An optimized deep learning architecture for breast cancer diagnosis based on improved marine predators algorithm

Breast cancer is a familiar frequent malignancy in females and the second most common leading cause of death in women. The global occurrence of breast cancer has increased over time, and more cases have been reported every year. In comparison to other malignancies, it is more common in women. If this disease is not detected early, it might lead to death [1]. Early diagnosis improves the chances of successful therapy and survival, but its diagnosis is time-consuming and frequently results in an agreement between pathologists. Computer-aided diagnosis (CAD) systems can improve diagnostic accuracy. Breast cancer can be classified as benign (non-hazardous) or malignant (threatening). Early detection of breast lesions and the distinction of malignant from benign lesions is critical for breast cancer prognosis [2]. Breast cancer affects approximately a million women globally each year, accounting for more than 25% of all female cancer occurrences. There is a tremendous need for early-stage development of new breast cancer methods with this exponential expansion. This motivates researchers to develop innovative methods for obtaining rapid and accurate diagnoses, ultimately extending patients’ lives [1]. Reviewing preliminary diagnostic information and gathering relevant data from previous information is key to detecting this disease early and reliably. Medical imaging and deep learning (DL) techniques will aid this procedure. Medical imaging performs a significant part in clinical illness diagnosis, therapy evaluation, and the detection of anomalies in many body components, including eye [3], lungs [3], brain [4], breast [5], and stomach [6]. Medical image research aims to classify an organ’s location, size, and characteristics in question, which is regarded as a viable way to obtain usable information from vast amounts of data. The most effective process to identify breast cancer is through medical imaging, including mammography images, histopathology images, magnetic resonance imaging (MRI), ultrasound, and thermography images [7].

Thermography images, also known as thermal imaging, have shown enormous promise in the early detection of breast cancer over the previous decade. Thermography images can help with other types of diagnosis by providing information about physiological changes [8].

Ultrasound imaging is another method for breast cancer detection that young women mostly use it. Noise levels might cause the process to fail when it attempts to detect microcalcifications and deeper breast tissue [9].

In addition to ultrasound and thermography screening, MRI is another approach for detecting cancer cells early. MRI uses magnetic instead of X-rays to produce extremely accurate three-dimensional (3D) transverse imaging [7].

Another imaging modality is called mammography images. Mammography is a very well and generally used approach for breast cancer screening, and it is the only image type that has been shown to decrease breast cancer mortality significantly [10]. It is an x-ray test image regarded as a reliable and accurate method for detecting breast cancer. [11]. In this paper, we used mammography images for breast cancer classification.

CAD systems have been used to aid clinicians in interpreting medical imaging to improve disease diagnosis. The critical role of a CAD system is feature extraction. Traditional feature extraction techniques have disadvantages because they lack flexibility [10]. Recently, DL approaches have been presented for breast cancer diagnosis.

DL is a category of machine learning and artificial intelligence that focuses on a complicated structure of image features due to its capacity to learn autonomously. DL approaches use various recently developed models to improve feature extraction from data. These models have been used in various medical fields [7, 12]. DL is composed of multilayer neural networks (NNs) that use raw input images to generate a hierarchical feature structure. Stack autoencoders, deep-Boltzmann machines, and convolution NNs (CNNs) are examples of common DL algorithms [13].

CNNs have had significant success in biomedical imaging, such as mitosis cell detection from microscopic images, tumor detection, segmentation of neural laminae, skin disease, immune cell detection and classification, mass detection, COVID-19 prediction [14] and classification in mammograms [15]. CNNs are a popular technique for object detection and image classification that involves layer-wise automatic feature extraction. The preparation of CNNs for classification objectives depends on the knowledge of hyperparameter tuning. Hyperparameters for each layer are different. Few studies, such as [16], have recognized the importance of the hyperparameters in achieving high performance with CNN architectures and the need to consider them as an optimization problem. Because a CNN model’s performance is determined by these hyperparameters [17], they need to be fine-tuned to achieve great results. The selection of hyperparameter values is frequently based on a mix of human expertise, trial and error, or a grid search method [18]. Training can take more days because of the computationally expensive core of CNN designs. Since the number of combinations grows exponentially with the increase in hyperparameters, the grid search method is usually not suited for CNN models. Tuning hyperparameters is a time-consuming effort for researchers; the number of layers of CNNs is increasing daily to cope with vast and complex datasets. It is not suitable to optimize hyperparameters manually at a reasonable cost. To improve them, different researchers have adopted different techniques. Some of them accepted their outcome, whereas others did not [17].

