Breast Cancer Detection using Mammography: Image Processing to Deep Learning

1. Introduction

Globally, breast cancer ranks as the most prevalent cancer and holds the second-highest mortality rate among women internationally. According to the WHO, half a million women worldwide die each year from breast cancer [1]. In 2018, breast cancer claimed the lives of 627,000 women, making up 15% of all cancer-related deaths. In Europe, there were 3.9 million new cases reported in 2018 – accounting for 25% of the world's cancer toll [2]. This research reviews breast cancer detection through mammography over some time, aligning with several UN Sustainable Development Goals (UDGs) [3] by improving early detection (goal 3), optimal health and well-being (goal 5), contributing to industry, innovation, and infrastructure through technological innovation and collaboration (goal 9) and fostering partnerships for collective action (goal 17).

Statistically, the increase in the death rate from breast cancer can be attenuated by adopting routine screening procedures at its early stages. A reliable technique for breast cancer screening is mammography. A mammogram can often reveal breast changes that could represent cancer years before physical symptoms develop. The goal of screening programs such as yearly mammography is to promote early cancer detection while the cancer is still clinically occult. Ductal carcinoma in situ (DCIS) is the earliest form of breast cancer, as it is confined to the ductal system and is not yet invasive. Over 90% of non-palpable DCIS is detected by microcalcifications alone [4]. Therefore, differentiating benign from malignant calcifications based on their radiographic appearance is very important [5]. Due to their small size (0.1-1.0 mm) and lack of background contrast on images, malignant calcifications have the potential to be difficult to observe and characterize [6]. Accurately identifying and classifying such calcifications on screening mammograms can be challenging. Previous computer-aided detection (CAD) systems were disappointing, yielding reduced accuracy, increased biopsies, and no definite improvement in cancer detection rate [7].

Optimizing the ability to differentiate benign from malignant calcifications is an area of great interest in Artificial Intelligence (AI) through Deep Learning (DL) [8]. DCIS often has a natural history of non-progression, with an estimated up to about 40% progressing to invasive carcinoma, according to some investigators [9]. Others have estimated the progression of DCIS to invasive cancer to be about 53% [10]. As DCIS is a pre-invasive and a non-obligate precursor to invasive carcinoma, with approximately half of the cases remaining non-invasive, greater emphasis is placed on detecting early invasive breast cancer. While there are many potential mammographic features associated with invasive breast cancer at screening mammography, the most common mammographic appearance is that of a mass [11]. The goal of mammography in detecting invasive breast cancer is early detection when it is 1 cm or smaller, with a 95% chance of 10-year survival compared to 85% and 60% survival at 1-2 cm and 2-5 cm, respectively [12]. Morphological features describing a mammographic mass's shape and margins help differentiate benign from malignant masses. These features have been exploited in DL using CNN for shape classification, yielding an accuracy of 80% on a public dataset [13].

Mammography uses low-dose X-rays to approximate the distribution of absorption coefficients depicted on imaging as structural density variation, revealing anatomy based on radiomic features used to translate/detect the presence of breast cancer. In this context, this review article uses large publicly available X-ray mammography datasets from different time frames to perform a detailed summary of breast cancer diagnosis methodology. Although previous studies have examined similar literature, most cover a specific range. They must be revised to examine a broad spectrum of methodologies over a considerable period. Examining the consistent improvements in technology and image quality can provide a better understanding of how best to detect and classify lesions with greater accuracy and efficiency.

Analyzing a mass manually is a time-intensive task, with the effectiveness of contrast enhancement during screening relying on both image quality and the radiologist's skill level. Screening techniques that are used include clinical and self-examination (CSE) of the breast, mammography, breast ultrasound, and magnetic resonance imaging (MRI) [1]. CSE is a good approach where limited clinical resources are available with high sensitivity (60-70%) and specificity (90-97%) [14,15]. Among US women, mammography screening has reduced mortality [16]. Although it has a positive impact, screening mammography has many false positives [17]. In addition, CSE performance varies based on the physician's experience and knowledge. This issue can be resolved with multiple clinical exams, but this approach is costly, arduous, and time-consuming.

Since 1990, researchers developed CAD tools to improve screening accuracy. While computer-aided mammography (CAM) has been in use for over two decades, the automatic detection of microcalcification still needs to be improved, mainly due to its low contrast, fuzzy nature, and poor discrimination from the contiguous pixels. Researchers have been using a variety of algorithms for the automatic diagnosis of microcalcification. Different techniques and methods are used for feature extraction and enhancement, such as wavelet transforms, segmentation, and clustering. The medical community is successfully applying advanced AI machine learning (ML) techniques, to improve accuracy in cancer screening procedures [18]. In this context, unsupervised learning methods have successfully aided radiologists in improving screening outcomes [19,20,21,22,23,24]. ML techniques have allowed for the detailed analysis of mammography images to be performed using feature extraction techniques like textural, statistical, and sub-band transformation, in addition to other traditional feature extraction techniques. These features tune classical ML techniques to classify unknown images into benign and malignant cases. These DL strategies represent learning-based techniques leading to automatic feature extraction for categorizing mammography images.

