Introduction

Breast cancer is the world leading cancer for females caused by the uncontrolled proliferation of cells in the breast and the accumulation of additional tissues known as a tumor. It is most frequently diagnosed cancer and the primary contributor to cancer-related deaths in the female population. The high incidence of breast cancer in Asia presents a substantial public health concern. The region experiences a significant prevalence of breast cancer cases, exhibiting varied patterns among its member countries. Asia exhibits the lowest age-standardized mortality rate (ASMR) and age-standardized incidence rate (ASIR) compared to other regions. However, the mortality-to-incidence ratio (M/I) in Asia, with a value of 0.32, surpasses the global average of 0.28, positioning it as the second-highest area worldwide in terms of M/IYip, Taib & Mohamed (2006). The breast cancer mortality rates in low- and middle-income countries are higher than in their high-income counterparts.

In Malaysia, a middle-income country, there were 21,634 cases of breast cancer discovered between 2012 and 2016, accounting for 34.1% of the cancers diagnosed in the country (Azizah et al., 2019). Given the potential severity of breast cancer, early detection and immediate implementation of appropriate treatment become essential in minimizing its impact. This emphasizes the critical significance of identifying the disease at an early stage and preventing its progression. Mammography screening is the most common and practical method of detecting breast cancer as it reduced breast cancer mortality by around 20% in women aged 50 to 70 (Christiansen, Autier & Støvring, 2022). The cornerstone of breast cancer diagnosis, mammography screening, plays a critical role in recognizing microcalcifications—tiny calcium deposits inside breast tissue. Although frequently an early indicator of breast abnormalities, these microcalcifications are not conspicuously apparent through symptoms. Mammography, on the other hand, greatly assists in the diagnosis of suspected early-stage breast cancers. With mammography lowering breast cancer mortality by roughly 20% in women aged 50 to 70, the focus on early diagnosis is directly related to identifying these microcalcifications.

By looking for microcalcification in the mammography images, breast cancer in its initial stages can be discovered. Microcalcification, a calcium deposit in the breast, appears as a scattering of white dots on a mammogram. They come in sizes ranging 0.1 to 1 mm (Cai et al., 2019). According toHakim, Prajitno & Soejoko (2021), microcalcification is an indication of breast cancer occurring in between 12.7% and 41.2% of women who completed mammography screening.Brahimetaj et al. (2022) estimated that between 85% and 95% of cases of ductal carcinomain situ (DCIS) were detected due to the existence of microcalcifications in roughly 55% of non-palpable breast cancers. Additionally, there is a strong correlation between a cluster of microcalcification as well as an increased probability of breast cancer (Azam et al., 2021). Invasive disease can be prevented by recognizing microcalcifications given that they are associated with premalignant and developing breast disease (Logullo et al., 2022). Therefore, the early intervention to detect breast cancer is identifying signs of microcalcification.

Yet it can be difficult and challenging to detect microcalcification because their distribution in the breast is unpredictable and their shapes are uncertain. Radiologists often missed or misdiagnosed microcalcification, therefore many healthcare professionals have had trouble spotting it. The microcalcifications can be observed as a cluster of several white dots and they can be identified in common benign lesions including fibrocystic alterations, inflammatory lesions, and breast abnormalities (Brahimetaj et al., 2022).

When considering a middle-income country such as Malaysia, where financial resources are limited, the incorporation of deep learning models into the healthcare infrastructure can offer a substantial enhancement to mammography screening. Through the effective detection of microcalcifications that might otherwise evade detection, these models make a valuable contribution to the timely detection of potential breast abnormalities. This is in line with the overarching goal of minimizing the detrimental effects of breast cancer through early detection. The potential for significant improvement in the diagnosis of breast cancer, especially during the critical early phases, could be realized through the synergistic collaboration of deep learning and mammography screening. This would enable the implementation of targeted and timely intervention strategies.

