Class-Balanced Convolutional Neural Networks for Digital Mammography Image Classification in Breast Cancer Diagnosis

Abstract

This study introduces a class-balanced Convolutional Neural Network (CNN) framework specifically designed for the binary classification of breast tumors in digital mammography. The proposed method systematically addresses the pervasive issue of class imbalance in medical imaging datasets by implementing advanced dataset balancing strategies, which resulted in a significant reduction in false negatives that is critical in early breast cancer detection. The proposed architecture is designed for high-resolution mammograms and employs regularization techniques, such as dropout and L2 weight decay, which are intended to enhance generalization and reduce the risk of overfitting. Comprehensive data augmentation and normalization further enhance the model’s robustness and adaptability to real-world clinical variability. Evaluated on the MIAS dataset, our balanced CNN achieved an accuracy of 98.84%, exhibiting both sensitivity and overall reliability. This work demonstrates that a class-balanced CNN can deliver both high diagnostic accuracy and computational efficiency, indicating potential for future use in clinical screening workflows. The system’s ability to minimize diagnostic errors and support radiologists with reliable, data-driven predictions represents an exploratory step toward improving automated breast cancer detection.

1. Introduction

Breast cancer remains one of the leading causes of mortality among women worldwide, accounting for a significant proportion of new cancer cases each year. Early detection is critical for improving survival rates, as timely intervention can dramatically reduce disease progression and treatment complexity. Traditional diagnostic approaches, such as manual interpretation of mammograms and ultrasound images, while effective, are often limited by human variability, fatigue, and the inherent complexity of medical imaging. These challenges can lead to misdiagnosis or delayed detection, underscoring the urgent need for reliable, automated solutions in clinical practice.

Recent advances in artificial intelligence (AI) and deep learning have revolutionized medical image analysis, offering unprecedented accuracy and efficiency in disease detection [1,2,3,4,5,6,7]. Convolutional Neural Networks (CNNs), in particular, have demonstrated exceptional capability in extracting hierarchical features from imaging data, enabling robust classification of benign and malignant lesions without manual feature engineering [8,9,10,11,12,13]. Beyond CNNs, hybrid architectures and ensembles have further enhanced performance, addressing limitations in interpretability and generalization. Studies using high-resolution mammograms, ultrasound images, and histopathology slides consistently report state-of-the-art results, highlighting the transformative potential of AI-driven diagnostic systems.

Recent advances in hybrid deep learning architectures demonstrate significant benefits from cross-layer feature interaction. Notably, the Feature Cross-layer Interaction Hybrid Method for Remote Sensing Scene Classification (FCIHMRT) [14] has demonstrated significant improvements in image classification tasks by enabling effective cross-layer feature fusion. While FCIHMRT was originally developed for remote sensing applications, its underlying principles, such as hybrid feature extraction and interaction, are highly relevant to medical image analysis. Integrating such hybrid strategies in medical imaging [15,16] can further enhance the discriminative power of convolutional neural networks, suggesting promising directions for future research in mammography image classification.

Although numerous deep-learning approaches have been applied to mammography, many challenges remain unresolved, including class imbalance, intra-class variability, and the critical need to minimize false negatives. This study focuses on the systematic integration and fine-tuned optimization of established deep-learning components in a way that directly addresses these practical challenges. By engineering a balanced and robust training process specifically for high-resolution mammography, the present work aims to demonstrate that carefully optimized combinations of class balancing, regularization, and augmentation can yield clinically meaningful improvements, particularly in reducing false negatives.

This research builds upon these advancements by developing and evaluating a CNN-based framework for automated classification of digital mammograms. The proposed model aims to deliver high diagnostic accuracy while maintaining computational efficiency, showing promise for future research, though clinical deployment will require extensive external validation across heterogeneous datasets and imaging environments. By reducing diagnostic errors and supporting radiologists with consistent and data-driven predictions, such systems can significantly improve breast cancer detection, streamline clinical workflows, and ultimately contribute to better patient outcomes.

This paper presents an innovative approach to automated breast cancer detection using deep learning methods designed for digital mammography. The study introduces a streamlined, highly effective CNN framework optimized for classifying mammogram images. By minimizing false negatives and improving diagnostic reliability, the proposed system aims to enhance early detection of breast cancer. Additionally, the automated and scalable tool supports clinical decision-making, reduces radiologist workload, and integrates smoothly into real-world medical environments. The key contributions of this work can be outlined as follows:

• Development of a robust CNN architecture specifically designed for high-resolution mammographic images, ensuring efficient feature extraction and superior classification performance;
• Integration of advanced regularization and optimization strategies (including dropout, L2 weight decay, and early stopping) to enhance generalization and reduce overfitting, addressing common limitations in medical imaging datasets;
• Comprehensive evaluation on real-world clinical data, demonstrating the model’s ability to achieve high accuracy and reliability in distinguishing benign from malignant cases;
• Focus on practical applicability, ensuring the proposed solution is computationally efficient and suitable for deployment in clinical screening workflows to support radiologists and reduce diagnostic errors.

