Breast Cancer Calcifications: Identification Using a Novel Segmentation Approach.

Breast cancer is a strong risk factor of cancer amongst women. One in eight women suffers from breast cancer. It is a life-threatening illness and is utterly dreadful. The root cause which is the breast cancer agent is still under research. There are, however, certain potentially dangerous factors like age, genetics, obesity, birth control, cigarettes, and tablets. Breast cancer is often a malignant tumor that begins in the breast cells and eventually spreads to the surrounding tissue. If detected early, the illness may be reversible. The probability of preservation diminishes as the number of measurements increases. Numerous imaging techniques are used to identify breast cancer. This research examines different breast cancer detection strategies via the use of imaging techniques, data mining techniques, and various characteristics, as well as a brief comparative analysis of the existing breast cancer detection system. Breast cancer mortality will be significantly reduced if it is identified and treated early. There are technological difficulties linked to scans and people's inconsistency with breast cancer. In this study, we introduced a form of breast cancer diagnosis. There are different methods involved to collect and analyze details. In the preprocessing stage, the input data picture is filtered by using a window or by cropping. Segmentation can be performed using k-means algorithm. This study is aimed at identifying the calcifications found in bosom cancer in the last phase. The suggested approach is already implemented in MATLAB, and it produces reliable performance.

1. Introduction

AI and machine learning are recently widely used in health care for the prediction of critical diseases like colorectal cancer, Alzheimer, fetal brain abnormality detection, and type-2 diabetes risk prediction, and the present study used AI and ML for breast cancer prediction.

Breast cancer is an abnormal development of malignant cells in the breast. Cancer spreads to other parts of the body if left untreated. Breast cancer, excluding skin cancer, is the most prevalent form of cancer among women in the United States, accounting for one in every three cancer diagnoses. In 2005, the United States was projected to have an estimated 211,240 new invasive cases of breast cancer among women. In 2005, about 1,690 additional male cases of breast cancer were anticipated. Breast cancer incidence increases beyond the age of 40. Women over the age of 50 have the greatest incidence (about 80% of invasive cases). Along with invasive breast cancer, women are projected to develop 58,590 new instances of in situ breast cancer in 2005. Around 88 percent of them will be diagnosed as ductal carcinoma in situ (DCIS). DCIS instances are being detected as a direct consequence of increasing mammography screening. Additionally, this screening technique identifies invasive tumors at a less advanced stage than would have happened otherwise.

Breast cancer was expected to claim an estimated 40,870 lives in 2005 (40,410 women and 460 males). Breast cancer is the second leading cause of mortality from cancer in women. According to the most current statistics, death rates decreased substantially between 1992 and 1998, with the greatest reductions occurring among younger white and black women.

Breast cancer constitutes one-third of all forms of cancers diagnosed by females. This is the primary reason of 18% of deaths from cancer worldwide. Up to now, breast cancer was the major cause of death amongst females; however, since 1985 it has been ranked second after lung cancer. It is not possible to prevent this cancer due to the lack of adequate knowledge regarding its cause. Nevertheless, modern treatment technologies are extremely efficient against breast cancer. Using these technologies, it is possible to detect this cancer in its early stage. Thus, removing cancer in its early stages is the most efficient way to get a change in the existing breast cancer situation. A technique named mammography is used to detect breast cancer in most reliable manner [1]. On the other hand, about 10 to 30% of female breast cancer patients who experience mammography have negative mammograms. The radiologist could not detect the cancer that was evident on second thoughts in about two-thirds of these false-negative mammograms. Maintaining interest in understanding a huge number of images that shows merely a small number of abnormalities is a complex task for a human radiologist. Therefore, it is required to develop computer-aided systems for diagnosing breast cancer in mammograms. Several ultrasound techniques have been proposed for designing automated breast detection methods. Among such techniques, few ultrasound techniques have been discussed in the section below.

1.1. Automated Breast Ultrasound Technique

The first ever designed technique for bosom cancer screening in females is the automated breast ultrasound technique. This technique was designed for those women who have dense breast tissue and for those individuals who had no breast biopsies or surgeries previously [2]. The detail resolution is sharpened by high center-frequency, and the distinct contrast differentiation is delivered simultaneously by the ultrabroadband performance.

The existing techniques for breast cancer detection, such as mammograms, ultrasounds, and biopsies, were time-consuming, necessitating the development of a computerized diagnostic system using machine learning technology. This approach incorporates algorithms that aid in tumor categorization and cell detection more precisely and efficiently.

The fast development of machine learning, particularly deep learning, continues to pique the attention of the medical imaging industry in using these methods to enhance cancer screening accuracy. Breast cancer is the second largest cause of cancer death in women in the United States, and mammography screening has been shown to decrease mortality.

1.2. Breast-Specific Gamma Imaging (BSGI)

For the recognition of the bosom tissues which collect higher levels of radioactive tracer through which radiation of gamma is emitted, a noninvasive diagnostic technology is designed known as Breast-Specific Gamma Imaging (BSGI). It is commonly studied as Molecular Breast Imaging (MBI). Once the radioactive tracers are administrated intravenously, a gamma camera is used to perform tests.