There are a few prior review articles on mammography and ML. In 2017, Abdelhafiz et al. published a review of 83 studies applying CNNs to mammography. Their review spanned a period of 1995 to 2017 and DL techniques for lesion localization and detection, risk assessment, image retrieval, high resolution reconstruction, and classification tasks. Both public and private datasets were included [25]. In 2019, Gardezi et al. reviewed DL techniques for mammogram datasets spanning a 5-year period, which discussed breast density estimation, mass detection, and mass classification, demonstrating how DL can contribute to the diagnostic performance of CAD systems [26]. In 2019, Geras et al. performed a review comparing conventional CAD, which uses prompts to demonstrate possible cancers, to AI-based CAD. Their review discusses the advantages of CAD system, which include fewer false positives and sensitivity equal to radiologists, allowing its use as a double reader and potentially as an independent first reader in the dismissal of normal cases [27]. More recently, Hickman et al. [28] published a meta-analysis of 15 studies in 2023 amalgamating 185252 cases read by ML algorithms versus human readers. The overall area under the curve for the receiver operating characteristic AUC (ROC) curve for ML was 0.89, and for the human readers was 0.85. They concluded that ML algorithms as stand-alone applications in screening mammography can exceed the performance of human readers in detecting breast cancer and enhance efficiency. Similarly, in a recent work, Khan et al. [29] carried out a systematic review of representation learning-based methods using whole-breast MRI scans without expert radiologists’ help to predict pathological complete response to neoadjuvant chemotherapy (NAC) in breast cancer. The immediate effects included reducing the toxicity of unnecessary NAC, with improved clinical outcomes.

Our review is on breast cancer diagnosis and detection in mammography from 1970 to 2023. We focused on the evolution of techniques over this period chronologically, including image preprocessing and enhancement methods, feature extraction methods with ML, and DL methods applied to mammography datasets. The value of this presentation is that it is oriented to professionals in the computing segment of healthcare, reaching beyond radiologists and clinicians. The main contributions of this article are:

• This work encompasses mammography based on early diagnosis and breast cancer detection, highlighting the significant achievements from 1970 to 2023 using three keywords, namely mammography, microcalcification, and breast tumor.
• The notion is to divide the work into three main basic strategies for early breast cancer detection: image processing-based techniques, ML-based solutions, and DL algorithms.
• Competing algorithms have been briefly discussed in the field of mammography.
• All datasets of mammography have been discussed.
• Recent state-of-the-art techniques have been compared and discussed.

This review is conducted carefully, following the preferred reporting items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. The PRISMA flow diagram illustrating the salient steps of the electronic literature collection process is shown in Figure 1. The search was carried out in 20, 16, 10, and 8-year intervals based on the keywords illustrated in Table 1. The last column shows the highest-citation articles in different periods for manual and computer-aided techniques. The compilation and analysis of mammography data involved an extensive online search of published literature from the Macquarie University library. The electronic bibliographic databases that have been used include Macquarie University library, Semantic Scholar, PubMed, Google Scholar, the Biological Science database, Springer Journals Complete - Open Access, the SciTech Premium Collection, Taylor & Francis Open Access, the Catalog of Open Access Journals, Wiley-Blackwell Open Access Titles, the Engineering Database, Wiley Open Content, the Computer Science Database, and IEEE Open Access Journals and Conferences. Shahzad Ahmad Qureshi and Aziz-ul-Rehman initially conducted the literature search, with subsequent reviewers independently assessing the article’s relevance based on their titles and abstracts relevant to the study.

The three main themes, manual-aided, ML, and DL are equally crucial for the early diagnosis problem. Also, the manual-aided articles (conventional image processing techniques) were the most cited due to their relatively longer presence in research databases. The converse is true for DL-based articles, which are more recent and have needed more time to accrue citations. We have concisely reviewed primitive to recent methodologies and techniques, including novel DL architectures-based ideas and approaches related to mammography for micro-calcification detection in the breast.

2. Materials and Methods

This study highlights explicitly the techniques used to detect abnormalities in breast cancer mammograms spanning 1970-2023 to provide researchers with a quick overview of employed methods. A basic breast image classification scheme is illustrated in Figure 2. We discussed the methods to detect breast cancer via mammography using (a) image preprocessing methods, including image enhancement methods, (b) different feature extraction methods with conventional ML approaches, and (c) DL-based convolutional neural network (CNN) methods. These methods are applied to different breast mammography datasets.

2.1. Preprocessing

Preprocessing is a complex task that involves cleaning, transforming, and preparing the dataset for further analysis. The quality and effectiveness of data-driven tasks can be significantly affected by various challenges, including dealing with missing values, scale dependency, outliers, feature selection, correlation removal, redundant information removal, class imbalance, data noise impregnation issues, and large datasets with space and time complexities. Each of these challenges necessitates careful consideration and domain-specific knowledge to select the most appropriate solution for the task at hand.

2.2. Image Processing Methods

Mammographic images, whether analyzed through self-examination or intelligent algorithms, can be challenging to interpret. The focus is primarily on the contrast between malignant and normal tissues, with tumors or microcalcifications appearing as white or dense regions. Enhancing these images to detect microcalcification involves a range of direct and indirect techniques. The complexity of this task is underscored by the following challenges:

Direct image enhancement means improving the perceptual quality of the image by establishing correct discriminative contrast using direct techniques. Rangayyan et al. [40] suggested the adaptive neighborhood technique, in which the local contrast of an image is computed. Tang et al. [41] proposed a technique called multiscale local contrast, which decomposes contrast using a wavelet domain, improving contrast details in different scales.

