Efficient Breast Cancer Diagnosis from Complex Mammographic Images Using Deep Convolutional Neural Network

Medical image analysis places a significant focus on breast cancer, which poses a significant threat to women's health and contributes to many fatalities. An early and precise diagnosis of breast cancer through digital mammograms can significantly improve the accuracy of disease detection. Computer-aided diagnosis (CAD) systems must analyze the medical imagery and perform detection, segmentation, and classification processes to assist radiologists with accurately detecting breast lesions. However, early-stage mammography cancer detection is certainly difficult. The deep convolutional neural network has demonstrated exceptional results and is considered a highly effective tool in the field. This study proposes a computational framework for diagnosing breast cancer using a ResNet-50 convolutional neural network to classify mammogram images. To train and classify the INbreast dataset into benign or malignant categories, the framework utilizes transfer learning from the pretrained ResNet-50 CNN on ImageNet. The results revealed that the proposed framework achieved an outstanding classification accuracy of 93%, surpassing other models trained on the same dataset. This novel approach facilitates early diagnosis and classification of malignant and benign breast cancer, potentially saving lives and resources. These outcomes highlight that deep convolutional neural network algorithms can be trained to achieve highly accurate results in various mammograms, along with the capacity to enhance medical tools by reducing the error rate in screening mammograms.


Introduction
Te predominant cause of cancer-related deaths among women globally is breast cancer [1,2]. According to a report by the World Health Organization's (WHO) cancer research institute, the International Agency for Research on Cancer, in 2018 globally, 17.1 million breast cancer cases were reported. Te number of cases is predicted to increase to double the amount by 2025 [3]. Breast cancer is a highly invasive tumor that primarily afects women [4]. Te high death rate among women makes it the second deadliest malignancy after lung cancer [5,6]. A study by the National Institute of Cancer in China found 1.67 million breast cancer and 522,000 deaths cases from 2008 to 2012 [7].
Despite extensive eforts from medical professionals and researchers, a defnitive method for treating breast cancer has yet to be established and reliable evidence for its prevention remains elusive [8][9][10][11]. Some components of breast cancer tissues are highly malignant and pose a severe danger to patients' lives as they can spread to other vital organs [12][13][14][15]. Te growth of mammary cells can lead to tumors in women. Tumors are classifed as benign or malignant based on the area, size, and location, using the BI-RAD scores [16,17]. Benign tumors are not life-threatening and can be treated through medication to prevent further growth [17,18]. Malignant tumors, on the other hand, can spread to other parts of the body via the lymphatic system or blood, making them much more dangerous [19][20][21][22]. Tis uncontrolled cell proliferation in the breast leads to the formation of malignant tumors, which can only be treated through surgery or radiation therapy [23,24].
Early detection of breast cancer is crucial for accurate diagnosis and analysis, and many researchers are turning to biomedical imaging to aid specialist radiologists. Various methods such as MRI, mammography, and ultrasound are utilized to identify breast carcinoma [25,26]. However, the large volume of images challenges radiologists in identifying potential cancerous areas. Terefore, an efcient automated method is needed, and computer-aided diagnostic (CAD) systems are being utilized in aiding radiologists in detecting cancerous breast tumors [27,28].
Increasingly, deep learning techniques are being applied to medical imaging to develop automated computer-aided diagnosis (CAD) systems [29][30][31][32][33][34][35]. Deep learning is considered the most efective method for detecting and classifying medical images [29,30]. With these techniques, the mammogram image's signifcant low to high-level hierarchical features can be directly extracted, making deep learning the most reliable medical imaging method [29]. Several CAD systems based on deep learning have been developed for breast lesions detection, which outperforms traditional systems [36]. Accurate detection of breast lesions is crucial for improving breast cancer diagnosis [29,31]. However, detecting these lesions can be challenging due to their varying texture, shape, position, and size. Deep learning and image processing methods have been proposed to overcome the limitations of conventional technology, which cannot perform automated identifcation [29]. Te fnal stage in the CAD model is the classifcation of breast lesions into benign or malignant, which is important in assessing the correctness of the diagnostic [30].
Currently employed methods for detecting breast cancer are slow, costly, and require extra eforts to run the radiology equipment. Accurately detecting breast cancer automatically from an image processing perspective is not easy. Hence, early diagnosis and proper treatment are deemed crucial. Terefore, an efcient screening system and automation are necessary for breast cancer detection due to the following reasons [12]: incorrect diagnoses and predictions, tumors appearing in low contrast areas, unreliable human diagnoses, overburdening of radiologists, human error in diagnosis, need for large training data to avoid overftting in deep learning algorithms, high computational complexity, and longer processing time for accurate tumor identifcation.
A novel breast cancer detecting system is proposed with an improved architecture that integrates deep convolutional neural network (DCNN) and breast mammogram images to address previous drawbacks mentioned above. Te proposed system intends to divide breast tumors into benign and malignant categories. Te system's performance is evaluated and compared with existing classifcation systems using a public mammographic image dataset named INbreast. Te new system includes transfer learning to fne-tune the pretrained DCNN and detailed results from experiments on the INbreast dataset. Te system's performance is evaluated using the following metrics: AUC, specifcity, accuracy, sensitivity, and F-1 score.
Te rest of the paper is organized as follows. Section 2 presents the related work. Section 3 provides the proposed approach for breast cancer detection. Section 4 presents experimental analysis, results and comparison with existing work. Section 5 concludes this paper and presents future work.