1.3. Scintimammography

For patients who have suspicious mammograms, the selection for biopsy can be improved which is helpful for mammography and physical examination.

1.4. Sonography

An imaging technique that uses waves of sound instead of ionizing radioactivity to a particular body part is known as breast ultrasound, also commonly called sonography. A small instrument similar to a microphone is placed on the skin to perform this test. Sound waves are emitted by this device, and the echoes bouncing off the body tissues are gathered. A monochrome image is generated on a computer monitor by converting the received echoes. To evaluate certain breast masses and to know if a certain region is a cyst or not, ultrasound is the best method known. By performing only physical exams, one cannot diagnose cysts accurately. To assist doctors in guiding a biopsy needle into few breast lesions, the breast ultrasound might also be used [3-5]. Computer-Aided Detection (CAD) systems: for helping radiotherapist in analyzing images and automating the reporting process, pattern recognition methods are used by Computer-Aided Detection (CAD) systems for ultrasound. The standardized bosom ultrasound reporting is promoted by these systems.

1.5. Computer-Aided Detection with MRI (Medical Resonance Imaging)

For helping the interpretations of radiologists of contrast-enhanced MRI, the Computer-Aided Detection was used. This CAD-with-MRI can be used as an alternative to mammography and screening testing, owing to its high efficiency in identifying breast lesions. Placing a sensor in contact with the breast with a touch array is classified as tactile breast imaging. As the clinician pushes the handheld sensor over the breast and underarm region, the data signals are transformed into multidimensional color pictures. Via such multidimensional color pictures, which appear on the computer screen, the clinician can clearly see the height, form, strength, and position of offender masses [6].

1.6. Electrical Impedance Scanning (EIS)

In conjunction to mammography, a confirmatory test was designed for Electrical Impedance Scanning (EIS). Small electrical currents are used by this device to detect the abnormal breast tissue. A conductivity map of the breast is created by the produced electrical current from malignant tissues which conduct more electricity as compared to normal tissue. The suspicious sites are identified automatically through this process. On the arm of a handheld device, an electrical patch that moves to the breast is used to transmit the electricity into the body. A probe on the surface of the skin is used to measure it.

1.7. Magnetic Resonance Elastography (MRE)

A phase-contract-based MRI method which is dependent on the quantitative contrast of mechanical properties of benign and malignant tissues is known as Magnetic Resonance Elastography (MRE). In comparison to normal tissues, the breast cancers are harder [7]. A known stressor can be applied to measure this difference, and then, the resulting deformation is measured [8].

1.8. Magnetic Resonance Imaging (MRI)

A nonintrusive imaging methodology in which the magnetic and radio frequency fields are used is known as Magnetic Resonance Imaging (MRI). Very elaborate, cross-sectional pictures of body are generated by MRI to image the body tissues [8]. No ionizing radiation is used by MRI inconsistent with CT, and this process is known to be safe.

1.9. Computer-Aided Tactile Breast Imaging

A tactile array sensor is deployed in tactile breast imaging for generating the data signals in connection with a part of the patient's body. These data signals correspond to pressure gradients that come across the parts of the tactile array sensor. With the gentle movement of the portable sensor all over the breast and underarm regions, the processing of data signals starts into multidimensional color images. These images are directly visible on a computer monitor in a realistic manner. This enables the doctor to analyze the size, shape, rigidity, and place of doubtful deposits at the same time [9].

1.10. Diagnostic/Radiologic Mammography

This radiology-based technique is applied to a person having the indications and symptoms of breast cancer or a patient with breast cancer history or a patient having a history of biopsy process, confirmed as noncancerous breast disease. This process is also applied on patients analyzed by the doctors on the basis of outcomes.

1.11. Digital Mammography

In recent times, radiologists are slowly replacing traditional SFM (Screen-Film Mammography) with FFDM (Full-Field Digital Mammography) for the diagnosis of bosom cancer. In the past, SFM was the major approach of bosom cancer screening in women who have no indication of breast cancer. In the SFM approach, breast images are acquired, displayed, transmitted, and stored as digital data for making them visible on the computer screen. These images are later printed and viewed using a light box. On the other hand, FFDM makes the image interpretation easier than the SFM approach. In this approach, different processes such as image acquisition, processing, review, and data storage are implemented in a direct manner.

2. Calcification

Sometimes, women find minute bits of calcium in their breast with age growth. These minute bits are known as “calcifications.” It is very difficult to feel these calcium bits due to their tiny size. However, these bits can appear on a mammogram in the form of tiny, bright white spots. These bits can be a reason of worry. Therefore, more tests are required. As an example, tense clusters or lines of minute calcifications may be the symptom of breast cancer. Calcifications are frequent. Around half of all mammograms in women age 50 and older have calcifications. There may be many reasons behind calcifications. Some reasons include older age, earlier wound, or an infection in the breast tissue. The calcium amount in a female's diet is not the cause of calcification. Earlier breast surgery or radiation therapy may also be the reasons of calcifications in women [10-13].