Thus, automatic hyperparameter optimization for CNN models is critical [19]. Meta-heuristic algorithms have significantly influenced hyperparameter optimization in several fields. They have been developed to address different real-world problems and have gained enormous interest in classification problems. Because meta-heuristic algorithms have great performance and are straightforward to implement, researchers have widely proved their experience to handle many types of challenging optimization issues in engineering, communications, industry, and social sciences [20]. In addition, they have been employed in biological information [21], chemical information [22], feature selection [23], task scheduling in cloud computing [24], image segmentation [25], global optimization [26], and as well as cost-effective emission dispatch [27]. There are various meta-heuristic algorithms, of which the marine predators algorithm (MPA) [28]. MPA is one of the most recent algorithms, released in 2020 by Faramarzi et al., [28], showing more quality results than several classical and the latest counterparts on various mathematical and engineering benchmark problems. Faramarzi et al. applied twenty-nine test functions to evaluate the MPA performance, and it showed high performance in different optimization problems. MPA has many advantages, including requiring the fewest amount of adjustable parameters, being simple to implement with powerful search capability, and flexible in altering the basic MPA version [29].

However, all meta-heuristic algorithms should strike a balance between exploration and exploitation; other solutions can be stuck in optimal solutions or fail to converge [30, 31]. Premature convergence is a problem for the MPA. As it splits the optimization iterations into three parts, the first part is only for exploration. The second part serves as a means of transitioning from exploration to exploitation. The third part is specified for the exploitation phase; this could affect the search process by causing the population to get stuck in local optimal. Accordingly, like other metaheuristic algorithms, the MPA has been improved. The no-free-lunch theorem states that no single method can solve all types of optimization problems. As a result of hybridization, various features can be combined into a single algorithm to address multiple challenges.

Thus, hybridization of many techniques from various scientific domains is required. Hybridization merges the benefits of various algorithms to create a more powerful, high-performing version with a promise of higher accuracy and performance. Notwithstanding MPA’s success in search processes, it might be improved in various areas to demonstrate its usefulness on more complex optimization projects. To enhance the MPA, we hybrid it with the opposition-based learning (OBL) strategy [32] to produce solutions from probable regions to explore the search space more thoroughly. OBL is one of the most efficient approaches to improve meta-heuristic algorithms [33]. It is merged with meta-heuristic algorithms in multiple ways to improve explorative searchability. After that, the proposed Improved Marine Predators Algorithm (IMPA) is used to optimize the CNN’s hyperparameters. As a result, automatic hyperparameter optimization for CNN architectures is critical. This paper proposes an improved meta-heuristic algorithm called IMPA using the OBL strategy to optimize the hyperparameters of the CNN architecture used for breast cancer classification. The proposed method is IMPA-ResNet50, which depends on a pretrained CNN model named ResNet50 to diagnose breast cancer from two mammographic datasets, the mammographic image analysis society (MIAS) and curated breast imaging subset of DDSM (CBIS-DDSM) datasets, using transfer learning (TL). The main contributions of this paper are as follows:The remainder of this paper is structured as follows: Sect. 2 presents some literature reviews. Section 3 explains the MPA, the OBL strategy, CNNs, and TL. Section 4 presents the improved version of the MPA (the IMPA method). Section 5 introduces details of the proposed model. The experimental results and performance analysis including the limitation of the proposed model are explained in Sect. 6. Finally, Sect. 7 concludes this paper and presents future directions.

This section provides a summary of previous work on breast cancer diagnosis. In [34], the authors presented a review of some works on breast cancer classification, and from this review, they conclude that CNNs achieve higher accuracy than multilayer perceptron (MLP) NNs. The authors in [15] proposed a DL technique based on TL. They used three pretrained models, namely GoogLeNet, VGGNet, and ResNet, to classify malignant and benign cells. They evaluate their approach on cytology images to see how well it works. In addition, in [35], the authors presented a deep-CNN model that incorporated TL to avoid overfitting occurring when dealing with small datasets. They evaluated the presented model’s performance using four datasets: DDSM, INbreast, BCDR, and MIAS datasets. The DDSM dataset showed 97.35% accuracy and 0.98 area under the curve (AUC). The INbreast dataset yielded 95.5% accuracy and 0.97 AUC, whereas the BCDR database yielded 96.67% accuracy and 0.96 AUC.