This paper is organized as follows. Section 2 presents the related work and background on deep learning and medical image analysis. Section 3 describes the proposed methodology, including the architecture and training of the Convolutional Neural Network. Section 4 details the experimental setup, evaluation, and the results obtained. The model’s performance is analyzed in terms of accuracy, precision, F1-score, and recall. Section 5 discusses the findings and their implications. Finally, Section 6 concludes the paper and outlines directions for future research.

2. Related Literature

Recent advances in artificial intelligence and deep learning have significantly transformed the field of medical image analysis, particularly in breast cancer diagnosis. This section reviews key studies and methodologies that have contributed to the development of automated diagnostic systems using neural networks and related techniques.

Utomo et al. [17] investigate the use of artificial neural networks (ANNs) with extreme learning machine (ELM) techniques for breast cancer diagnosis, addressing key limitations of traditional backpropagation (BP) ANNs such as slow training, parameter tuning, and susceptibility to local minima. Using the Breast Cancer Wisconsin Dataset, the authors compare ELM ANN with conventional BP ANN, evaluating performance through sensitivity, specificity, and accuracy. Their results demonstrate that ELM ANN consistently outperforms BP ANN in terms of sensitivity and accuracy, while BP ANN shows slightly higher specificity in some cases. The study highlights the superior generalization and computational efficiency of ELM ANN, suggesting its promise for intelligent medical decision support systems. However, the authors note that further work is needed to enhance interpretability and user accessibility, such as integrating ELM with decision trees and developing user-friendly interfaces for clinical deployment.

Bhardwaj and Tiwari [18] explore the application of genetically optimized neural networks for breast cancer diagnosis, presenting a methodology that integrates genetic algorithms with neural network architectures to enhance classification performance. The study provides a detailed account of the optimization process, including the encoding of network parameters and the evolutionary strategies employed to select optimal configurations. Notably, the authors illustrate the comparative advantages of their approach through schematic diagrams and ROC curve analyses, demonstrating improved accuracy and robustness over conventional methods. However, while the genetically optimized neural network shows promise in terms of diagnostic precision, the article also highlights certain limitations, such as the computational complexity associated with genetic optimization and the need for extensive validation on diverse datasets. The work contributes to the ongoing discourse on intelligent diagnostic systems by emphasizing the potential of evolutionary computation to refine neural network models for medical applications, yet it underscores the importance of balancing model sophistication with practical considerations in clinical deployment.

Masud et al. [19] investigate the application of CNN architectures for breast cancer diagnosis, focusing on the ability of these models to automatically extract and learn discriminative features from medical images. The study provides a technical overview of CNN-based methodologies, highlighting their strengths in handling complex image data and their adaptability to various diagnostic tasks. Through schematic representations and performance metrics, the authors demonstrate the effectiveness of CNNs in improving diagnostic accuracy compared to traditional approaches. However, the article also addresses challenges such as the need for large annotated datasets, computational demands, and the interpretability of deep learning models in clinical settings. While the results underscore the promise of CNNs for breast cancer detection, the authors emphasize the importance of further research to enhance model transparency, generalization, and integration into routine medical practice.

Sureshkumar et al. [20] introduces a hybrid computer-aided diagnosis (CAD) system that combines CNNs with a pruned ensembled extreme learning machine (HCPELM) for breast cancer detection and analysis. The authors address key challenges in mammogram interpretation, such as image enhancement, segmentation, feature extraction, and classification, by integrating advanced pre-processing steps, including artifact and pectoral muscle removal. Their hybrid approach leverages convolutional layers for spatial feature extraction and fully connected layers for classification, while the HCPELM component enhances model efficiency and generalization through pruning and ensembling. The system is evaluated on the MIAS database and demonstrates competitive performance, achieving a breast image recognition accuracy of 86%, which is comparable to or better than several benchmark deep learning models, including Faster R-CNN, Mask R-CNN, and hierarchical CNN architectures.

Meng et al. [21] propose an interpretable breast ultrasound image classification algorithm that integrates CNNs with Transformer architectures to improve both accuracy and transparency in computer-aided diagnosis. By using tumor boundary masks as prior knowledge, the model guides attention to relevant regions, combining local feature extraction from CNNs with global modeling from a multi-scale Swin Transformer. The ensemble approach and a gradient-free interpretation method (Score-CAM) further enhance performance and explainability. Tested on the BUSI dataset, the model achieves state-of-the-art results, with an accuracy and F1-score of 0.987, outperforming standalone CNN and Transformer models. While the method demonstrates clear advantages, its reliance on manual annotation and architectural complexity may limit scalability in clinical practice. Overall, this work underscores the value of combining domain knowledge, advanced deep learning, and interpretability for breast cancer diagnosis in medical imaging.

Dequit and Nafa [22] present a comprehensive study on the application of CNNs for the early detection of Invasive Ductal Carcinoma (IDC) in breast cancer, leveraging a large dataset of high-resolution microscopic images. Their approach stands out for its integration of a user-friendly graphical interface, designed to facilitate the adoption of deep learning tools by healthcare professionals. The authors meticulously curated a dataset comprising over 277,000 image patches, enabling robust model training and evaluation. The proposed CNN model achieved a high accuracy rate (98.24%) in distinguishing between benign and malignant tissue, underscoring the potential of deep learning for clinical decision support. Notably, the study critically compares CNN-based methods with traditional machine learning approaches such as XGBoost, highlighting the superior ability of CNNs to automatically extract hierarchical features from complex image data without the need for manual feature engineering. The work demonstrates that the synergy of advanced image processing and deep learning can significantly enhance diagnostic accuracy, while also emphasizing the ongoing need for improved interpretability, dataset diversity, and automation in annotation to further advance the field.