Related Work
Breast cancer diagnosis in modern medical procedures often involves using mammography images [37]. A summary of recently developed systems for breast cancer diagnosis using mammogram images is presented in this section.
Structured support vector machine (SSVM) and conditional random feld (CRF) are two structured prediction techniques proposed in [38] to classify mass mammograms. Both approaches used potential functions based on deep convolution and belief networks. Te results demonstrated that the CRF method outperformed the SSVM method in training and inference time. Authors in [29] utilized fourfold cross-validation on X-ray mammograms from the INbreast dataset to estimate a full-resolution convolutional network (FrCN). It resulted in an F1 score of 99.24%, an accuracy of 95.96%, and a Matthews correlation coefcient (MCC) of 98.96%. In another study [39], the BDR-CNN-GCN approach was proposed by combining a graphconvolutional network (GCN) with a basic 8-layer CNN that includes batch normalization and dropout layers. Te fnal BDR-CNN-GCN model was formed by integrating the twolayer GCN with the CNN. Tis method was tested using the MIAS dataset, and successful results were obtained with a 96.10% accuracy level.
Authors in [40] proposed modifying the YOLOv5 network for identifying and classifying breast cancers, with the algorithm run using specifc parameter values. Te modifed YOLOv5 was compared with a faster RCNN and YOLOv3, achieving an accuracy of 96.50% and an MCC value of 93.50%. Te diverse features (DFeBCD) method was proposed by [41], which classifed mammograms into two categories normal and abnormal. Tey used two classifers, an emotion learning-inspired integrated classifer (ELiEC) and SVM, with the IRMA mammography dataset. Te ELiEC classifer outperformed SVM, achieving an accuracy rate of 80.30%. In [30], a deep-CNN model that utilized transfer learning (TL) was introduced to prevent overftting when working with small datasets. DDSM, MIAS, BCDR, and INbreast were used to assess its performance. INbreast dataset achieved an accuracy of 95.5%, the DDSM dataset 2 Computational Intelligence and Neuroscience achieved an accuracy level of 97.35%, and the BCDR database achieved a 96.67% accuracy level. Authors in [42] for extracting features from breast mammograms utilized lifting wavelet transform (LWT). Feature vectors' size was reduced using linear discriminant analysis (LDA) and principal component analysis (PCA). Te classifcation was performed using the moth fame optimization and extreme learning machine (ELM) approach with MIAS and DDSM and datasets, achieving accuracy of 95.80% and 98.76%, respectively. In addition, researchers have also trained the CNN Inception-v3 model on 316 images, resulting in a sensitivity of 0.88, specifcity of 0.87, and an AUC of 0.946 [43]. Furthermore, in [44], a CNN and TL classifcation method was proposed to evaluate the performance of eight fne-tuned pretrained models. Authors in [45] presented a hybrid classifcation model using Mobilenet, ResNet50, and Alexnet with an accuracy level of 95.6%. In [46], four diferent CNN architectures (VGG19, InceptionV3, ResNet50, and VGG16) were utilized for model training using 5000 images, while prediction models were evaluated on 1007 images.
Authors in [47] utilized alpha, geostatistics, and diversity analyses forms in their proposed breast cancer detection method. Tey employed the SVM classifer on MIAS and DDSM databases, which resulted in a detection accuracy level of 96.30%. Te SVM classifer and gray level cooccurrence matrix (GLCM) were employed by [48] for detecting breast cancer abnormalities in the MIAS data set. Teir method achieved an accuracy of 93.88% and surpassed the performance of the k-nearest neighbour (kNN) algorithm. Authors in [49] used AlexNet and SVM to enhance classifcation accuracy with data augmentation techniques. Te method achieved 71.01% accuracy, which increased to 87.2% with SVM and was evaluated on DDSM and CBIS-DDSM datasets.
A DenseNet deep learning framework extracted image features and classifed cancerous and benign cells by feeding them into a fully connected (FC) layer. Te efectiveness of this technique was evaluated by adjusting the hyperparameters [50]. An algorithm named DICNN was developed by Irfan et al. [51], which uses a dilated semantic segmentation network and morphological operation. Combining these feature vectors with SVM classifcation yielded an accuracy of 98.9%.
Although prior breast cancer detection and classifcation systems have improved information extraction, several issues still need attention, such as low contrast in tumor location, high memory complexity, long processing time, and the need for a large amount of training data for deep learning approaches. In response to these problems, we propose a new approach to breast cancer detection and classifcation, which will be discussed in detail in the following section.