Calcifications are the signs that appear in the primary stage of bosom cancer that is not palpable. Calcifications are generally related to the ductal carcinoma in situ (DCIS). However, calcification may also occur in invasive cancers. It is defined as a main sign of malignant tumors in the breast imaging treatment. About 41.2% of women are depicted with calcifications as the only indication of cancer in screening programs. Ultrasound images show greater part of calcifications. It is very difficult to visualize some minor patterns. But ultrasound RF signals provide more effectual information. This information includes frequency and scattered property. This information can replicate the calcification features in a comprehensive manner. The high likelihood of tumor in the radiological analysis of breast lesions is indicated by breast calcifications. The American College of Radiology performed the design of the Breast Imaging Reporting and Data System (BI-RADS). Standardizing reporting on breast imaging and providing transparency on the understanding of studies related to breast imaging were the main objective of this process [14-17]. In order to interpret calcifications, a set of rules are implemented in the BI-RADS atlas. These guidelines provided support to the radiologist in differentiating doubtful calcifications from normally benign variations [18, 19]. The examples include vascular and skin calcifications. The patients with doubtful calcifications were suggested for more medical evaluations, such as a biopsy. This can unintentionally instigate to false-positive results due to the doubtful appearance of calcifications related to benign disease. DCIS are the common type of cancers detected by calcifications. Among these cancers, less than 20% are low grade cancers.

2.1. Types of Classifications

Calcifications can be categorized into two types. These types are known as microcalcification and macrocalcification. A brief description of these calcification types is provided below:

Macrocalcifications: on a mammogram, macrocalcifications look large and have a round shape. These types of calcifications are generally not associated with cancer. Therefore, there is usually no need of reporting

Microcalcifications: microcalcifications are very small in size. These types of calcifications may emerge in clusters. These types of calcifications are generally benign or noncancerous. However, it can be an indication of breast cancer. In general, doctor scan detects any change in these calcifications over time. Sometimes, it is required to carry out follow-up tests for cancer detection

Many times, it is not easy to identify whether microcalcifications are benign (noncancerous) or an indication of breast cancer. Therefore, it is required to get multiple images and generally from another mammogram. These images provide support in the identification of microcalcifications as benign, most likely benign, or doubtful for cancer. Microcalcifications as benign or most likely benign demonstrate that the area is not cancerous. On the other hand, more follow-up tests are required for suspicious microcalcifications.

2.2. Calcification Morphology

Breast calcifications can be easily classified by the mammography. The benign and malignant causes are also revealed under it. Imaging evaluation is also required in the detection of calcification in the breast. The main technique that is used for more analysis of breast calcification is known as magnification mammography. It recognizes in the projections named craniocaudal and mediolateral as it permits differentiating the benign milk of calcium from the other troublesome deposits. Magnified mammography allows the clear analysis of morphology, increases the sharpness of an image, decreases noise, and distributes the calcification. Breast calcification morphology evaluations and assessments can be helpful to determine the patient management. To decide whether the calcifications are benign, unspecific, and warranting imaging or biopsy, or suspicious and requiring any biopsy, the morphology and size of calcification are required. In the breast tissues, the larger calcium deposits are called macrocalcifications. These particles are larger than 2 mm and are usually associated with the benign process and associated to fat necrosis, radiation therapy, or mastitis in the plasma cell. The calcification bits that are associated with a malignant process and are smaller than 0.5 mm are called microcalcifications. To know whether the calcifications are benign, likely benign, or malignant, analyzing breast calcification morphology is beneficial [20].

Figure 1 shows the various types of morphology of calcifications. The morphology of breast tissue calcification is further divided into two parts: benign calcification and suspicious calcification. The benign calcification usually does not entail magnification or warrant enhanced monitoring. In certain cases, the calcifications requiring magnification for characterization and histological examination are called suspicious calcifications. The morphology of calcifications has been described as follows.

Figure 1

Morphology of calcifications.

2.2.1. Typically Benign Calcifications

The following are the typical benign calcifications:

Vascular calcifications: these calcifications correspond to deposits of calcium that are present in the mammary artery wall. These calcifications are seen on the one side generally in a bilateral form. These calcifications present in parallel paths that are corresponding to the lateral perspective of every vessel wall. These are related to tabular structures which are in correspondence to the vascular path. These calcifications are popular among patients over the age of 50 with diabetes and chronic kidney dialysis deficiency. It tends to diminish with the utilization of hormone replacement therapy

Skin (dermal) calcifications: the calcifications which are similar to the small calcifications of the sebaceous gland are called skin calcifications. They are generally related to inflammatory processes that are frequent and numerous. The morphology of these calcifications can be polygon and round with a radiolucent center. These calcifications are present in the inflammatory fold, armpit, or areola. Their size normally lies between 1 and 2 mm

Milk of calcium calcifications: these calcifications match with microparticles of calcium oxalate that are present in secular dilatations of TDLU (macro or micro). These calcifications are placed in the breast's central and posterior region, which is called the teacup

Popcorn calcifications: the calcifications which are inclined to amalgamate, indicative of fibroadenomas over time, are called popcorn calcifications. These calcifications are thick, intense, and dense. The outlines of well-designed nodes that are related to calcifications are essential to notice and can be seen in small fibroadenomas. They may be located in the periphery of the nodule when initiated. The size of these calcifications is larger than 2-3 mm

Dystrophic calcifications: these calcifications demonstrate the sign of fat necrosis as the result of a noxa to the breast tissue. This is due to the fact that the medical record of a patient plays a significant role in the identification of disease. These calcifications generally occur after surgery and radiotherapy. In these conditions, these calcifications generally appear nearby to the surgical scar, probably after 3 to 5 years of surgery [21].