In addition to the aforementioned approaches, several recent studies have explored breast-cancer classification using diverse methodological directions that inform the present work. Rouhi and Jafari [23] developed hybrid segmentation–classification processes that integrate region-based and contour-based segmentation with ANN-based decision models. Dong et al. [24] introduced a Dual-Contourlet Transform framework that captures multi-resolution texture characteristics, subsequently classified using an improved KNN scheme. Chougrad et al. [25] examined transfer-learning strategies with deep CNN architectures such as VGG16, ResNet50, and InceptionV3, highlighting the benefit of fine-tuning for mammographic image analysis. Muduli et al. [26] proposed a lightweight CNN model designed to accommodate both mammographic and ultrasound modalities, focusing on architectural efficiency and automated feature learning. Finally, Jafari and Karami [27] investigated multi-CNN feature extraction combined with machine-learning classifiers, emphasizing the role of feature selection in improving diagnostic robustness. These representative studies provide a diverse methodological backdrop, and a comparative summary of their performance is later presented in the Section 4.5.

3. The Proposed Research Design

3.1. Medical Diagnosis with Deep Learning

In recent years, deep learning has emerged as a transformative approach for automated diagnosis in various domains, particularly in medical imaging and pattern recognition tasks [28,29,30,31]. Deep learning models, especially those based on neural networks, have demonstrated remarkable capabilities in learning complex representations directly from raw data, thereby reducing the need for manual feature engineering. This paradigm shift has enabled the development of diagnostic systems that can achieve expert-level performance in tasks such as image classification, object detection, and anomaly identification.

Deep learning offers the ability to analyze complex medical data with unprecedented accuracy. By using large datasets of medical images, electronic health records, and sensor data, deep learning models, particularly convolutional neural networks, can automatically learn to identify patterns and features that may be difficult for human experts to discern. This capability has led to significant advancements in areas such as disease detection, classification, and prognosis, often matching or even surpassing human-level performance in specific diagnostic tasks.

Despite these achievements, the integration of deep learning into clinical practice presents several challenges. The need for large, high-quality labeled datasets, concerns about model interpretability, and ensuring consistent performance across diverse patient populations remain active areas of research. Nevertheless, ongoing developments in explainable AI and hybrid modeling approaches are helping to address these issues, paving the way for deep learning to become an integral part of future diagnostic workflows and ultimately improving patient outcomes.

In the context of this study, deep learning techniques are employed to analyze digital mammography images and provide accurate diagnostic predictions. Specifically, CNNs are employed to automatically classify mammographic images as benign or malignant, leveraging the rich visual information present in medical imaging data. The use of real mammography datasets, such as the MIAS, Mammographic Image Analysis Society database, enables the model to learn discriminative features directly from the images, without the need for manual annotation of tumor shapes or sizes.

The diagnostic process involves several key stages: image preprocessing, data augmentation, and class balancing to address dataset imbalances, followed by the design, training, and evaluation of the CNN model. Preprocessing steps include resizing and normalization of images, while augmentation techniques such as random rotations and flips increase the diversity of the training data. The model is trained to distinguish between benign and malignant tumors, with performance evaluated using metrics like accuracy, recall, and F1-score. By automating feature extraction and classification, deep learning offers a scalable and reliable tool to support clinical decision-making and enhance the accuracy of breast cancer diagnosis from mammographic images.

3.2. The Architecture of Convolutional Neural Networks

CNNs are a class of deep learning models that have demonstrated exceptional performance in image classification tasks, including medical image analysis [32,33,34,35]. In this work, mammographic images are preprocessed as single-channel grayscale matrices of size 128 × 128, normalized to ensure consistent intensity distributions across the dataset. The CNN architecture (Figure 1) exploits local receptive fields and weight sharing, enabling the network to learn hierarchical feature representations directly from pixel data.

The initial layers of the network consist of convolutional operations, where a set of learnable filters is applied to the input image. Each filter performs a weighted sum over a local region, producing a feature map that highlights the presence of specific visual patterns. The convolution operation for a filter $j$ in layer ℓ is defined as

$$z_{j,i}^l=\sum _{u=1}^K \sum _{v=1}^K W_j^l\left(u,v\right)x_{i+u,i+v}^{\left(l-1\right)}+b_j^l$$

(1)

where $W_j^l\left(u,v\right)$ is the filter weight at position $\left(u,v\right)$ for filter $j$ in layer ℓ, $x_{i+u,i+v}^{\left(l-1\right)}$ is the input from the previous layer at position $i+u,i+v$, $b_j^l$ is the bias term for filter $j$ in layer ℓ and $K$ is the Kernel size. Following convolution, a non-linear activation function is applied element-wise. The Rectified Linear Unit (ReLU) is commonly used:

$$x_{j,i}^l=RELU\left(z_{j,i}^l\right)=max\left(0,z_{j,i}^l\right)$$

(2)

where $x_{j,i}^l$ is the activated output at position $i$ for filter $j$ in layer ℓ.