In the indirect contrast enhancement technique, the image histogram is modified instead of directly manipulating the image contrast. The most popular indirect image contrast enhancement techniques are based on histogram modeling. Significant variants of this technique are: In the histogram equalization (HE) technique, grey levels are distributed equally in the output, whereas input image mapping is proportional to the cumulative intensity [42]. This technique enhances the contrast of images globally, and the abnormal regions are filled close to contrast values. Therefore, an increase in the lower local contrast rises to higher contrast levels, and unnatural artifacts are seen in the output image. An adaptive histogram computes the difference in histograms against distinct regions of an input image. Multiple histograms are used to reallocate the contrast values of an image so that local contrast and edges can be enhanced in each region. In the adaptive and contrast limited adaptive histogram equalization (ACLAHE), an image's enhancement amount is equalized by computing the histogram of sub-matrices and truncated to limit the enhancement gray levels [43]. In the minimum mean brightness error bi-histogram equalization (MMBEBHE) technique, the image is divided into two parts depending on a threshold value. The threshold is calculated by finding all possible separated ranges of intensity values in an image and then finding the Absolute Mean Brightness Error (AMBE) followed by HE. In the case of an inhomogeneous background of a mammogram, local-based enhancement techniques can perform better. Such a popular technique is fixed-neighborhood statistical enhancement [44]. It estimates the background of an input image by computing the statistical properties of a pixel in the neighborhood and suppresses the estimated background. As a result, an increase in the local contrast of an image occurs.

2.3. Classfiication

Feature extraction in clinical images is an important pre-processing step for ML and DL algorithms. It helps reduce redundant data and an algorithm's learning time. This section highlights feature extraction methods, followed by a brief overview of some of the most prominent methods proposed in the past.

2.3.1. Feature Extraction for ML

Feature extraction reduces the dataset's dimensionality by condensing it into smaller subsets for processing. An undesirable effect of a large dataset is the buildup of many variables, requiring much computational power. Various feature extraction methods are illustrated in Figure 3.

Several researchers suggested that different local features in mammographic images can be extracted for classification. Zhang et al. [45] divided local mammogram features into two classes, i.e., spatial (average and standard deviation intensities of foreground and background) and morphological features (moment, compactness, and Fourier descriptor). Spatial, morphology, and cluster description features were used to designate mammogram clusters. Davies et al. [46] proposed a technique using a few local mammographic features (mean intensity level, area, the ratio based on the area to the squared maximum-linear dimension, the strength of edge, and shape parameter).

2.3.1.1. Textual Feature Extraction

Textural features are repeated or identical arrangements of patterns of pixels’ intensities that refer to characteristics of an object, such as its appearance, shape, size, density, coarseness, and contrast. Different techniques for extracting texture-based features are the surrounding region dependence method [47], spatial gray-level dependence method [48], gray-level run-length method [49], and gray-level difference method [50].

2.3.1.2. Intensity-based Feature Extraction

It consists of first-order statistics using features like mean, mode, variance, and standard deviation to find the intensity and its variation [51]. Here, a large set of sampling features can be extracted from a pixel's local neighborhood to train learning algorithms. However, the spatial arrangement of a pixel, a key factor for detecting and classifying abnormal tissues in mammograms, needs to be considered.

2.3.1.3. Multi-scale Feature Extraction

The disadvantage of using texture-based features is that all methods focus on the single scale of a pixel. To extract features, Laws [52] has explained a system in which linear transformation and energy computation perform better than the texture features of second-order statistics. Over time, many researchers have proposed improved methods for better performance by domain transformation, like wavelet transform-based methods [53,54,55], Gabor Filter Bank Methods [56], and Laplacian of Gaussian Method [57].

2.3.2. ML Approaches

Extracting the most discriminative feature set is an essential factor in classification [58]. The latest ML algorithms can detect and classify normal and cancerous tissues (benign or malignant). A brief overview of important classifiers is given by:

• k-Nearest Neighbor (k-NN) Classifier

It is used to categorize benign, malignant, and normal tissues. It finds the distance between unobserved data and all the known data in $k$-classes, and the unknown instance is given the class name depending on the shortest distance from the known data classes [59]. The Minkowski distance, d_M, is a generalized form for the Euclidean- (L1-norm) and the Manhattan-(L2-norm) distances that are used with the k-NN classifier, as given by:

$$d_M={\left({\sum }_{i=k}^k\left|{x_i-y_i}^q\right|\right)}^{1/q}$$

(1)

The number of features is $k$, for object $x$ the value of feature is $x_i$ whereas for object $y$ the value of feature is $y_i$. The value of q defines the Manhattan and Euclidean metrics as one and two respectively.

• Decision Tree Classifier

In diagnosing breast cancer, a decision tree (DT) [60], composed of decision and leaf nodes, is a commonly used classifier because it is easy to define and map the rules from the root to a leaf node. The entropy or information gain computes the homogeneity in the dataset given by:

$$E\left(S\right)=-{\sum }_{x\in X}P\left(i\right)\times \log P\left(i\right)$$

(2)

The dataset is represented by$\mathrm{S}$, $\mathrm{X}$represents class cardinality and ith class probability is $\mathrm{P}\left(\mathrm{i}\right)$. Zero entropy means perfect classification.

• Naïve Bayes (NB) Classifier

The NB is a classifier that categorizes the data statistically by considering all features. Given the feature set$X=x_1+x_2+x_3$, NB computes the posterior probability (probability to be assigned to the unknown instance) of class$c$ as given by. The probability density function of class c is represented by$P\left(x_i\right|c)$, whereas the class prior is given by $P\left(c\right)$. NB performs well for large datasets, and sometimes, due to the feature set, it outperforms other classifiers [61]. The problem with NB is zero frequency, which occurs when the class and feature combination frequency becomes zero.