The researchers in [36] presented an approach called BDR-CNN-GCN that combines a graph-convolutional network (GCN) with a CNN. A basic eight-layer CNN was used that was integrated with batch normalization and dropout layer. The final BDR-CNN-GCN model was developed by combining this model with a two-layer GCN. The performance of this model was tested on the MIAS dataset, achieving 96.10% accuracy. Furthermore, in [37], the researchers presented a CNN Inception-v3 model that was trained on 316 images, yielding 0.946 AUC, 0.88 sensitivity, and 0.87 specificity. In addition, in [38], the authors proposed a classification method using CNN and TL. The purpose of the paper was to assess how eight fine-tuned pretrained models performed. In [39], the authors proposed a hybrid classification model: Alexnet, Mobilenet, and Resnet50. They achieved 95.6% accuracy from these hybridized models. In [40], the researchers used four different CNN architectures, namely InceptionV3, VGG16, ResNet50, and VGG19, including 5000 images, for model training (benign: 2500 and malignant: 2500). In addition, the prediction models were tested by 1007 images (benign: 788 and malignant: 219).

In the same context, in [41], the researchers proposed a TL model for classifying histopathological breast cancer images. The authors used the ResNet-18 model as a backbone model with a block-wise fine-tuning method. The performance of the model was improved using data augmentation techniques and global contrast normalization. In addition, in [42], the authors proposed a classification method for thermal images, which combines thermal images of different views using a CNN model. The method achieved 97% accuracy and 0.99 AUC, with a 100% specificity and 83% sensitivity. In [2], the authors presented a DL framework (DenseNet) that extracts image features and feeds them into a fully connected (FC) layer to classify cancerous and benign cells. The efficiency of the technique was evaluated through hyperparameter adjustment. In [43], the authors proposed a deep classification algorithm for mammography imaging named CNNI-BCC (CNN improvement for breast cancer). The CNNI-BCC model classifies breast images into malignant, benign, and healthy. They achieved 90.50% accuracy and 90.71% specificity. In [44], the authors demonstrate that TL provides improved performance. They employed Inception-v4 [45] pretrained with the ImageNet and DDSM datasets.

Furthermore, in [46], the authors used AlexNet, along with the support vector machine (SVM), to improve classification accuracy and using data augmentation techniques to increase input images. Performance is evaluated on two datasets: DDSM and CBIS-DDSM. The accuracy of the method is 71.01%, and when using SVM, the accuracy becomes 87.2%. In addition, in another work [47], a deep-CNN was proposed in which an MLP was employed in the FC layer to classify mammography images into benign, malignant, and normal. The authors employed a bilateral filter with vector grid computing to maintain edge information in the preprocessing step. They also use hyperparameter tuning to evaluate performance. The results demonstrate that hyperparameter tuning of the final layers produces 96.23% overall accuracy and 97.46% average accuracy. In addition, in [48] used the CBIS-DDSM dataset to develop an automatic mammogram classification method based on TL and data augmentation. ResNet was fine-tuned to produce good results, and achieved accuracy was 93.15%.

To be specific, despite the encouraging results achieved by the CNN architectures in detecting breast cancer, the large number of hyperparameters is a barrier to attaining improved results. Thus, the use of hyperparameter optimization for CNN architecture is essential to enhance the performance of CNNs. In this paper, an optimized CNN model based on an improved MPA (IMPA) algorithm was proposed for breast cancer classification, which may aid health professionals in breast cancer diagnosis.

This section explained the concept, the mathematical representation of the marine predators algorithm (MPA), along with the opposition-based learning (OBL) strategy, the convolutional neural networks (CNNs) architecture, and transfer learning (TL).

In this section, the proposed IMPA is explained. When examining the performance of the original MPA, it is clear that it does not adequately explore all search space solutions. Furthermore, because it divides the optimization phases into three discrete portions, it suffers from a convergence rate. Thus, the original version of the MPA is improved using the OBL strategy (IMPA) and used to optimize the hyperparameters of the pretrained CNN architecture. Algorithm 1 presents the pseudo-code for the IMPA. The MPA’s diversity was improved in the search process using the OBL strategy during the initialization phase to improve the search operation as follows:where \(\mathrm{Opp}_s\) is a vector produced by applying OBL, \(lb_a\), and \(ub_a\) are lower and upper bounds of the ath component of Y, respectively.

The phases of the proposed IMPA are described in the next subsections:

This section introduces the proposed IMPA-ResNet50 model based on TL from a pretrained CNN architecture. The pretrained model employed in this proposed model is ResNet50. The IMPA algorithm is used to optimize the hyperparameters of the pretrained CNN model to get the best performance of this model. The ResNet50 model was trained using TL methods after determining the best values for the parameters. After the model has been trained, it is verified using a different test set. The test set is then used to validate the fully trained model. The proposed model has split into four main phases, as shown in Fig. 3. These phases operate in this order:

In the first phase, the datasets were enhanced and divided into two training and test sets. Also, multiple data augmentation procedures have been applied to increase training sets. The proposed model has been applied to two datasets, namely CBIS-DDSM and MIAS. In the second phase, the IMPA is used to optimize the hyperparameters in the pretrained CNN architecture (ResNet50). In the third phase, ResNet50 was completely trained using the values of the hyperparameters determined in the second phase that helped the architecture accurately diagnose the test set in the next phase. The phases of IMPA-ResNet50 will be presented in detail in the following sections.