This non-linearity enables the network to model complex relationships and helps prevent the vanishing gradient problem, facilitating the training of deeper architectures. To reduce the spatial dimensions of the feature maps and introduce invariance to small translations, max-pooling layers are interleaved with convolutional layers. Max-pooling selects the maximum value within a local window, typically of size 2 × 2, effectively downsampling the feature maps and reducing computational complexity.

After several convolutional and pooling stages, the resulting feature maps are flattened into a one-dimensional vector, which is then passed to one or more fully connected (dense) layers. These layers learn non-linear combinations of the extracted features and perform the final classification. For binary classification tasks, such as distinguishing between benign and malignant mammographic findings, the output layer consists of a single neuron with a sigmoid activation function, yielding a probability $p\in \left[\mathrm{0,1}\right]$ for the positive class. The binary cross-entropy loss function is used to optimize the network parameters:

$$L_{BCE}=-\left[y logp+\left(1-y\right) log\left(1-p\right)\right]$$

(3)

where $y$ is the true label (0 for benign, 1 for malignant) and $p$ is the predicted probability for the positive class. The Adam optimizer is employed for its adaptive learning rate and robust convergence properties.

To prevent overfitting, regularization techniques such as dropout and L2 weight decay are incorporated. Dropout randomly sets a fraction of activations to zero during training, while L2 regularization penalizes large weights, promoting smoother decision boundaries. Early stopping is also utilized, halting training when the validation loss ceases to improve, thereby preserving the best-performing model.

Data augmentation strategies, including random rotations and flips, are applied to increase the diversity of the training set and enhance generalization. Class balancing methods, such as undersampling the majority class or applying class weights, are implemented to address class imbalance, which is common in medical datasets.

In summary, the proposed CNN architecture processes normalized single-channel mammograms through successive convolutional, activation, and pooling layers, followed by dense layers and a sigmoid output for binary classification. The network is trained using binary cross-entropy loss and the Adam optimizer, with regularization and augmentation techniques to ensure robust and generalizable performance. This approach leverages the strengths of CNNs in capturing both low-level and high-level visual features, making it well-suited for the challenging task of mammographic image classification.

3.3. The Classification Problem in Mammography

Mammographic image classification presents unique challenges in medical image analysis. While digital mammography is widely adopted for breast cancer screening due to its ability to capture detailed images, its diagnostic effectiveness can be compromised in patients with dense breast tissue. Other imaging methods, such as ultrasound and MRI, are often employed to complement mammography. Ultrasound assists in differentiating between types of lesions and guiding biopsies, whereas MRI is reserved for cases involving elevated risk or inconclusive findings. However, these imaging techniques are not without limitations, since their diagnostic accuracy can be affected by subjective interpretation and inherent constraints in sensitivity and specificity, sometimes resulting in incorrect or missed diagnoses [36,37,38]. These ongoing challenges underscore the necessity for more advanced and consistent diagnostic solutions [39].

Unlike natural images, mammograms often exhibit subtle differences between benign and malignant findings, requiring models to detect fine-grained patterns. Furthermore, mammography datasets are typically imbalanced, with significantly fewer malignant cases compared to benign ones. This imbalance can bias models toward the majority class, increasing the risk of false negatives that is a critical concern in clinical practice, as missed cancer diagnoses can have severe consequences. To address these challenges, our approach incorporates comprehensive data preprocessing, augmentation, and class balancing techniques. By enhancing the diversity of the training data and ensuring balanced class representation, the model is better equipped to learn discriminative features and generalize to unseen cases. The binary classification task is thus framed to maximize sensitivity (recall), while maintaining high specificity and overall accuracy, supporting reliable and early breast cancer detection.

4. Experimental Evaluation and Results

The experimental evaluation was designed to validate the effectiveness of the proposed CNN framework for automated classification of mammographic images. This stage focuses on assessing the model’s ability to accurately distinguish between benign and malignant cases under realistic conditions. To ensure a rigorous analysis, the experiments incorporated standardized preprocessing, class balancing, and augmentation techniques, aiming to address common challenges such as dataset imbalance and variability in image quality. The evaluation not only measures overall accuracy but also emphasizes clinically significant metrics like sensitivity and precision, which are critical for reducing missed diagnoses and false alarms in breast cancer screening. The following subsections describe the dataset, experimental setup, performance metrics, and results obtained from the conducted trials.

4.1. Dataset and Preprocessing

The experimental analysis utilized the Mammographic Image Analysis Society (MIAS) [40] dataset, an open-access collection of 329 grayscale mammograms annotated as benign or malignant. It is widely utilized in research on automated breast cancer detection because it includes both normal and abnormal cases with clearly defined labels, supporting the development of algorithms aimed at early diagnosis. Following dataset inspection, as summarized in Figure 2, it was observed that several samples were not labeled as benign or malignant. This is due to the presence of patients or individual breast images without detectable abnormalities. To ensure reliable model development, these unlabeled or non-tumor images were excluded. The resulting updated dataset, illustrated in Figure 3, comprised 322 images used for training and evaluation.