ResNet, a novel deep learning (DL) framework unveiled in 2015, addresses complexities inherent in deep networks through the utilization of skip connections within residual blocks. ResNet-50, an instance of the ResNet architecture, has garnered considerable attention in domains such as mammogram analysis. Studies byShiri Kahnouei et al. (2022) andLeong et al. (2022) demonstrated the effectiveness of ResNet-50 in mammogram-based calcification detection and segmentation, achieving impressive accuracies of 96.7% and 97.58%, respectively. This showcases ResNet-50’s potential for advancing medical diagnosis and intervention through precise image analysis.Montaha et al. (2021) have built upon VGG16, developing models like BreastNet18 for early-stage breast cancer identification. GoogLeNet, a 22-layer architecture created by Google researchers, achieved high accuracy rates of 93.3% on the ImageNet dataset.Jhang (2018) explored the use of GoogLeNet with class activation mapping for detecting suspicious microcalcification regions in mammogram images.Sharma & Mukherjee (2020) also compared the classification performance of GoogLeNet and AlexNet in microcalcification detection.

The application of these deep learning models to the detection and segmentation of microcalcifications has been the primary focus of recent research. However, there remains a need for comprehensive comparative analyses to better understand the strengths and limitations of these models, as well as their practical implementation in clinical settings. In this context, ensemble learning emerges as a promising approach to enhance the performance of deep learning model in breast cancer detection. Ensemble learning is a method that combines multiple base models to produce a more powerful model. There are several advantages in utilizing ensemble learning. First of all, ensemble models combine outcomes from multiple base models and produce higher performance outcomes. This is achieved by the ensemble methods in leveraging the strength of different algorithms resulting in a better overall outcome performance compared to individual models. In addition, ensemble models help in reducing the risk of overfitting by capturing irrelevant patterns from training data. Ensemble methods can reduce the model biases and variance and improved the generalization performance of the ensemble models (Mohammed & Kora, 2023). In our study, we aim to improve the detection of microcalcifications in mammogram images by developing an optimized ensemble model. Specifically, we utilize an averaging technique to combine two baseline deep learning models and form an optimized ensemble model, thereby enhancing the reliability and effectiveness of breast cancer detection.

Materials and Methods

Our study proposes the implementation of an ensemble model that integrates the advantageous features of individual pre-trained convolutional neural network models to efficiently tackle the issue of detecting microcalcification. In order to address the existing research gap, this study undertakes an exhaustive comparative analysis to ascertain the merits, drawbacks, and particular suitability of each model in detecting these nuanced yet critical indicators of breast abnormalities in their early stages. Furthermore, the objective of the project is to investigate the viability and practical execution of incorporating these models into clinical environments to improve the precision of breast cancer detection and their practical implications in the healthcare sector.

Data preprocessing

Preparing the input dataset for the DL algorithms was a critical step in the pipeline. It involved a series of preprocessing steps to ensure the mammogram images are in optimal condition for further analysis. These steps encompassed various techniques and transformations aimed at enhancing the quality and suitability of the data. The process of preprocessing mammogram images played a vital role in facilitating accurate and effective analysis using DL algorithms.

• (a)

  Artifact removal

• InFig. 1, the dataset of normal images from mini-MIAS and INBreast underwent two crucial processes: the Otsu segmentation method and MorphologicalEx method. The Otsu segmentation method played a vital role in automatically identifying the optimal threshold value for segmenting the mammogram images. This segmentation process helped in separating the desired regions from the background noise. Subsequently, the MorphologicalEx method, also known as Morphological Erosion, was employed to effectively eliminate the labeled regions from the images. This meticulous step ensured that the presence of free labeling does not interfere with the subsequent image cropping process. By implementing these techniques, the dataset was prepared for further preprocessing steps.

• (b)

  Image cropping and resizing

• After the removal of artifacts, the images underwent cropping and resizing. This crucial step ensured that the images were appropriately sized at 224 × 224, aligning them with the requirements of the DL algorithms for efficient classification. By standardizing the image size, the DL algorithms could effectively perform their classification operations with optimal accuracy and efficiency.

• (c)

  Image filtering

• Then, an adaptive median filter was employed to enhance the quality of the mini-MIAS and INBreast cropped images, as well as the ROI images containing microcalcification from CBIS-DDSM. This filter effectively mitigated noise, resulting in improved mammography image quality. The application of the adaptive median filter demonstrated its efficacy in reducing noise and enhancing image clarity.

Classification model development using deep learning approach

A deep convolutional neural network (D-CNN) model is developed by finely tuning its hyperparameters to accurately detect microcalcifications in mammogram images. Instead of training the CNN model from scratch, transfer learning is applied, utilizing pre-trained CNN models such as ResNet-50, GoogLeNet, VGG16 and AlexNet architecture from the torchvision library. To achieve optimal results, critical hyperparameters such as the number of epochs, learning rate, batch size, and step size are carefully adjusted. The input images underwent normalization to the standard normalization parameters specified by ImageNet ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]). This normalization was performed because the transfer learning model employed was pre-trained using the ImageNet database. Normalization is necessary to restore the original color of the input cell regions that have been stained.