Round calcifications/punctuate: these calcifications like their name are round in shape. These calcifications can be of different sizes. These calcifications may appear at the lobular level. They become round shaped if their size becomes larger than 0.5 mm and punctuate in case of smaller size. They generally represent calcium oxalate deposits. These calcifications generally occur in people above 40 years of age

Ring calcifications: these calcifications are referred to as “eggshell” and “radiolucent centered.” They represent encapsulated cystic lesions that contain fat in the liquid state, even though they have also been defined on the surface of plain cysts

Suture calcifications: these calcifications correlate to calcium collected in the material of suture. These are the linear or tubular calcifications. These calcifications may generate knots. The patients who have undergone radiotherapy are more prone to these calcifications. This is due to the fact that the damage caused by radiation changes the treatment and holds up the reabsorption of catgut that provide a matrix for calcium masses

2.2.2. Calcifications of Suspicious Morphology

The following are the suspicious morphology calcifications:

Coarse heterogeneous calcifications: these types of calcifications are generally referred to as irregular, heterogeneous, and irregular and defined calcifications having the tendency to combine. These calcifications may present in the breast stroma or ducts. Generally, these calcifications produce benign lesions, e.g., reactionary fibroadenomas, regions of fibrosis, or trauma. In the other condition, they signify dystrophic calcifications in growth. These calcifications are generally referred to as benign in cases where they are manifold and mutual

Amorphous calcifications: these calcifications are referred to as “powder,” “cloud,” or “cottony.” These calcifications are very small in size, generally less than 0.1 mm. These calcifications neither can be calculated nor can their shape be determined. This is the reason these calcifications are known as “amorphous” (shapeless). In order to classify them as amorphous, such calcifications must not tap in firmly agile projections, because in such a case, they will become milk of calcium calcifications. Most of these calcifications are benign, like those generated in fibrocystic variations, particularly when they are dispersed and bilateral. On the other hand, these calcifications present a segmental or linear dispersion in group form [22].

Fine pleomorphic calcifications: these calcifications are also referred to as “crushed stone.” These calcifications have various shapes and sizes and are angled and heterogeneous. The diameter of these calcifications is around 0.5 and 1 mm which is lower than heterogeneous coarse calcifications

Fine linear or branched calcifications: these calcifications are small in size, generally below 0.5 mm. These calcifications are thin, linear, and generally discontinuous with rough boundaries. These calcifications are produced in calcified necrotic debris inside a duct contaminated by the carcinoma. This means that these calcifications provide calcium molds in an uneven duct

3. Machine Learning Algorithms

Machine learning algorithms have been generally used in the diagnosis of breast cancer over the years. These algorithms forecast to get different imminent from data patterns. Machine learning is a part of artificial intelligence (AI). Machine learning uses an array of statistical, probabilistic, and optimization devices for learning and improving performance in an automatic manner. Machine learning algorithms can be implemented on new data and past experiences with no explicit programmed instructions. The data is analyzed using both statistics and machine learning. Machine learning deals with compound, large, and high dimensional data. Machine learning algorithms generally have the ability to extract main features and potential rules. It is very challenging to extract these characteristics and rules with the help of conventional statistics. In contrast to statistics approaches, machine learning algorithms are more suitable for breast cancer datasets [23].

Machine learning algorithms are prearranged into taxonomy on the basis of the required result of the algorithm. Some common types of algorithms are as follows:

Supervised learning: in supervised learning, the algorithm produces a function for the mapping of inputs to required outputs. Classification issue is the one average definition of the supervised learning task. In these algorithms, the beginner is desired to learn (to estimate the behavior of) a function. This function turns a vector into one of numerous classes by searching numerous input and output instances of the function

Unsupervised learning: unsupervised learning performs the modeling of input sets. In these algorithms, specified examples do not occur

Reinforcement learning: reinforcement learning trains the algorithm by utilizing the arrangement of reward and punishment

The following are some popular machine learning algorithms.

3.1. Naive Bayesian (NB) Networks

Naive Bayesian are extremely clear Bayesian networks. These networks are made up of directed acyclic diagrams with just one parent and numerous children. The parent represents the unobserved node while the children represent observed nodes. These networks are based on strong supposition of independence among child nodes regarding their parent. Therefore, the independence model (naive Bayes) depends on estimation. In contrast to other more innovatory learning algorithms such as ANNs, Bayes classification models are generally less accurate.

3.2. Support Vector Machines (SVMs)

Support Vector Machines are the latest supervised machine learning algorithms. The Support Vector Machine directly corresponds to traditional multilayer perceptron neural networks. These algorithms are built on the idea of a margin on the either side of a hyperplane. This hyperplane divides two data classes. These algorithms maximize the margin and hence create the biggest feasible distance among the separating hyperplane and the instances on either side for reducing the expected generalization error.