To ensure consistency and computational efficiency, all images were resized to a fixed resolution and normalized prior to model training. These preprocessing steps standardize pixel intensity distributions, reducing variability across samples. Additionally, data augmentation techniques, including random rotations and horizontal flips, were applied to increase dataset diversity and reduce overfitting.

Figure 4 presents a representative grayscale mammography image, providing visual context for the type of patterns the proposed model is designed to recognize, which is particularly important for collaboration with medical professionals.

The MIAS dataset’s detailed annotations and inclusion of both normal and abnormal cases make it highly suitable for developing and validating automated breast cancer detection algorithms. Additional analyses were performed to summarize tissue type distribution, tumor size, and abnormality categories, providing context for the model’s learning environment.

4.2. Data Analysis

A thorough analysis of the MIAS dataset was conducted to provide both model developers and end users with a clear understanding of the data characteristics and the conditions under which the model was trained. This analysis focused on three main aspects: tissue type distribution, tumor size, and abnormality categories.

First, the dataset’s tumors were categorized into three distinct tissue types, as illustrated in Figure 5:

• Fatty;
• Fatty–glandular;
• Dense–glandular.

Understanding the distribution of these tissue types is important, as tissue composition can influence the visual appearance of mammograms and potentially affect classification performance.

In addition to tissue type, the dataset includes information about tumor radius. Since the original mammographic images have a resolution of 1024 × 1024 pixels, tumor radii are reported in pixels for consistency (see Figure 6). Analysis of the data revealed no clear correlation between tumor size and tissue type, suggesting that the model must rely on a combination of visual features, such as texture heterogeneity and structural irregularities, rather than size alone, for accurate classification.

Finally, the dataset encompasses six categories of abnormalities, providing a diverse set of diagnostic challenges. These categories, depicted in Figure 7, are as follows:

• CALC: Calcifications.
• CIRC: Well-defined (circumscribed) masses.
• SPIC: Spiculated masses.
• MISC: Other ill-defined masses.
• ARCH: Architectural distortion.
• ASYM: Asymmetry.

By analyzing these characteristics, we ensure that the model is trained and evaluated on a dataset representative of real-world clinical scenarios, which is essential for developing robust and generalizable diagnostic tools.

4.3. Experimental Setup

The proposed CNN was implemented and trained using Python 3.12.12 with the TensorFlow/Keras deep learning framework. All experiments were conducted on a workstation equipped with an NVIDIA GPU to ensure efficient training and inference.

For compatibility with the CNN and to reduce computational cost, all mammographic images were resized from 1024 × 1024 to 128 × 128 pixels and normalized. Each image was assigned a binary label: 0 for benign and 1 for malignant. To increase dataset diversity and improve model robustness, data augmentation was applied, specifically through random rotations at varying angles. This approach is widely validated in medical imaging research, as it simulates real-world variability in mammogram acquisition and helps the model generalize better to unseen data. By artificially expanding the dataset in this manner, we reduce the risk of overfitting and ensure that the trained CNN can reliably recognize tumors across a broader spectrum of imaging conditions. This process expanded the dataset to 19,440 images, significantly enhancing the effective sample size.

Figure 8 presents sample images from both benign and malignant classes, offering visual examples of the data used for model training and highlighting the diversity of mammographic patterns encountered.

The model architecture described in Section 3.2. consists of three convolutional layers with ReLU activation functions, two max-pooling layers, dropout layers for regularization, and a final dense layer with a sigmoid activation for binary classification. To further enhance generalization and prevent overfitting, L2 regularization and early stopping were incorporated into the training process. Early stopping was triggered if the validation loss did not improve for three consecutive epochs (patience = 3), ensuring retention of the best-performing model.

Figure 9 illustrates a 3D schematic representation of the proposed CNN architecture, illustrating the sequence of convolutional, pooling, dropout, flatten, and dense layers. The visualization was generated using Python to provide an intuitive overview of the network structure and data flow. The layers are as follows:

• Conv2D: Extracts local spatial features from input images using learnable convolutional filters.
• MaxPooling2D: Reduces spatial dimensions by retaining dominant features, improving computational efficiency.
• Dropout: Randomly deactivates neurons during training to reduce overfitting and improve generalization.
• Flatten: Transforms multidimensional feature maps into a one-dimensional feature vector.
• Dense: Performs high-level feature learning and final classification based on extracted representations.

The training parameters are summarized as follows:

• Optimizer: Adam.
• Loss Function: Binary Cross-Entropy.
• Batch Size: 32.
• Epochs: 50 (with early stopping based on validation loss).
• Learning Rate: 0.001.

The feature extraction stage comprises three convolutional layers with ReLU activation functions. The first convolutional layer uses 32 filters of size 3 × 3, followed by a second convolutional layer with 64 filters of the same kernel size. A third convolutional layer with 64 filters is applied after the first max-pooling operation. All convolutional layers employ a 3 × 3 kernel to capture local spatial features while maintaining a manageable number of parameters. Feature map dimensionality is progressively reduced using two max-pooling layers with a pool size of 2 × 2.

To avoid overfitting, dropout regularization with a rate of 0.25 is applied after the convolutional block and again after the fully connected layer. The network includes a dense layer with 64 neurons and ReLU activation to learn high-level feature representations, followed by a final sigmoid-activated output neuron for binary classification.