A comprehensive summary of each deep learning architecture can be found inTable 2. Upon completion of the training process, the model’s performance is evaluated using performance metrics (accuracy, precision, recall, F1-score, area under curve from true positive ratevs false positive rate graph) and confusion matrix as summarized inTable 3. The confusion matrix considers true positives (accurate microcalcification detection), true negatives (correct classification of negative cases), false positives (incorrectly labeled as positive), and false negatives (missed microcalcification detection). Analyzing these metrics allows for a robust evaluation of the model’s overall accuracy.

Table 2:

Overview of deep learning models in this study.

Model	Descriptions
ResNet-50	ResNet-50 is a variant of the ResNet model, which has 48 convolution layers along with 1 MaxPool and 1 Average Pool layer. ResNet-50 is an artificial neural network (ANN) that forms networks by stacking residual blocks. Identity connections take the input directly to the end of each residual block and learn the features to be classified as the desired output.
GoogLeNet	The GoogLeNet network is an impressive deep convolutional neural network comprising 22 layers. The GoogLeNet architecture is designed to handle input images with a resolution of 224 × 224 pixels. It was specifically engineered to be a computational powerhouse, surpassing or competing with existing networks at the time of its creation.
VGG16	VGG-16 is a convolutional neural network architecture with convolutional layers of 3 × 3 filters with stride 1 and constant padding and maxpool layers of 2 × 2 filters with stride 2 used. This arrangement of convolutional and max pool layers is followed uniformly throughout the entire architecture. The end has 2 F.C. (fully connected layers) followed by a softmax for output. The number 16 in VGG16 indicates that it has 16 layers with weights.
AlexNet	AlexNet is made up of five convolutional layers, three max-pooling layers, two normalization layers, two fully connected layers, and one softmax layer. Convolutional filters and a nonlinear activation function ReLU are used in each convolutional layer. The pooling layers are used for maximum pooling. Because of the presence of fully connected layers, the input size is fixed.

DOI:10.7717/peerj-cs.2082/table-2

Table 3:

Model evaluation metrics.

Metrics	Formula	Description
Accuracy	${\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$	A fraction of correctly classified cases over the total samples.
Precision	${\displaystyle \frac{TP}{TP+FP}}$	A proportion of positive classes were correctly classified.
Recall	${\displaystyle \frac{TP}{TP+FN}}$	A ratio of all positive samples correctly predicted as positive.
F1-score	$2\mathrm{x}{\displaystyle \frac{\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{x}\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}{\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}+\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}}$	A combination of precision and recall as their harmonic mean.

DOI:10.7717/peerj-cs.2082/table-3

The classification model optimization through ensemble model

The classification model was refined through the development of an ensemble architecture. Through a thorough evaluation, the two top-performing models distinguished by their high accuracy and minimal computational requirements, were selected. These models are then integrated into the ensemble architecture, employing a weighted averaging approach to combine their outputs. This process ensures a more resilient and accurate final classification outcome as illustrated inFig. 2.

The developed DL models of AlexNet, GoogLeNet, VGG16, and ResNet-50 were constructed and tested for their capability to detect microcalcifications in mammogram images. The process of individual testing was initiated by training and evaluating each model on their accuracy and proficiency in discerning these subtle yet vital characteristics. Specifically, the following models were subjected to comprehensive evaluation underwent rigorous testing. The AlexNet model was chosen due to its innovative contributions to image classification, whileVGG16, was compared due to its ability to be distinguished by its depth and resilient performance. GoogLeNet, which is renowned for its inception modules; and ResNet-50, which utilizes skip connections to ensure robust training. The Pseudocode of the optimization through development of ensemble model as shown inFig. 3.

Results

As shown inTable 4, the performance of each DL algorithm can be evaluated using the following performance metrics: accuracy, precision, recall, F1-score, AUC, and training time. These metrics offer valuable insights regarding the quality and effectiveness of the models utilized for microcalcification detection in mammogram images. The best results of each performance metric were made bold.