Methodology
In this section, the processes used for implementing our proposed scheme are described in depth. Te system consists of the following steps: (1) image enhancement, (2) image segmentation, (3) feature extraction and the selection, and (4) feature classifcation. Te proposed system is illustrated in Figure 1.

Dataset.
Tis study used a digital breast X-ray database named INbreast to implement the proposed CAD approach. Te INbreast dataset is a public database that contains more recent FFDM images. It typically has an image size of 3328 × 4084 pixels. It contains 115 patients' cases along with 410 mammograms with both craniocaudal (CC) view and a mediolateral oblique (MLO) view. Of these 115 patients, 90 had mammograms taken of both breasts, totaling 360 images, while the other 25 had only two mammograms taken each. In total, 410 mammograms were produced from 115 patients, including cases of normal, benign, and malignant breasts. 107 cases with breast lesions were used from the MLO and CC views for evaluation purposes.

Convolutional Neural Network.
Tis subsection will examine the fundamental structure of all convolutional neural network (CNN) architectures. CNNs are deep neural networks used for image recognition and classifcation. In recent years, CNNs have become a crucial tool in image analysis, especially for identifying faces, text, and medical imaging. CNNs have a long history of success in image classifcation and segmentation, frst developed in 1989. CNNs replicate the human brain's visual information processing by incorporating layers of "neurons" that only respond to their local surroundings. Tese networks can understand the topological aspects of an image through a combination of convolutional, pooling, and fully connected (FC) layers. Te architecture of a CNN is shown in Figure 2.

Convolutional Layers.
Te convolutional layers are assembled into feature maps based on local connections and weight distribution principles. A flter bank, a group of weights, connects neurons in a feature map to corresponding local regions in the preceding layer. Each feature map uses a diferent flter bank, and all the units in the map share the same flter row. Tis weight distribution and local connection help reduce the number of parameters by utilizing the close relationship between neighboring pixels and location-independent image features. Te output of the weights is then sent to an activation function, such as ReLU or Sigmoid. Tis activation function enables the nonlinear transformation of the input data, which is necessary for the following processing stages. Figure 2, the pooling layer follows the convolution layer and uses subsampling to integrate the features from the convolutional layer into a single layer semantically. Tis layer's primary objective is to decrease the size of the image by combining pixels into one value while preserving its features. In this layer, typical operations include max as well as main pooling.

Fully Connected Layer.
Te last layer in CNN is the dense classifcation layer, which is responsible for determining the category of input data based on extracted features from CNN. Te number of units in the FC layer is the same as the number of diferent classifcations (categories).

Proposed Workfow.
Tis section provides the proposed workfow for breast cancer diagnosis using a deep convolutional neural network.

Image Enhancement.
Image enhancement refers to increasing contrast and suppressing noise in mammogram images to assist radiologists in detecting breast abnormalities. Various image enhancement methods exist, including adaptive contrast enhancement (AHE). AHE improves the local contrast and reveals more image details, making it a helpful technique for enhancing both natural and medical images [52]. However, it may also result in considerable noise. In this paper, we utilized the contrast-limited adaptive histogram equalization (CLAHE) technique, a form of AHE, to enhance image contrast [52]. A drawback of AHE is that it can over-enhance the images due to the integration process [49]. To mitigate this issue, CLAHE is used as it limits the local histogram by setting a clip level, thus controlling contrast enhancement. Figure 3 illustrates an image enhanced by the CLAHE algorithm.

Tresholding Method.
One of the simplest image segmentation methods is the thresholding method [55,56]. Te pixels of the image are split according to their intensity level. Te global threshold is the most commonly used thresholding technique [57]. It is accomplished by setting a threshold value (T) constant throughout the image. Te output image is derived from the original image based on the threshold value.