All trainable layers are initialized using a random normal initializer with mean 0 and standard deviation 0.02. L2 regularization is applied to penalize large weights and improve generalization. The model is optimized using the Adam optimizer with a learning rate of 0.001 and trained using binary cross-entropy loss.

The final model comprises 117,505 trainable parameters, corresponding to a compact architecture suitable for small-scale medical imaging datasets while maintaining sufficient representational capacity.

The dataset was divided into input and output variables, as well as training and testing subsets. The resized images served as model inputs, while the corresponding binary labels constituted the outputs. The dataset was split using an 80/20 ratio, where 80% of the samples were used for training and 20% for testing. The test set remained unseen during training and was used to evaluate model performance on unseen data.

To prevent data leakage and ensure a fair evaluation, the dataset was first split into training and test sets. Data augmentation was applied exclusively to the training set and only after the train–test split, while the test set remained completely untouched and was used solely for performance evaluation. This protocol guarantees that no augmented samples derived from the training data were present in the test set, thereby preserving the validity and reliability of the reported results.

This setup was designed to achieve high diagnostic accuracy while maintaining computational efficiency, making the approach suitable for real-world screening workflows.

4.4. Evaluation Metrics

To measure the performance of the proposed CNN model, four widely recognized metrics were employed: accuracy, precision, recall, and the F1-score. Accuracy expresses the ratio of correctly classified samples to the total number of observations, serving as an overall indicator of performance. Precision measures how many of the instances predicted as positive are truly positive, which is particularly important when reducing false alarms is critical. Recall (also referred to as sensitivity) captures the proportion of actual positive cases that the model successfully identifies, making it essential in scenarios where missed detections must be minimized. The F1-score combines precision and recall into a single value by computing their harmonic mean, offering a balanced perspective when both error types matter. These measures offer a clearer perspective on model performance than accuracy alone, emphasizing the balance between false positives and missed detections.

Performance assessment is based on the confusion matrix and its derived components. The confusion matrix summarizes the relationship between actual labels and predicted classifications. It serves as a standard tool for evaluating model performance. An example of a confusion matrix for a binary classification task is shown in

Table 1. A typical confusion matrix.

	Predicted Positive	Predicted Negative
Actual positive	True Positive (TP)	False Negative (FN)
Actual negative	False Positive (FP)	True Negative (TN)

. True positives (TP) denote correctly recognized positive cases, while true negatives (TN) correspond to correctly identified negative cases. False positives (FP) occur when negative samples are incorrectly labeled as positive, and false negatives (FN) represent positive cases that the model fails to detect. The mathematical definitions of these measures are given in Equations (4)–(7).

$$Accuracy={\displaystyle \frac{TP+TN}{TP+FP+TN+FN}}$$

(4)

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(5)

$$Recall/Sensitivity={\displaystyle \frac{TP}{TP+FN}}$$

(6)

$$F1-score=2·{\displaystyle \frac{Sensitivity·Precision}{Precision+Sensitivity}}$$

(7)

These metrics collectively provide a deeper understanding of the trade-offs between false positives and false negatives, which is crucial for reliable breast cancer detection.

4.5. Results and Analysis

The proposed CNN model was evaluated on the MIAS dataset using the metrics described in Section 4.4. After training and validation, the model consistently demonstrated strong performance in distinguishing benign from malignant cases.

Given the limited size and inherent class imbalance of the original dataset, data augmentation was a critical component of our methodology. By applying random image rotations, we simulated real-world variability in mammogram acquisition, which is widely recognized as a valid approach in medical imaging research. This process expanded the dataset to 19,440 images, substantially enhancing data diversity and reducing the risk of overfitting. The augmented dataset was then split into training and testing subsets using an 80/20 ratio.

In the initial experiment, the dataset was used in its original, slightly imbalanced form. Using early stopping, training was terminated at the 19th epoch out of the initially scheduled 50 epochs, as validation loss increased for three consecutive iterations. The final selected model corresponded to epoch 16, achieving a test accuracy of 0.9396. The final model weights were restored from the epoch corresponding to the lowest validation loss, ensuring optimal generalization and preventing overfitting.

Figure 10 and Figure 11 illustrate the accuracy–epoch curves, which show effective feature learning and convergence during training. Both training and test accuracy increase steadily during the initial epochs, indicating effective feature learning and convergence. The close alignment between the two curves suggests good generalization and limited overfitting. Training was terminated using early stopping once no further improvement in validation performance was observed, ensuring optimal model selection while preventing unnecessary training.

Figure 12 presents the confusion matrix, providing a detailed assessment of the model’s performance in cancer classification. The high number of true positive predictions (2363) indicates the model’s strong ability to accurately detect malignant cases, which is essential for early diagnosis and effective clinical intervention. Furthermore, the limited number of false negative instances (152) suggests that the model infrequently fails to identify existing malignancies, thereby minimizing the risk of missed diagnoses and enhancing overall patient safety. However, the model showed slightly better accuracy in predicting benign tumors (class 0), indicating a relative weakness in detecting malignant cases (class 1).

Table 2. Evaluation metrics for the first model.