• Support Vector Machine

The principle attribute of SVM is maximum marginalization [62], which classifies the malignancy of breast cancer. SVM is a linear classifier that constructs a linear hyper-plane of $\mathrm{N}$-dimensional space to categorize the linear data by maximizing the margin between two classes. Non-linear mapping functions can be used to link the data (source) to a higher dimensional feature space (target) if the data is non-linear. Training data is used to compute the weighted vector ${\mathrm{w}=[{\mathsf{\omega }}_1+{\mathsf{\omega }}_2+{\mathsf{\omega }}_3+\cdots ]}^{\mathrm{T}}$ and the bias b for discrimination boundary is given by,

$$d\left(X,w,b\right)={\sum }_{i=1}^n{w}_i{x}_i+b$$

(3)

Where $\mathrm{d}\left(\mathrm{X},\mathrm{w},\mathrm{b}\right)$ is the decision function that is optimal when$\mathrm{d}\left(\mathrm{X},\mathrm{w},\mathrm{b}\right)=0$. For non-linear data, it is mapped to a high dimensional space using a non-linear linking function as given by,

$$K\left(x_i,x_j\right)=\phi {\left(x_i\right)}^T\phi \left(x_j\right)$$

(4)

, where $\mathrm{K}({\mathrm{x}}_{\mathrm{i}},{\mathrm{x}}_{\mathrm{j}})$ is the kernel function that maps the non-linear mapping function$\mathsf{\phi }\left(\mathrm{x}\right)$. Different studies suggested changing the dataset and adding more constraints (hard margin), the SVM produced good results on mammographic images.

2.3.3. Deep Learning Approaches

DL is a sub-domain of ML and is the basis of many of the approaches recently reported by many investigators. The basic idea relates to the boosted hidden layers used in shallow neural networks by many of the other time-to-time solutions exploited for the early diagnosis of breast mass abnormality.

DL has an edge over ML due to its self-feature extraction capability (representation learning), significantly improving breast cancer diagnosis [63]. A deep neural network (DNN) is constituted of “n” layers and “m” hidden neurons. Weights are associated with each input and output link in the neural network. Regulating weights in training helps neural networks learn significant information. The information flows in the learning phase to make it a self-learning network [64]. We have discussed numerous DL-based methods for breast cancer classification and segmentation that use the variants of CNN belonging to one of the three categories (Figure 1): (a) Novel CNNs express impressive results in mammography detection and classification, (b) Transfer Learning (TL) consists of developing a model to use the preset parameters, with salient variants to make the training compatible after adjusting the input size equal to the size of the mammogram with several neurons in the concluding layer set equal to the class-cardinality of the target database, (c) DL Features Extraction solutions are based on dynamic features extraction using the DL models and employed further by conventional ML algorithms to carry out classification tasks. The TL models can also be used for deep feature extraction using customized datasets. An overview of salient DL methods by different authors is given by:

• Bayesian Neural Network

In DL, the Bayesian neural network (BNN) extracts the uncertainties using the Bayesian model, and it refers to the uncertainty of the model output [65]. Specht [66] proposed a Probabilistic Neural Network (PNN) based on the Bayesian classification. PNN consists of three layers (Input, Radial Basis, and Competitive layer), described in the study. PNN classified the mammograms into three classes (normal, benign, and malignant) based on the input vector. The feature vector was extracted from a mammogram through Discrete Wavelet Transform (DWT) and then applied classification. Seventy-five (75) images were selected for the experimental purpose, and the accuracy achieved by the authors was 90%.

• Back Propagation in DL

A back propagation neural network (BPNN) is a supervised learning network called a feed-forward neural network with backward propagation of error to adjust the neurons’ weights. This neural network consists of an input layer, one or two hidden layers (no hard rule for layers’ selection), and, in the end, an output layer. In the BPNN, during the training phase, the features are fed to the neurons for learning. Then, the algorithm fine-tunes the weights based on the error rate (i.e., loss function) from the previous epochs to converge quickly in the backward direction. Zhang et al. [67] proposed a neural network that reduced the false positive rate of microcalcification detection in real cases of 100 patients’ mammograms. It consists of a three-stage method. In the first stage, all the microcalcifications were detected. In the second stage, FP detection was removed from the first stage’s output. In the third stage, the BPNN with Kalman filter was used for the classification of microcalcification in mammograms.

• Convolutional Neural Network

CNN is a sub-category of deep neural networks that achieves impressive results in mammography for detecting and classifying microcalcifications and is currently the most popular neural network [68]. The classification decision is taken in the decision layer by computing the prediction error or loss function. The first CNN was developed by Fukushima et al., known as “Recognition,” and was the very first neural network model used for medical image analysis [69]. The class with minimum loss is declared the classifier decision. Sahiner et al. [70] proposed a backpropagation CNN to classify mammograms. Sub-regions, regions-of-interest (ROIs), are extracted for CNN input. ROI patches are extracted by either averaging, subsampling, or texture feature extraction methods. Lo et al. [71] introduced the Multiple Circular Path Convolutional Neural Network (MCP-CNN), which initially gathers information from the identified regions in mammograms and subsequently processes it as features using CNN. Fonseca et al. [72] used an SVM-CNN classifier for breast cancer categorization. Su et al. [73] suggested a method for classifying the breast cancer named “Fast Scanning CNN,” where the algorithm computed the pixel-wise image segmentation to remove the redundant information (which increased the original CNN complexity). Jiao et al. [74] presented a method that utilized deep features for breast mass classification. The method combined the intensity-based features with deep features extracted by CNN from the original image and was finally used for classification.