Region-Based Segmentation Methods.
It is a simple approach compared to other methods, as it involves dividing an image into diferent sections based on predetermined. Compared to others, it is a straightforward method because it entails separating an image into diferent sections based on predetermined criteria [58]. Tere are two primary kinds of region-based segmentation: (1) region splitting and merging and (2) region growing. Region growing allows the removal of a region from an image using defned criteria, such as intensity. It involves selecting a starting seed point. It is important to note that unlike region growing, region splitting and merging work on the entire image [59].
In the present study extracting the region of interest (ROI) involves using both thresholding and region-based techniques. Te tumor in the INbreast dataset samples cites moreira2012inbreast is labeled by a white bounding box, as shown in Figure 4. For extracting ROI, the tumor region is frst determined by setting a threshold value based on the white color pixels in the image. Te threshold for all images is determined to be 80 after several attempts, independent of tumor size. After identifying the greatest area inside this threshold within the image, the tumor is automatically cropped. Figure 4 shows ROI extracted using threshold and region-based methods.
Te method for extracting ROI can be summarized in four steps: (1) Tresholding the grayscale mammogram image to create a binary image. (2) Labelling and counting the binary image objects, then retaining only the largest one, which is the tumor, as defned by the white bounding box. (3) Assign the largest area within the threshold value to "1" and the rest a value of "0." (4) Multiply binary image with original mammogram image for obtaining fnal ROI without including other parts of breast or artifacts.

Feature Extraction and Selection.
Numerous methods exist for feature extraction. Due to their exceptional performance, deep convolutional neural networks (DCNN) garnered signifcant interest in recent years. Consequently, the DCNN is utilized in this paper.

Deep Convolutional Neural Network. Te success of DCNN in image classifcation and analysis has been
documented in various studies [60,61]. Convolutional neural networks (CNNs) are composed of multiple trainable stages that culminate in a supervised classifer and feature maps [62]. Tree primary types of layers are employed to build CNN structures: convolutional, pooling, and fully connected (FC) layers [63]. Te ResNet50 CNN classifcation model categorizes breast cancer as benign or malignant in this work.

Feature Learning through Transfer Learning.
Machine learning has various feature learning methods (FL), allowing a system to automatically identify the representations required for feature detection, prediction, or classifcation from a preprocessed dataset [64]. Tis implies that  In deep learning, transfer learning (TL) is a widely-used technique that enables the utilization of a pretrained network for new prediction or classifcation tasks. Tis is achieved by adjusting the parameters of the pretrained network with randomly initialized weights for the new task. TL typically results in faster training than starting from scratch and is considered an optimization that saves time and improves performance, as stated in [65]. For this purpose, transfer learning is utilized to fne-tune ResNet50 CNN. Tis involves using pretrained weights from the ImageNet dataset [66] for retraining after preprocessing the collected dataset. Te network parameters and hyperparameters are optimized during this process.
3.4.6. Classifcation. Te features are taken from ResNet-50 and processed via a fully connected (FC) layer with a 40% dropout rate to avoid overftting [67,68]. Tis layer is then activated with the rectifying function, ReLU. All negative values are set to zero in the input matrix, while other remains unchanged. Te use of ReLU leads to faster and more reliable convergence than a sigmoid activation function during training deep networks [69]. Te output layer comprises a sigmoid function (binary classifer) to provide class probabilities. Te sigmoid function normalizes the input into two outcomes, i.e., malignant vs. benign [70].

Evaluation and Results
Te proposed deep convolutional neural network for mammogram imaging undergoes examination and validation in this section. Information about benchmark datasets, assessment metrics, and comparisons to other leading techniques are also covered.

Image Acquisition Process.
Te proposed system's performance is evaluated using digitized mammogram images from the INbreast dataset [71]. Te database is used to demonstrate the efciency and reliability of the proposed method for identifying breast cancer. INbreast dataset includes 336 mammogram images, with 269 abnormal and 69 normal images, where 220 are benign and 49 malignant cases. Tables 1 and 2 show the distribution of mammography images.

Metrics of Performance. Te purpose of cross-validation
is to improve efciency, validate performance, and assess the results from the dataset. To assess the classifcation efciency of the proposed method, multiple metrics are utilized such as confusion matrix, accuracy, sensitivity, specifcity, error rate, F1 score, and area under the curve (AUC). All these metrics act as benchmark values for comparing the proposed method against previous algorithms [72]. Tese measurements are defned as follows.    Computational Intelligence and Neuroscience

Specifcity.
Te chances that the test will correctly recognize the patient who has the disease is shown in the following equation:

Sensitivity.
Te chance that the test will correctly recognize a patient with the disease is shown in equation:

F1
Score. It is a weighted average of precision and recall used for assessing the classifer's performance. It considers both false positives and negatives in its calculation, as shown in the following equation:

Area Under the Curve (AUC).
AUC is the classifer's ability to distinguish between benign, normal, and malignant mammograms.