3.3. Decision Trees

Decision trees (DT) perform the classification of instances by arranging them on the basis of value of characteristics. Every node within a decision tree characterizes an attribute in an instance selected for classification. A value that the node can presume is represented by each branch. The classification of instances is performed at the beginning of the root node. Later, these instances are arranged based on their attribute's values. Decision tree learning is utilized in data mining and machine learning. Here, this approach makes use of a decision tree as a predictive model. This model performs the mapping of observations about an object to provide the target value of the object. Classification trees or regression trees are the more expressive names for these types of classification models.

3.4. k-Means Algorithm

The k-means is a frequent clustering technique. This algorithm is in the category of partitioning-based clustering techniques. The k-means clustering algorithm is utilized in a lot of fields. These fields include information retrieval, computer vision, and pattern recognition. In the k-means algorithm, representing every cluster is represented by the mean (called the centroid) of its data points. The aim of this algorithm is to partition set of n objects into k (essential number of clusters which is provided as input to the technique) clusters. This result in the intracluster similarity is high and intercluster resemblance is low. The object mean value in a cluster is utilized for similarity measurement. The initial centroids of the cluster are selected at random. On the basis of the Euclidean distance between the centroid cluster and the object, the objects in clusters are migrated iteratively by transferring them from one cluster to another by this technique. After this transposition of objects into clusters, the object's mean value is again shown for the cluster. This occurrence gives the innovational centroid of the cluster. This process is often applied till the merging of function of principle.

3.5. Linear Regression

Linear regression is a supervised machine learning algorithm. This algorithm generates continuous predictions having a consistent slope. This algorithm is employed for predicting values in a steady range, instead of endeavoring to categorize them into different classes. This algorithm can be divided into two categories. These categories are known as simple linear regression and multivariable linear regression. The first category makes use of a conventional slope-intercept format. In this format, variables m and b signify variables that the algorithm will attempt to “learn” for generating the most detailed prediction outputs. In the following equation (1), variable x denotes input data while y stands for our prediction:

A multivariable linear equation is somewhat complex than the simple linear regression. In equation (2), w refers to the coefficients or weights that the classifier will try to understand:

In this equation, x, y, and z variables signify the features or different parts of information, regarding every observation.

3.6. Random Forest

Random forest is an adaptable and user-friendly machine learning algorithm. This algorithm, most of the time, generates good outcomes even without the tuning of a hyperparameter. This algorithm is used in various applications due to its unfussiness and range. It is a supervised learning algorithm. The “forest” it constructs represents an ensemble of DTs (decision trees). Generally, the training of DTs is carried out with the “bagging” method. This bagging approach is based on the concept that the overall result can be increased by combing several learning models. This algorithm can be adopted for both classification and regression issues, and this is the major benefit of this algorithm. Random forest has almost similar hyperparameters like a decision tree or a bagging classification model. RF (random forest) makes the model more random with the growth in the number of trees. In order to split a node, this algorithm looks for the optimum feature among a random subset of features rather than looking for the leading feature. This generates outputs with a broad range that usually improves the performance of the model.

3.7. ANN (Artificial Neural Network)

ANN is also named as Artificial Neural Network. In general, this model is used for computation. This model considers the structures and functions of biological neural networks in its work. Also, the information flow affects the configuration of the ANN. Therefore, variations in neural network generally depend on input and output. ANN is generally considered as nonlinear statistical data. This implies that a complicated relationship is defined between input and output. This results in the generation of a variety of patterns. ANN is also known as a neural network. Artificial Neural Networks (ANNs) are completely linked neural networks with multiple layers having a brain-like structure. They are composed of an input layer, numerous hidden layers, and an output layer. Each node in one layer is coupled to each another node in the subsequent layer. The depth of the network can be increased by adding the number of layers that are hidden.

3.8. KNN (k-Nearest Neighbours)

KNN is a simple machine learning algorithm. This algorithm does the storing of all present cases and does the classification of novel cases on the basis of a similarity measure such as distance functions. KNN as a nonparametric algorithm is quite popular for statistical estimation and pattern recognition. In this algorithm, a bulk number of votes of its neighbours classify a case, while the case is allotted to the class most frequent amid its k-nearest neighbours. A distance function is used to measure the k-nearest neighbours. The case is generally allotted to the class of its nearby neighbour when k = 1. Many distance metrics are used to discover the k-nearest data points to the unknown sample. These matrices include the Euclidean distance, Minkowski distance, and Manhattan distance. The Euclidean distance among data points X and Y is signified as

The Manhattan distance among X and Y data tuples is measured as

The Minkowski distance among data tuples X and Y is measured as

4. Literature Review

MengyunQiao et al. [23] recommended a new automatic calcification detection technique. The recommended technique is on the basis of Ultrasound Radio Frequency (RF) signals by fusing quantitative multiparameter [23]. There were mainly four steps included in the recommended technique. The first step was the selection of the Region of Interest (ROI). Extracting numerous attributes on sliding windows across the whole ROI was the second step. The third step included the classification of a window with or without calcifications with the help of the adaptive boosting classification model. In the last step, a threshold filter provided the detection result. A database of 130 skilled doctors confirmed bosom cancers with calcifications which were used for carrying out tests. The recommended technique attained a standard accuracy rate of 88% in contrast to manual details. The tested results depicted that the automated RF signals feature system could support radiologists in the detection of cancer calcifications in a more accurate manner. This system provided more direction for the concluding result.

