T P+F P}$     (4)

F1 score $F 1=2 \times \frac{P \times S v}{P+S v}$     (5)

• TN: True negatives (correctly identified negatives)
• TP: True positives (correctly identified positives)
• FN: False negatives (positives incorrectly identified as negatives)
• FP: False positives (negatives incorrectly identified as positives)

This study investigated the potential of deep learning–based CAD systems for breast cancer diagnosis through an ensemble of VGG16, ResNet50, and MobileNetV2. The proposed approach achieved strong performance across multiple benchmark datasets, offering valuable support as a “second opinion” to radiologists. However, these results must be interpreted with caution. In particular, the small size and limited diversity of datasets such as INbreast restrict the ability to generalize findings to larger and more heterogeneous clinical populations. Moreover, while many studies, including ours, report improved results compared to the state of the art in controlled testing setups, this does not guarantee robustness in real-world clinical deployment. AI models remain highly sensitive to variations in patient cohorts, imaging protocols, and annotation quality, which often limit their applicability outside the training environment.

Future work should therefore prioritize validation on larger, multi-center, and demographically diverse datasets, along with systematic analyses of model robustness under varying acquisition and labeling conditions. Beyond reporting incremental performance gains, it is equally important to highlight scenarios where AI systems underperform, as such insights are crucial to building clinically reliable tools. Shifting the notion of “novel contribution” from ever-increasing model complexity toward a deeper understanding of AI limitations and failure modes will ultimately help bridge the gap between promising experimental results and safe, trustworthy deployment in medical practice.