Arevalo et al. [75] devised a hybrid approach where they employed CNNs to learn data representations in a supervised manner rather than extracting feature maps directly from mammographic images. Rezaeilouyeh et al. [76] presented a microscopic breast cancer classification model based on CNN. In this model, the Shearlet Transform (ST) was applied to images to get the feature vector of Shearlet coefficients fed to CNN for classification. Jadoon et al. [77] developed two deep neural network-based classification models for normal, benign, and malignant classes using CNN-Discrete Wavelet and CNN-Curvelet transform. The features extracted are fused and fed to the CNN for the classification task. Jaffar [78] introduced a method in which the mammograms are first enhanced; then, the CNN is used for feature extraction, followed by SVM-based classification. The outcome of their experimentation is discussed in the forthcoming section.

• Regions with CNN (R-CNN)

Zhu et al. [79] used a fully convolutional network (FCN) to segment the masses in mammograms using a Conditional Random Field (CRF) model. By leveraging prior positional information, the method empirically estimates the ROIs, enhancing the accuracy of ROI predictions. Wang et al. [80] introduced a hybrid method for classifying benign and malignant breast cancer, which involves cropping breast masses and extracting clinical features from multi-view patches of mammograms. CNN was employed to concentrate on regions associated with semantic-based lesions. Gastounioti et al. [81] explained an ensemble method for breast cancer classification. They used lattice-based techniques for textural feature maps and fed them to CNN for multi-class categorization.

R-CNN is a CNN-based network that detects and classifies data on a regional basis regardless of the class. For this reason, it is also referred to as a Region Proposal Network (RPN). R-CNN reduced the convolution time by selecting a region, which is considered an advantage of using R-CNN. Ribli et al. [15] explained a Faster R-CNN method used for classifying the breast cancer. Faster R-CNN used the ROI pooling technique to extract features from the ROI fed to the VGG-16. This method produced bounding boxes accompanied by a confidence score classifying the benign or malignant cancer. Chiao et al. [82] proposed an advanced variant of RPN called Mask R-CNN, which is used to detect and segment mammographic images. Instead of extracting features from the ROI pooling method, the Mask R-CNN employed ROI alignment (ROI-Align) procedure. ROI-Align feature extraction tackled the spatial information loss encountered in Faster R-CNN. Lastly, CNN was used for detection and classification tasks.

• Long Short-Term Memory (LSTM) Neural Network

LSTM is a Recurrent Neural Network (RNN) type that can learn from the reference point instead of from scratch (i.e., error feedback to the input). The reference point is any middle layer whose output can be used as input. As in CNN, the learning must be started from scratch. Gradient vanishing is the major problem in RNN, which is overcome by LSTM, as proposed by Hochreiter et al. [83]. Nahid et al. [84] used LSTM to classify microcalcifications and the formation of masses. They transformed images into 1-D vector format and then converted them into time-series data, followed by LSTM training. The maximum accuracy of 84.4% was achieved when using Softmax at the decision layer.

3. Results

Breast cancer detection has been a conventional problem in medical science since the 16th century. It is still considered the deadliest disease due to its complex morphological structure. In the early years, breast cancer diagnosis included a self-examination method, which produced many false results. With the advent of mammograms, cancer detection based on imaging became feasible. However, microcalcification or abnormal tissues are very subtle and unstable, making it hard for experienced professionals to detect cancerous tissues early. In recent years, image processing techniques have been replaced by computer-aided techniques that help radiologists and experts make more reliable decisions; intelligent algorithms further add classification (to classify cancerous types: benign or malignant), and these techniques produce state-of-the-art, accurate, and more reliable results.

3.1. Datasets for Mammography

Mammography datasets play a vital role in designing and developing AI-based solutions, thereby attributing confidence in the scope of the results along with the robustness of the model. Table 2 illustrates large datasets used for conventional machine and DL techniques, along with the corresponding references, total instances, and categories used for classification. The Mammographic Image Analysis Society (MIAS) dataset is one of the pioneering datasets in mammography. It contains 322 mammographic images (digitized at 50 microns pixel edge) [85]. These images are labeled as benign, malignant, and normal. The MIAS dataset has been extensively utilized for developing and validating CAD systems and algorithms to classify breast lesions. It is a widely cited benchmark data set using mammographic images for breast cancer detection research.

Similarly, a popular MIAS method is the Mini-MIAS data set [86] (http://peipa.essex.ac.uk/info/mias.html). The Mini-MIAS is a well-known case study in breast screening. It contains 322 mammograms reduced to 200 μm pixel edges and cropped/padded so that each image is 1024 × 1024 pixels. These mammograms are listed as safe and sensitive, making them valuable for training and experimental computer-aided detection programs. The dataset has been widely used to develop and evaluate algorithms for breast cancer detection and classification, which is a benchmark in medical image analysis