KosmiaLoizidou et al. introduced a new approach for bosom microcalcification (MC) detection [24]. For this purpose, sequential sequences of digital mammograms were used. Improving the detection accuracy of microcalcification through the subtraction of prior images was the main objective of this work. In this work, a novel dataset, with accurate indication of microcalcification positions was formed particularly for this study. Initially, their commended technique temporarily subtracted mammograms following by demon-based registration. This recommended technique efficiently eliminated the fixed areas and microcalcifications. In the next step, the classification of the microcalcifications was done as cancerous or doubtful with the help of subtracted pictures. The classification task was performed by selecting several different features. The testing of four different classification models was done with leave-one-patient-out cross-validation. These obtained outcomes revealed that it was advantageous to add temporal subtraction to CAD (Computer-Aided Design) systems for providing support to radiologists in the efficient detection of breast microcalcifications.

Shi et al. [25] recommended a completely computerized channel of mammogram image processing. Initially, this channel used a gradient weight map for estimating the skin-air edge [25]. Afterward, this channel used unsupervised pixel-wise labelling without requiring any prelabelled regions for detecting the pectoral-breast edge. At last, a new texture filter was utilized for detecting calcifications within the breast area. Several tests were conducted on the Mammogram Image Analysis Society database. The tested results demonstrated that the recommended technique achieved an accuracy rate of 97.08% and 96.15%, respectively, for bosom edge segmentation and detection of microcalcification. Moreover, somewhat better accuracy rates were attained on Full-Field Digital Mammography image databases. For wide-ranging similarity assessments, computation was made of the Jaccard and Dice indexes among breast areas that are segmented, and the ground truths were incorporated as well. This concept had the ability to provide precious assistance for analyzing mammograms in the hospital.

Hossain [26] introduced a computerized approach and performed the segmentation of microcalcification in the mammogram images. Initially, the image was enhanced by implementing the preprocessing applications of images [26]. Afterward, the segmentation of the breast area was done from the pectoral area. A clustering algorithm called fuzzy C-means had been implemented in this work for detecting the doubtful regions. This clustering algorithm performed the division of doubtful regions into the negative and positive patches. This process eliminated the physical labelling of the required area. The training of an improved U-net segmented network was done using positive patches. These patches contained microcalcification pixels. At last, the trained network was used for microcalcification segmentation areas from the mammogram images in an automatic manner. The recommended technique achieved an average accuracy rate of 98.5%. This accuracy rate proved the superiority of the recommended technique over state-of-the-art techniques.

Warren et al. presented a study for investigating the impact of image processing in the recognition of tumors in images of digital mammography [27]. For this purpose, hologic amorphous selenium detectors were used to gather two hundred seventy pairs of bosom images from eight systems. The image processing type influenced the recognition of clusters of calcifications. In digital mammography, the image processing played a significant role in the identification of calcification clusters as per the achieved tested results. Standard image processing proved its efficiency in the identification of calcification clusters in the context of three image-processing adaptations and the examined system. In the future work, the impact of cancer detection would be considered in the selection of the appropriate type of image processing technique.

Cruz-Bernal et al. [28] recommended a new technique for the identification of calcifications. The recommended technique was constructed on cluster prominence (cp) feature evaluation on mammograms [28]. The cluster prominence feature was analyzed deeply in this work. The results demonstrated that it was extremely discriminative and allowed more modelling of the calcifications in contrast to other features. This approach provided less classification error. False detection was the main error among these misclassifications. This error was called false positives. In contrast to false-negative errors, this error was less serious. The results of comparative analysis depicted that the recommended approach performed better as compared to other approaches. This technique was recommended as an instrument for providing support to radiologists in the diagnosis process. In the future, a new web application would be developed to provide more flexible assistance to the end customer. More tests of this approach would be carried out with the help of an IN breast dataset and a novel description of the cluster prominence (cp) feature to ameliorate the outcomes in terms of time.

Cai et al. modified the end-to-end training of Deep Convolutional Neural Networks (CNN) to distinguish microcalcification for diagnosing breast cancer [29]. The images from two different medical colleges were gathered for this purpose. It was analyzed that feature selection was necessary to perform for deep features using manual attributes. A trained model was utilized for extracting deep features for the comparison of conventional techniques using a similar benchmark. This model demonstrated the capability of deep CNN in the capturing of similar attributes for the manual attributes. The obtained results demonstrated that Convolutional Neural Networks could easily distinguish microcalcification. In addition, the recommended technique validated that the conventional morphological features could play a significant role in directing the features of Convolutional Neural Networks for achieving highly accurate microcalcification classification. Quintanilla-Domínguez et al. proposed an efficient technique for the diagnosis of microcalcifications in digitized mammograms [30]. The recommended approach mainly considered collaboration of the image processing, pattern recognition, and artificial intelligence. There were mainly four phases included in the recommended technique. These faces were identified as image selection, image enhancement, feature extraction, and image segmentation. The feature extraction stage was based on mathematical morphological operations by implementing coordinated logic filters. The image segmentation stage was based on clustering algorithms like k-means and self-organizing maps. At last, a classification model such as an artificial metaplasticity multilayer perceptron was used for the classification purpose. The recommended system constituted an efficient approach for detecting microcalcifications. The tested outcomes revealed that the recommended technique could successfully detect microcalcifications. The recommended system achieved accuracy and specificity of 99.93% and 99.95%, respectively.