	Precision	Recall	F1-Score	Support
class 0	0.94	0.96	0.95	2466
class 1	0.92	0.89	0.91	1422
accuracy			0.93	3888
macro avg	0.93	0.93	0.93	3888
weighted avg	0.93	0.93	0.93	3888

summarizes the evaluation metrics for the first model. The macro average evaluates performance by giving equal importance to all classes, whereas the weighted average reflects overall performance by weighting each class according to its sample frequency. Support refers to the number of samples belonging to each class in the evaluation dataset. It indicates how many true instances of each class are used to compute the corresponding evaluation metrics.

In medical diagnosis, evaluation metrics must be interpreted in a clinical context. False negatives are particularly critical in breast cancer screening, as missed malignant cases may lead to delayed diagnosis, disease progression, and reduced treatment effectiveness. Therefore, minimizing false negatives is essential for patient safety and early intervention. Conversely, false positives, while less immediately dangerous, may result in unnecessary follow-up examinations, biopsies, increased healthcare costs, and psychological distress for patients. An effective diagnostic model must balance these two error types, aiming to reduce missed cancers while maintaining an acceptable false positive rate to support reliable and clinically meaningful screening decisions.

To further improve diagnostic reliability, a second experiment was conducted using a class-balanced dataset, obtained through random undersampling of the benign class. Specifically, benign samples were randomly removed without any clinical or visual selection criteria, so that an equal number of input instances was available for both benign and malignant tumor classes.

This strategy was adopted to isolate and evaluate the effect of class balance on model behavior, particularly with respect to false negative reduction, while avoiding the introduction of additional hyperparameters or task-dependent bias. Training was performed using early stopping; optimization terminated at epoch 24, with the model selected from epoch 21 achieving a test accuracy of 0.9884.

Figure 13 and Figure 14 show stable convergence in the accuracy–epoch curves of the second model, with training and test accuracy evolving in parallel. This behavior indicates improved generalization due to dataset balancing, while early stopping was applied to select the optimal training epoch and prevent overfitting.

The confusion matrix (Figure 15) demonstrates a significant reduction in false negatives (16) and false positives (45). Minimizing false negatives is particularly critical, as these represent cancer cases incorrectly classified as healthy, potentially delaying diagnosis and treatment.

Table 3. Evaluation metrics for the second model.

	Precision	Recall	F1-Score	Support
class 0	0.99	0.98	0.98	1925
class 1	0.98	0.99	0.98	1963
accuracy			0.98	3888
macro avg	0.98	0.98	0.98	3888
weighted avg	0.98	0.98	0.98	3888

presents the evaluation metrics for the balanced model. The results indicate a nearly balanced class distribution, with 1925 samples for class 0 and 1963 samples for class 1. This balance enables the more uniform contribution of both classes to the overall performance metrics:

• The high accuracy indicates that the model correctly classified the vast majority of samples;
• Recall (99.0%) is particularly significant in medical diagnostics, as it reflects the model’s ability to minimize false negatives—critical for early cancer detection;
• Precision (98.5%) shows that the model effectively reduces false positives, which helps avoid unnecessary follow-up procedures;
• The F1-score confirms a balanced trade-off between precision and recall, reinforcing the robustness of the approach.

The confusion matrix revealed very few misclassifications, with only a small number of benign cases incorrectly flagged as malignant and an even smaller number of missed malignant cases. This outcome demonstrates the effectiveness of the implemented class balancing and augmentation strategies.

To improve the robustness of the evaluation and reduce potential bias introduced by a single train–test split, the proposed models were evaluated using multiple random train–test splits. Specifically, the dataset was randomly partitioned into training and test sets following the same split ratio for 15 independent runs. For each run, the model was trained from scratch and evaluated on the corresponding test set. Performance metrics were then aggregated across all runs, and the final results are reported as mean ± standard deviation, providing a more reliable estimate of model stability and generalization. As shown in

Table 4. Performance comparison across 15 random train-test splits (mean ± standard deviation).

Models	Accuracy
Model 1 (without class balancing)	92.20 ± 0.77
Model 2 (with class balancing)	97.13 ± 0.83

, incorporating class balancing consistently improved both predictive accuracy and robustness across different data partitions.

For completeness,

Table 5. Comparative performance analysis for representative MIAS-based studies.

Study	Model Type	Accuracy
Rouhi & Jafari (2016) [23]	Hybrid segmentation + feature extraction + ANN/KNN/SVM	90.94%
Dong et al. (2017) [24]	Dual-Contourlet Transform (Dual-CT) + Improved KNN	94.14–95.76% (depending on feature & classifier)
Chougrad et al. (2018) [25]	Transfer-learning with deep CNNs (VGG16, ResNet50, InceptionV3)	98.23%
Muduli et al. (2022) [26]	Lightweight CNN (four-layer architecture)	96.55%
Jafari & Karami (2023) [27]	Multi-CNN feature extraction + ML classifiers (NN, kNN, RF, SVM)	94.5%
Aguerchi et al. (2025) [12]	Multi-modal unified CNN	98.97%
Proposed approach	Class-balanced CNN for digital mammography	98.84%

provides an overview of representative MIAS-based approaches and their reported performance, illustrating the diversity of methodologies employed in the literature. These prior works vary substantially in preprocessing processes, image resolution, class definitions, data balancing strategies, augmentation methods, and the train–test splits used. Because of these methodological inconsistencies, direct numerical comparison with the proposed model would not be reliable or fair. Nevertheless, presenting these studies together offers useful context and highlights how heterogeneous the MIAS evaluation landscape remains.