DDSM (Digital Database for Screening Mammography) is a comprehensive database specially developed for research in mammographic image analysis. It contains a collection of 2620 scanned film mammography examinations labeled normal, benign, and malignant, as well as associated clinical data and ground truth descriptions [87] (http://www.eng.usf. edu/cvprg/Mammography /Database.html). It is a valuable resource for developing advanced image processing techniques such as detection and classification algorithms for breast cancer detection Due to its size, diversity, and fine resolution, the DDSM dataset is widely used in academia and engineering in a research environment. The Curated Breast Imaging Subset of the DDSM (CBIS-DDSM), a standard for performance and accuracy, is a curated dataset designed for breast cancer detection and diagnostic screening. The CBIS-DDSM data set consisted of 10239 standard mammograms to facilitate the creation and evaluation of computer screening algorithms, and other image analysis techniques developed for mammography [88] (https://wiki.cancer image database.net/display/Public/CBIS-DDSM). Due to its comprehensive description, image quality, and focus on analysis and diagnostic information, the CBIS-DDSM dataset has become a widely used benchmark in mammographic image analysis. The IRMA (Image Retrieval in Medical Applications) dataset from RWTH Aachen University in Germany serves as a collection of mammogram patches designed to assess the accuracy of mammogram patch classification methods [89]. It includes datasets from MIAS, DDSM, Lawrence Livermore National Laboratory (LLNL), and regular images from RWTH Aachen. The dataset provides detailed information about the images, classifying them based on background nodes and the nature of abnormalities detected in the mammogram patch. The mammogram patches in the dataset are standardized as 128 × 128 pixels, with 931 normal, and 584 abnormal images.

The BancoWeb LAPIMO dataset is a mammographic image dataset specially developed for research purposes in breast cancer detection [90] (http://lapimo.sel.eesc.usp.br/bancoweb). Based on 1400 images from 320 patients, this dataset is designed to support the development of ML/ DL algorithms and other image analysis techniques tailored for mammography. The digitized mammograms are normal, benign, and malignant, with detailed annotations and ground truth labels indicating the presence of various types of lesions, such as masses and calcifications. The INBreast is a mammographic image dataset designed to support research in breast cancer detection and diagnosis [91]. It was created by the Faculty of Medicine of the University of Porto, Portugal, and consists of a collection of 410 mammograms obtained from both screening and diagnostic examinations (https://www.kaggle.com/datasets/ramanathansp20/inbreast-dataset). This dataset includes images with numerous breast abnormalities, such as masses, calcifications, and asymmetries, along with detailed annotations and ground truth labels provided by expert radiologists. Over and above, the King Abdulaziz University Breast Cancer Mammographic Dataset (KAU-BCMD ) [92]

(https://www.kaggle.com/datasets/asmaasaad/king-abdulaziz-university-mammogram-dataset)is a mammographic image dataset designed to support research in breast cancer detection and diagnosis. This dataset consists of a collection of digital mammograms accompanied by detailed annotations and ground truth labels provided by three radiologists. The ordinal categorization of cancer is based on the breast imaging reporting and data (BI-RAD) classification system. The VinDr-Mammo, a Vietnamese dataset of digital mammography developed by VinBigData, is a research initiative by Vingroup JSC aiming to advance healthcare through ML [93] (https://doi.org/10.13026/br2v-7517). This dataset comprises an extensive collection of digital mammograms, categorized on the ordinal states of the BI-RAD classification system for extensive lesion-level annotations and damage-level assessment, with detailed annotations and ground truth labels provided by the three expert radiologists. The acronyms for some recent DL techniques are illustrated in Table 3 with the dataset used and the method adopted.

3.2. Image Processing

The subsection describes some image processing-based studies used for breast cancer analysis by mammography. In this study, salient image-based techniques are briefly reviewed in the initial part, where prominent algorithms are surveyed. In this context, Verbeek et al. [30] analyzed data according to the age groups of women, where the odds ratio of 0.48 (confidence interval= 95%) was used. Further, the study was conducted for unscreened objects as well. The fatality rates of both, screened/unscreened subjects, were then compared based on breast cancer. Rangayyan et al. [40] employed the Adaptive Neighborhood Contrast Enhancement (ANCE) method, represented the digitization down-sampling to an adequate pixel size and consequently degraded images affecting the radiologist’s decision. The use of the ANCE technique assisted the radiologist’s performance. The area parameter of an “enhanced mammogram” was found to be 0.6745 which was higher than the digitized and original mammograms. In another study, Carney et al. [32] correlated Hormone Replacement Therapy (HRT) with age and microcalcification density, and its role in screening accuracy performance. By adjusting breast density ranges between (62.9-87.0)% from highly dense to fatty breasts, they found that the sensitivity increased from 68.6% to 83.3% with age. Matsubara et al. [99] used an adaptive thresholding scheme dividing the mammogram so that the tissues in it are divided into one of the three classes based on histogram analysis ranging from fatty to dense classes. The affected regions containing potential masses are detected by using multiple threshold values. Dominguez and Nandi [100] based segmentation on the conversion of mammograms to binary images at selected threshold levels. They used 30 gray levels with a minimum step of 0.025 with 8-bit gray levels between 0 and 1. The segmentation sensitivity was 80% with the proposed method.