Guo et al. summarized the different ultrasound imaging techniques as well as their clinical applications in which the breast cancer patients were managed [31]. Several newly designed and previously proposed techniques were discussed in order to review their performance outcomes. The final sections of this review also compare certain techniques in terms of their outputs achieved for bosom cancer diagnosis. Thus, to perform bosom cancer diagnosis and treatments, integrating the new technologies was considered important. Ultrasonic quantitative analysis was possible through CAD. To diagnose breast ultrasound, a reliable and operator-independent approach was provided here. The review showed that for managing breast cancer patients, the newly designed ultrasound imaging techniques provided certain important tools which proved to be highly beneficial.

5. Problem Formulation

The calcification of breast cancer from the ultrasound image is the major issue of medical image processing. The calcification normally is the early indication of bosom cancer. The calcification of bosom cancer has various steps like image preprocessing, segmentation, textural feature extraction, and classification. In the preprocessing phase, the ultrasound image can be denoised using the filter data. In the second phase, the textual feature analysis algorithm will implement breast cancer calcification. The technique of region-based segmentation called k-means will be implemented. In the last step, the HMM classification algorithm will be implemented for calcification detection. The different objectives of this research study are to evaluate and examine various breast cancer calcification techniques. Further is to design a novel segmentation approach for the breast cancer calcification and also to implement the designed method and contrast it with extant techniques in terms of accuracy and execution time.

6. Research Methodology

This research work relates to calcification of the bosom cancer. The calcification indicates the premature stage of the breast cancer. The phases of calcification detection are described in Figure 2. Detecting the calcium from breast ultrasound image is the purpose of this research work. The entire process of diagnosis requires only a few steps. First, the image must be converted into the gray-scale image where the preprocessing step is being utilized. Furthermore, a region-based segmentation widely known as k-means is used to segment images. Extraction of the textural function is done by applying the GLCM algorithm in the following step. Finally, the term HMM classification is applied, which contributes to calcium prediction. The various steps of the detection of calcium cells are discussed as follows.

Figure 2

Phases of research methodology for the current work.

6.1. Preprocessing Phase

Preprocessing is the intermediate step in performing calcium cell detection. Here, the input image data is transformed into the gray-scale image to perform the latter operations that follow.

6.2. Segmentation

To segment the input signal, a region-based segmentation method named as k-means clustering algorithm is deployed. There are diverse scenarios in which the k-means algorithm, which is a clustering approach centered on partitions, is applied. Use this algorithm to describe every other cluster using the mean also called as centroid of the data points. The principal objective of this algorithm is to partition the n number of items into k number of clusters. The high similarity of the intracluster and the low similarity of the intercluster are considered as the outcome. To evaluate the resemblance, the mean value of the objects accessible within a cluster is set. Primary centroids of the cluster k are chosen in a randomized order. On the basis of the Euclidean distance among the centroid cluster and the target, by this approach, the objects in clusters are moved iteratively by moving them from one cluster to another. After this redistribution of objects in clusters, for each cluster, the value of the mean of the objects is again determined. This effect provides a vital centroid cluster. This method is repeatedly put in before the theory convergence function. Usually, the square error principal is used. These criteria can be interpreted as

This function defines a group of objects of m size for each object. The mean of the cluster which is defined by the k implies partitioning such n objects into k groups, in order to decrease the condition of the square error. The condition tends to compress the clusters of k. The basic method of k-means clustering is demonstrated by the following algorithm:

k random centroids are picked to build the very first cluster from the input data

For every data point, the distance from all centroids is determined. Its membership shall be attributed to the nearest centroid

For recalculation of the new cluster centroids, use the sum of the all-data points allocated to the cluster

Repeat the second step before convergence

6.3. Feature Extraction

In the third step, the section of interest is selected from the segmented image. In this work, the GLCM (gray-level cooccurrence matrix) algorithm is used to carry out the study of textural characteristics, which can be used to statistically test texture. The texture of the image is distinguished by the GLCM algorithm rule. The defined spatial relation is determined for the development of a GLCM algorithm along with the pixel pair derivation in an image. In addition, in the following step, the statistical measures are obtained from the created matrix. In MATLAB, the gray comatrix function is initiated using the GLCM. This algorithm allows for the possibility of an intensity value i to happen in a particular spatial relation to a pixel with the j value. The pixel of interest and the pixel to its immediate right are allocated by default via the spatial relation. The spatial relationships between two pixels may therefore be defined.