In overall, ResNet-50 was preferred in this study as it has high detection accuracy in detecting microcalcifications in mammogram images. The remarkable performance of ResNet-50 in detecting microcalcifications in mammogram images can be attributed to its distinctive architecture, specifically the integration of skip connections. By establishing connections in this way, the model can effectively learn to preserve important information from the input layers through to the deeper layers. This prevents the loss of crucial details during training, which can happen in traditional deep neural networks due to the vanishing gradient problem. ResNet-50 tackles this issue by ensuring that gradients can flow smoothly through the network, allowing it to identify subtle patterns, such as those indicating microcalcifications, more accurately.

The particular benefit of ResNet-50 in mitigating the vanishing gradient issue has significant importance in the context of microcalcification detection. In mammography pictures, the identification of these indications typically presents as subtle, necessitating the model’s ability to distinguish detailed patterns across several layers. The skip connections in ResNet-50 enable the preservation and transmission of important information, allowing for the effective collection and learning of small but significant traits associated with microcalcifications, even in the deeper levels of the network. The preservation of complex features is crucial for achieving precise detection, resulting in improved precision for recognizing small but diagnostically important abnormalities seen in mammography images.

The architectural superiority of ResNet-50 lies in its ability to preserve the accuracy of characteristics that are crucial for detecting microcalcifications across its network layers. This attribute plays a pivotal role in the exceptional performance of ResNet-50 in this task. Additional investigation aimed at maximizing the efficiency of the training period while maintaining the remarkable performance of this method has the potential to significantly improve its practical applicability in clinical environments. This would facilitate the detection process by reducing the time required, while ensuring that accuracy is not compromised.

The presentation of the confusion matrix for the DL models inFig. 4 produced significant insights regarding their performance. A comparative analysis was undertaken to assess and contrast the performance of every deep learning algorithm with the objective of determining which model is superior to the others. The confusion matrix was of the utmost importance in providing a thorough comprehension of the performance of the models. The information it furnished was crucial and can be utilized to assess the precision and efficacy of the models in detecting microcalcifications.

Figure 4 illustrates that GoogLeNet exhibited the highest percentage value of TP, suggesting that it is the most effective model for accurately identifying microcalcifications in mammogram images. AlexNet and ResNet-50 both had comparable percentage values of TP; however, ResNet-50 exhibited a smaller FP value in comparison to AlexNet. This suggested that AlexNet exhibited a propensity to incorrectly classify non-microcalcifications as microcalcifications on a more frequent basis. In addition, ResNet-50 acquired a greater value of TN and a lesser value of FN in comparison to VGG16. The findings of this study demonstrated that ResNet-50 exhibited a high degree of accuracy in identifying mammogram images devoid of microcalcifications, with no instances of misclassification involving the microcalcifications themselves. Overall, ResNet-50 performed the best in the confusion matrix, as its proportion of TN was the highest and its proportion of FN was the lowest.

Discussion

Conclusions

Our research represents a significant progression in the accurate detection and classification of microcalcifications in mammogram imagesvia the construction and implementation of the proposed ensemble model comprising optimized GoogLeNet and ResNet-50. The model’s precision and dependability were enhanced due to their outstanding individual performances and distinctive architectural details, which led to this decision. The integration of these models into an ensemble not only combined a wide range of capabilities but also facilitated increased resilience and a more comprehensive comprehension of the complex attributes that are crucial for prompt intervention in breast cancer. Moreover, our research is distinguished by its focus on average confidence scores, specifically in the classification of normal and microcalcification cases. The proposed model attained the average confidence scores of 0.9305 and 0.8859 in classifying microcalcification and normal cases respectively. This particular aspect enhances the comprehension of the model’s dependability, thereby making a substantial contribution to its precision and applicability in the field of clinical diagnosis. Our research is distinguished by the thorough integration of datasets and the in-depth analysis of average confidence scores within classes. This establishes a standard for future models in the critical field of breast cancer detection, which can be more refined and dependable. In comparison to microcalcification images from CBIS-DDSM, the datasets for normal images from mini-MIAS and INBreast are comparatively limited, which is the primary limitation of this study. The resultant distribution of the datasets for each class is imbalanced. By incorporating a broader range of datasets, one can obtain a more substantial and representative sample, which may enhance the performance and generalizability of deep learning algorithms, ultimately leading to more dependable and superior model performance. Further research could investigate automated approaches to cropping mammogram images in order to facilitate the development of an end-to-end classification system.

Supplemental Information