This section describes and analyzes the results to validate the proposed IMPA-ResNet50 model’s performance for classifying mammography of breast cancer. This section is structured as follows: Sect. 6.1 presents the datasets used in this paper. Section 6.2 presents the platform used in this paper. Section 6.3 presents the experimental parameter settings for the IMPA. Section 6.4 presents the results of the CBIS-DDSM dataset, and Sect. 6.5 describes the results of the MIAS dataset. Section 6.6 presents the comparative results of the proposed IMPA-Resnet50 with four other meta-heuristic algorithms paring with the ResNet50 architecture: GSA-ResNet50, HHO-ResNet50, WOA-ResNet50, and MPA-ResNet50. Eventually, the limitations of the proposed model are illustrated in Sect. 6.7

Deep Learning is one of the essential techniques in medical imaging classification. Convolution Neural Networks are examples of common DL techniques used for biomedical image classification that involves layer-wise automatic feature extraction. The preparation of CNNs for classification objectives depends on the knowledge of hyperparameter tuning. Hyperparameters for each layer are different. Because these hyperparameters determine a CNN model’s performance, they need to be fine-tuned to achieve great results. It is not suitable to optimize hyperparameters manually because manually selecting them is quite a complex and time-consuming task. Meta-heuristic algorithms have significantly influenced hyperparameter optimization in several fields. This paper proposes a novel classification model for breast cancer based on a hybridized pretrained CNN architecture (ResNet50) and an improved meta-heuristic optimization algorithm. The marine predators algorithm (MPA) is the used optimization algorithm, and we improve it using the OBL strategy to improve exploitation and avoid getting stuck in the local optimal. The Improved Marine Predators Algorithm (IMPA) is used to find the best values for the hyperparameters of the ResNet50 architecture, resulting in a model called IMPA-ResNet50. To the best of our knowledge, this is the first paper using the IMPA as an optimization algorithm to optimize the hyperparameters for the ResNet50 architecture for breast cancer classification. The proposed IMPA-ResNet50 model includes four phases: (1) the data preprocessing and augmentation phase, (2) hyperparameter optimization phase, (3) learning phase, and (4) performance evaluation phase. The proposed model is compared with other state-of-the-art methods and other CNN approaches. The comparison results showed the effectiveness of the proposed model in diagnosing breast cancer. The evaluation is performed on two mammography datasets: the curated breast imaging subset of DDSM (CBIS-DDSM) and the mammographic image analysis society (MIAS). To demonstrate the effectiveness of the IMPA in determining the best values for the hyperparameters of the ResNet50 model that can achieve the most significant accuracy, we compare it with the MPA-ResNet50 model, also compare it with the pretrained CNN ResNet50 model based on the manual search. The first experiment applied to the CBIS-DDMS consisted of 5283 images for two classes, normal and abnormal. The obtained accuracy was 98.32% in the testing phase according to the IMPA-ResNet50 model and 95.95% accuracy according to the MPA-ResNet50 model, which means the improved version for the MPA (i.e., IMPA) significantly improves performance when used to optimize the ResNet50 hyperparameters compared to the original MPA. The second experiment was performed using another dataset (MIAS) consisting of 1290 images after applying the data augmentation techniques. The obtained accuracy was 98.88% according to IMPA-ResNet50 95.95% accuracy according to MPA-ResNet50. Also, compared to four meta-heuristic algorithms: GSA, HHO, WOA, and the original MPA algorithm, the results of the comparison showed the effectiveness of the proposed algorithm. The proposed method’s performance has been assessed using five measures: accuracy, sensitivity, specificity, precision, and F-score. The results showed that the proposed model gets better results than the other competing algorithms, which means using the IMPA as a metaheuristic optimization algorithm to determine the ResNet50 architecture’s hyperparameters boosts the classification models to achieve the best performance for a breast cancer diagnostic.

In future work, multiple datasets with more images will be used and evaluated in the proposed model. Also, various pre-trained models such as DensNet201, DensNet121, and Inception will be used to classify breast cancer in conjunction with the proposed IMPA algorithm. In addition, various metaheuristic algorithms will be used for hyperparameter tuning. Also, we will test the proposed method’s performance in solving different medical image classification problems and different diagnosis applications, applying the IMPA algorithm to solving other medical and engineering issues, and using it as a feature selection method. Furthermore, various feature extraction methods will be used with CNN architecture to improve the classification accuracy further. Transfer learning with other models can enhance the performance.