Zheng et al. [101] introduced a region growth mechanism for initial boundary conditions of adaptive contouring of the mass region in the mammogram. The final contour of mass was achieved by using a contour algorithm working runtime dynamically. For 85 queried regions, in an observer preference study, a scheme was based on referenced regions divided into two sets, and the previously randomly selected regions were compared with the queried sets. In 54.1% of the examined regions, the four observers chose the reference image set that was visually more similar when compared to the queried region. Zou et al. [102] proposed a deformable model for parametric contouring using partial differential equations for the determination of gradient vector flow field, yielding an accuracy of 82.6%. After the enhancement of mammographic images with adaptive histogram equalization, the GVF field component with the larger entropy was used to generate the ROI. Yuan et al. [103] employed a Full-Field Digital Mammography (FFDM) image to highlight abnormal masses from the surroundings using a dual-stage procedure. Firstly, the initial contour is determined using index-based segmentation using radial gradients. Secondly, a contour model is used dynamically for the region segmentation to find the abnormal mass contour. The distance feature outperformed using the leave one out method using lesion-size, and -contrast, and distance features, yielding the AUC (ROC) = 0.86. Similarly, Hassanien and Ali [104] proposed an algorithm for segmenting abnormal masses using fuzzy sets with Pulse Coupled Neural Networks (PCNN). Before segmentation, the fuzzy histogram hyperbolization is applied as a filter, followed by PCNN. Although, the image processing methods are very old, they are being used as preprocessing support to exploit the feature space for AI-techniques.

3.3. Machine Learning

This section surveys the ML-based approaches for the detection and classification of breast cancer. In this context, numerous methods have been accumulated spanning over time, emerging for mammography datasets focusing on different aspects of the problem domain. Davies and Dance [46] studied three-step mammogram enhancement before classification. A TP of more than 85%was reported with an image count of 78 during the testing phase. The Frame Texture Classification Method (FTCM) was used for the classification of abnormalities in mammograms. Local thresholding was used, where the calcification was segmented from the background for further classification of images. These studies reported the true positives (TP) of 100% by testing the methodology on 50 images. Rad et al. [55] used a multi-wavelet-based features extraction technique and features used for training the classifiers and achieved an accuracy of 85%.

Caldwell et al. [105] computed the correlation between the fractal dimensions to mammograms by experimenting with NB and SVM classifiers. The study reported an accuracy of 84% (radiologist-approved) using 70 mammographic images. Zheng et al. [106] used a Bayesian network in which an acyclic graph is used to compute probability influences between features. By experimenting with a dataset of 433 images and 12 features, the authors reported AUC (ROC) curve equal to 0.87. The Nijmegen database was used in this study to test the methodology on 40 images. In another study, 180 images from the Nijmegen LLNL/UCSF database were tested by the k-NN classifier. It was trained on three categories of extracted features (a combination of statistical and multi-resolution) and reported an accuracy of 80% [107]. Cao et al. [108] investigated the additional benefit of peri-calcification areas in contrast-enhanced mammography for distinguishing between breast lesions appearing solely as calcifications on standard mammograms. They analyzed radiomic characteristics from both low-energy and recombined images within the calcification sites and their surrounding peri-calcification regions. These regions were defined by extending the annotation margin radially with gradients ranging from 1 mm to 9 mm. ML models were employed to categorize calcifications as either malignant or benign. In another effort, Prinzi et al. [109] presented a radiomic signature aimed at effectively distinguishing between healthy tissue, benign microcalcifications, and malignant microcalcifications. Radiomic features were extracted from a proprietary dataset that included 380 samples of healthy tissue, 136 samples of benign microcalcifications, and 242 ROIs with malignant microcalcifications. Following this, two separate signatures were identified for detecting healthy tissue versus microcalcifications and for the classification of benign versus malignant microcalcifications. Various ML models, including SVM, RF, and XGBoost, were used as classifiers. The performance of the models was assessed, with XGBoost achieving an AUC (ROC) of 0.830 for healthy tissue classification, 0.856 for benign microcalcifications, and 0.876 for malignant microcalcifications.

Yoen et al. [110] investigated factors linked to abnormality scores generated by AI software. They performed a retrospective search in a database to identify a series of asymptomatic women who had undergone breast surgery from (2016-2019. Preoperative mammograms were evaluated using AI software (LunitINSIGHT) to assign abnormality scores. A score greater than 10 indicated positive detection of an abnormal lesion. A general linear model was employed to analyze the mammographic and pathological findings along with clinical repercussions associated with the AI software scores. Similarly, Malek et al. [111] introduced a method for filtering and extracting features utilizing persistent homology (PH), a robust mathematical tool for analyzing intricate datasets and patterns. Instead of directly operating on the image matrix, the filtering process is conducted on the diagrams derived from PH, enabling the identification of significant image characteristics amidst noise. These filtered diagrams are then transformed into vectors using PH features. Supervised ML models were trained on the MIAS and DDSM datasets to assess the discriminative features in classifying the benign and malignant classes and to determine the optimal filtering level. The findings of this study demonstrate that selecting appropriate PH filtering levels and features can enhance the accuracy of early cancer detection classification.

Khalid et al. [112] introduced an effective DL model to detect breast cancer in computerized mammograms of varying densities. Their approach involved feature selection by elimination of high-bias and repeating features. The model was evaluated using 3002 merged images who underwent digital mammography between (2007-2015). Additionally, six classification models were applied for breast cancer diagnosis, including RF, DT, k-NN, LR, and SVM. The study demonstrates that the proposed model achieved high efficiency, requiring minimal computational resources while maintaining a high accuracy score. Li et al. [113] used adaptive and multi-scale processing techniques to create a novel CAD mass detection system that improved sensitivity, specificity, and robustness against fluctuations in mammography. Using the dataset, they integrated hard and soft categorization through a modified fuzzy decision tree and committee decision-making method. This approach enabled them to achieve an AUC (ROC) of over 0.9. Iseri et al. [114] developed a novel method for the classification of microcalcification clusters employing a multi-window-based statistical analysis (MWBSA) method. This method uses a two-stage software framework as a computational search and analysis system, with an artificial neural network (ANN) as a classifier. Experimental results on different datasets with MIAS and DDSM are also presented. Datasets showed that the MWBSA-based strategy was as successful as other well-known methods such as GLCM and Wavelet techniques. Their classification task achieved a 97% accuracy rate. Rampun et al. [115] presented a novel investigation of ML performance by examining probability outputs in conjunction with classification accuracy score and AUC for ROC curve. They used the CBIS-DDSM database1872 micro-calcification clusters. They experimented on Random Forest (RF), Multi-layer Perceptron (MLP), Logistic Regression (LR), Naive Bayes (NB), Bayesian Network (BNet), k-NN, Alternate Decision Tree (ADTree), Logistic Model Trees (LMT), AdaBoostM1(AdaBoost) and SVM.