Results and Discussion
For this study, a subset is taken from the INbreast dataset, and each sample is increased to four images. During the experiment, 60% images were used for training, and the remaining 40% were used for testing. Te samples were frst subjected to enhancement and segmentation according to the procedures described in the "Methodology" section. Afterward, features were extracted from the samples using a CNN. Finally, all the samples were classifed using ResNet-50. Te proposed DCNN method categorizes mammogram images of breast tumors into benign or malignant. A dataset named INbreast is used for experimentation. Table 3 displays the classifcation accuracy achieved by the proposed ResNet-50 method across the INbreast database. From the INbreast dataset, 132 benign and 29 malignant image samples were selected for training, and 20 malignant and 88 benign for testing. Te resulting accuracy is 93%. Te proposed ResNet-50 approach is also compared quantitatively with previously existing algorithms. Te study's results revealed that the presented approach outperformed these algorithms with high accuracy, specifcity, F1 score values, and sensitivity.
As shown in Table 3, the proposed approach demonstrated improved results on the INbreast database with an accuracy of 93.0%, specifcity of 93.86%, and sensitivity of 93.83%. It outperforms other methods in terms of accuracy. Although the accuracy achieved by [16] is slightly higher at 91.0%, the proposed approach still exhibits the best performance compared to the other methods. Compared to existing methods, the proposed approach enhances breast cancer detection and classifcation performance. It can potentially be used for real-time evaluation and to support radiologists in automating the analysis of mammogram images. However, performance may vary when the same method is applied to diferent datasets due to factors such as background noise, lighting conditions, occlusion, overftting, and the nature of the method.
Te performance of the presented approach is also evaluated using the confusion matrix and ROC curves. Figure 5 illustrates the confusion matrix on the INbreast data set. AUC, a crucial statistical metric in the ROC curve, is computed using the INbreast data set. Metric in the ROC curve is calculated INbreast data set. ROC curves were constructed based on true positive rate (sensitivity) and false positive rate (1-specifcity) rates, controlled by the threshold of the obtained probability maps. Figure 6 shows the ROC curve graph. Table 4 presents our proposed system's results of breast cancer detection. Te proposed approach achieved an F1 score and AUC of 93.03% and 93.02%, respectively, on the INbreast database.
In recent years, breast cancer detection and classifcation applications have gained widespread use in the medical feld, making the diagnostic process more accurate [76,77]. Te goal of the proposed method is to enhance clinical diagnosis by enhancing the detection of breast cancer. Te opinions of two medical specialists were gathered based on the accuracy level generated by our proposed algorithm. Tese experts expressed their appreciation for the improved results of ResNet-50 compared to other approaches. To sum it up, the proposed approach enhances performance compared to other methods and can be utilized for real-time evaluations along with helping radiologists automate the evaluation of mammograms.

Conclusion
Te proposed system aimed to detect malignant breast masses and classify benign and malignant tissues in mammograms. A novel computer-aided detection (CAD) system is proposed, which involves thresholding and region-based segmentation techniques. A region-based method with a threshold of 80 determines the largest area included in this threshold. A deep convolutional neural network (DCNN) is utilized during feature extraction. Specifcally, the ResNet-50 is retrained to classify the mammograms into two classes (malignant or benign), and its parameters were modifed to classify breast mammograms. Te proposed approach is applied to the INbreast database to evaluate its performance of the proposed approach. Te proposed method achieved an accuracy of 93.0%, specifcity of 93.86%, AUC of 93.02%, a sensitivity of 93.83%, and an F1 score of 93.03%, which are extremely satisfying results. Te proposed method surpasses the detection and classifcation of mammograms, delivering more precise results and improved visual outcomes compared to other systems. Te proposed system efciently detects and classifes malignant breast masses with reduced computation time and produced satisfactory results. Alternative networks, such as deep convolutional networks (VGG) and AlexNet architecture, will be proposed for future development. In the future, we intend to extend this work by collecting large datasets on breast cancer in diferent age intervals to detect cancer in its early stages.

Data Availability
Te (Breast Cancer Diagnosis) data used to support the fndings of this study are included within the article.

Conflicts of Interest
Te authors declare that they have no conficts of interest.