6.4. Classification

In the last step, the calcium cell prediction is built on the function derived with the GLCM algorithm, using the HMM classification algorithm. Among many classification algorithms, HMM is one of the most common algorithms. The HMM algorithm can be quickly applied, and results are satisfactory. In addition, this algorithm needs no contextual knowledge about the data set for classification purposes. In this algorithm, the classification depends upon resemblance. This algorithm searches for the closest k-sample for a novel sample classification. In HMM, the process of classification initiates with a collection of data. The data set contains a number of characteristics which define the collection of data. There are two separate sets that are originated by dividing both the training data sets and the test sets. Using the training set, the data is inured to the proposed algorithm, and the accuracy of the algorithm is determined via the test sets. A related collection of training data is used by the HMM classifier for the classification of novel samples. For this cause, HMM is also known a local learner. There is no learning process of this algorithm. This algorithm holds all the samples of the training allocated to it as data. The measurements are conducted during the analysis of a research sample. The training sample inside the HMM algorithm is defined as data points set accessible in the n-dimensional space. When classifying the unknown samples, the nearest k data points are found in the n dimension space. Various distance metrics are used to classify the k-nearest data points.

7. Results and Discussion

In this research paper, we have implemented the idea in MATLAB software. This software is generally utilized for model simulation. MATLAB includes numerous toolboxes, such as computer vision, neural networks, and classification. MATLAB uses different kinds of programming languages such as C, C++, and python. As shown in Figure 3, default interface is shown for the breast cancer calcification. The calcification process detects the calcium cells from the ultrasound images. As shown in Figure 4, the ultrasound image is taken as input on which we need to perform calcification process detection. The further image can be preprocessed to convert it into the gray scale. As shown in Figure 5, the input image can be converted into gray scale. The technique of the median filter is applied which can increase the quality of the image. The technique of k-means segmentation is applied which segments the image. As shown in Figure 6, the classification is the last part which detects calcification from the input image. The calcification process detects the calcium cells using the technique of HMM.

Figure 3

Default interface of the proposed system.

Figure 4

Input images of the proposed system.

Figure 5

Filtering technique of the proposed system.

Figure 6

Classification process of the proposed system.

7.1. Result Analysis

7.1.1. Accuracy

The classification accuracy i for a specified program may be explicated as the ratio of total patterns characterized precisely to the total number of patterns. The following is the equation:

The total number of accurately classified pattern cases is represented by variable “t” in the above formula. In this equation, “n” defines the total number of sample sets.

7.1.2. Execution Time

The execution time is represented as the difference between the end time and the start time.

The end time represents that time when the algorithm stops executing while the start time represents the beginning time of execution of algorithms.

7.1.3. Precision

Precision is represented as the total number of extracted positive events more than the total number of cases declared positive using classification methods.

7.1.4. Recall

Recall value is illustrated in the context of the total number of true-positive cases extracted more than the total number of positive cases in the set.

7.1.5. F1 Score

The prejudiced standard of precision and recall represents the F1 score.

Table 1 shows the accuracy comparison, and Table 2 shows the execution time comparison.

Table 1

Accuracy comparison (in percentage).

Image no.	Multiclass SVM (%)	HMM (%)
1	95	96
2	96	97.2
3	96.5	97.56
4	96	96.5
5	97	97.8

Table 2

Execution time comparison (in seconds).

Image	No multiclass SV	M (%) HMM (%)
1	2.23	1.23
2	2.78	1.22
3	1.22	0.98
4	2.89	1.34
5	2.78	1.45

Figure 7 displays the improvement in the accuracy of the proposed technique than the existing algorithm. The proposed algorithm showed the accuracy level of 97%. At the same time, the existing algorithm achieved 97% accuracy rate.

Figure 7

Accuracy comparison of the proposed system.

Figure 8 depicts that the proposed and existing algorithms contrast in terms of execution time. The proposed method utilizes less execution time than the existing algorithm.

Figure 8

Execution time comparison of the proposed system.

8. Conclusion

Image processing is a method that is used to extract and process the pixel-shaped data that is included inside representations of objects. For the purpose of identifying the calcification cells, a machine learning method is used in this study. Several procedures must be completed in order to do the calcification analysis at the same time. The clustering method k-means is used to conduct segmentation based on area, and it is applied. It is necessary to apply the k-means technique to calculate the Euclidean distance between the center points and to identify where they are located. The points with the same distance between them are clustered together in a cluster. Remove the textural features of a picture using a function extraction technique known as the GLCM algorithm, which is a function extraction method. Using the GLCM method, the 13 textural characteristics of the input image are removed from the picture. After that, the HMM classifier is utilized to perform classification in the last phase. The HMM classifier takes advantage of the gap in order to carry out a classification. The identification of calcium cells is used to describe the performance of this classification system. The suggested HMM-based methods and existing SVM-based strategies for illness prediction are compared in terms of the degree of accuracy with which they can forecast disease. According to the results, the degree of accuracy has improved by up to 10% when compared to the previously established norm. The usage of the HMM classification system for calcium cell identification is one of the potential methods that may underpin the extension of this study, and the technique that has been proposed can be further improved. The suggested approach may be compared to existing calcium cell detection and categorization techniques that have been developed.