Despite the methodological differences across published MIAS-based studies, the proposed model achieves a consistently high accuracy of 98.84%, demonstrating strong and reliable diagnostic performance within a controlled and reproducible experimental setting. This effectiveness is driven by the combined use of data augmentation, regularization, and class balancing, which together help reduce overfitting and improve generalization on a small dataset. These design choices support a consistent classification process and highlight the potential of CNN-based systems as practical decision-support tools for radiologists, helping to reduce diagnostic variability and uncertainty in breast cancer screening.

5. Discussion

The results of this study demonstrate the effectiveness of class-balanced CNNs for digital mammography image classification in breast cancer diagnosis. The proposed framework achieved high accuracy, precision, and recall, confirming its potential as a reliable tool for supporting radiologists in clinical practice.

A key finding is the significant impact of dataset balancing on model performance. By addressing the inherent class imbalance in mammography datasets, the model was able to reduce false negatives, which is a critical factor in early breast cancer detection. This improvement is particularly important, as missed diagnoses can have severe consequences for patient outcomes. The use of data augmentation and regularization techniques further contributed to the model’s robustness, minimizing overfitting and enhancing generalization to unseen data.

The architecture’s simplicity and computational efficiency make it suitable for integration into real-world screening workflows. Unlike more complex hybrid or ensemble models, the streamlined CNN design facilitates faster training and inference, which is advantageous for large-scale deployment in clinical settings. Moreover, the automated feature extraction capabilities of CNNs eliminate the need for manual annotation or handcrafted features, reducing the workload for medical professionals and enabling scalable diagnostic support.

Despite these strengths, several limitations should be acknowledged. The study relied on the MIAS dataset, which, while widely used, may not fully represent the diversity of mammographic images encountered in different populations or imaging devices. Future work should include validation on larger and more heterogeneous datasets to assess the model’s generalizability, as well as direct benchmarking against other recent image classification methods. Additionally, while the model achieved high performance metrics, interpretability remains a challenge. Future work should incorporate explainability techniques, which can help radiologists interpret the model’s decisions.

In addition to these points, it is important to acknowledge that the MIAS dataset is relatively small and originates from a single imaging environment, which limits the diversity of mammographic appearances and patient characteristics represented in this study. Although data augmentation helped increase variability, synthetic transformations cannot fully replicate the complexity of real-world imaging conditions. Furthermore, the absence of external validation on independent datasets restricts our ability to assess broader generalizability. The use of random undersampling, while practical for balancing classes, may also lead to the exclusion of subtle benign cases that could be clinically relevant. These limitations should be considered when interpreting the results, and future work will aim to address them by incorporating larger, multi-center datasets and exploring more advanced imbalance-handling strategies.

Nevertheless, it is important to highlight that the contribution of this work lies in the integration of established deep learning techniques for digital mammography. By prioritizing practical applicability and transparent evaluation, we provide a foundation for reliable AI-assisted screening workflows. The study demonstrates that innovation in medical AI can arise from the thoughtful adaptation and optimization of existing methods to meet the stringent demands of real-world healthcare environments.

In comparison with related literature, the proposed approach matches or exceeds the diagnostic accuracy of more complex models, while maintaining computational efficiency and ease of deployment. The focus on class balancing and regularization addresses common pitfalls in medical image analysis, such as overfitting and bias toward the majority class.

In summary, this work highlights the promise of class-balanced CNNs for reliable and efficient breast cancer diagnosis from digital mammograms. Continued research should focus on improving interpretability, validating across diverse datasets, and exploring integration with other diagnostic modalities to further advance automated breast cancer screening.

6. Conclusions

This study presents a class-balanced CNN framework for the binary classification of breast tumors in digital mammography. Through a comprehensive pipeline that included data preprocessing, augmentation, strategic class balancing, and architectural optimization, the proposed CNN model achieved high diagnostic performance, with accuracy exceeding 98% on a balanced dataset. Key techniques such as dropout, L2 regularization, and early stopping contributed significantly to reducing overfitting and improving the model’s generalization capabilities.

A critical outcome of this research, enabled by the systematic application of class balancing and data augmentation, was the significant reduction in false negatives, an essential factor in medical diagnostics, where undetected malignancies may have severe clinical consequences. The results emphasize that the reliability of deep learning models in healthcare applications is not solely dependent on architecture, but heavily influenced by the quality, balance, and representativeness of the input data.

This work suggests that class-balanced CNN systems have the potential to support radiologists in the future; however, this will require additional studies using larger, multi-center datasets. While these models are not intended to replace human expertise, they can enhance diagnostic workflows, reduce fatigue-related errors, and facilitate early cancer detection, especially in challenging or ambiguous cases.

Future work may include experimentation with transfer learning, larger and more diverse datasets, and deployment in real clinical settings with integration into radiology workflows. This study lays the groundwork for more reliable, efficient, and scalable AI-driven solutions in breast cancer screening.