Further, Fanizzi et al. [116] suggested a binary classification model for differentiating tissues in digital mammograms to support radiologists in their work. They specifically examined the impact of various techniques on the feature selection procedure concerning the chosen features and learning performances. They extracted textural features using Haar wavelet decompositions for each ROI, as well as interest spots and corners that were found using the Minimum Eigenvalue Algorithm (MinEigenAlg) and Speeded Up Robust Feature (SURF). Next, a subset of features was chosen using two distinct feature selection methods, namely filter and embedding methods, used to train an RF binary classifier. For the benign/malignant and normal/abnormal situations, the prediction performance achieved the accuracy and median AUC (ROC) values as (97.31% and 88.46%), and (98.16% and 92.08%), respectively. Vy et al. [117] created a machine-learning classification model that would use clinical factors, mammography results, ultrasound results, and histopathological features to distinguish between ductal carcinoma in situ (DCIS) and minimally invasive breast cancer (MIBC). Using a trained XGBoost algorithm, tumors were classified as DCIS or MIBC using the five most significant clinical characteristics including calcification on mammograms. XGBoost model achieved an AUC (ROC) of 0.93 and an accuracy of 84%, comparable to that of a skilled radiologist in distinguishing between DCIS and MIBC and providing patients with the most excellent possible therapy options.

Sarvestani et al. [118] assessed the practicability and precision of automatically separating images of microcalcifications in the breast tissue. The decision tree classification approach was used to classify the breast tissue microcalcification clusters that have been identified. ANN was employed to identify the benign and malignant forms of segmented ROI clusters. The DDSM was used to train the proposed system. After training, the model led to an improved generalization of 93%. Lin et al. [119] showed that an automated deep-learning pipeline can facilitate early breast cancer diagnosis for mammography microcalcifications detection and classification. The system's development and testing utilized a total of 4,810 images from various centers. For both the training and test sets, the overall classification accuracy values for differentiating benign and malignant breasts were 0.7237 and 0.8124, respectively. Their automated artificial intelligence system was claimed to have the potential to enhance the decision-making abilities of clinicians by helping them identify, diagnose, and treat breast cancer more effectively. Fu et al. [120] proposed a two-step detection model. First, the location and nature of potential microcalcifications were determined using a mathematical model. It was investigated that the proposed model could accurately detect micro-calcification events when tested in the Nijmegen University Hospital database. Second, after extracting the features of each putative microcalcification, the sequential forward search (SFS) method was applied to identify sensitive attributes for microcalcifications. The classification performance of SVM and general regression neural network (GRNN) in terms of AUC (ROC) was 0.98 and 0.97, respectively, using the test data set.

Computer-assisted breast cancer detection from mammograms was introduced by Golobardes et al. [121]. In the first phase, several imaging techniques were used to extract the microcalcification features from the mammograms. In the second stage, automatic diagnosis was obtained by applying several ML algorithms. The authors examined the use of these characteristics as a classification system to differentiate between benign and malignant microcalcifications in mammograms taken from the mammography database of the Girona Health Area. They achieved a maximum accuracy of 78.57%. Similarly, Alolfe et al. proposed a four-step approach to developing intelligent diagnostic systems [122]. These steps were (a) the selection of ROIs, (b) the use of wavelet decomposition as the basis for the feature extraction step, (c) the selection of features, and (d) the classification of the results. The SVM classifier and the voting k-NN classifier were utilized in the classification stage. The MIAS mammographic datasets comprising 322 mammograms were used to assess the suggested method. SVM yielded the best accuracy of 87.5% of the potential classifiers tried, whereas k-NN produced the best accuracy of 75%. Using the Nijmegen and MIAS mammographic databases, Papadopoulos et al. [123] demonstrated a novel automated system for characterizing microcalcification clusters in three phases (cluster detection, feature extraction, and classification). A rule-based system, an ANN, and an SVM were constructed and assessed during the classification stage using the ROC analysis. A comprehensive comparison of evaluations of some ML-based techniques is illustrated in Table 4.

3.4. Deep Learning

Convolution Neural Network (CNN) is mainly used for DL. A CNN is constituted of n layers and m hidden neurons. Weights are associated with each input and output link in the neural network. Regulating weights in training helps neural networks to learn significant information, known as representation learning. The information flows in the learning phase to make it a self-learning network. CNN achieves impressive results in mammography for classifying the microcalcifications. Depending on the complexity of mammograms, features are hidden locally in mammograms that exhibit similar information, so it is tough to recognize abnormalities. The changes made in the CNN architectures were mainly in the input size, depth and size of kernels, activation functions, surviving fractions of neurons in the drop-out layer, fully connected layers’ specifications, and the decision layer.

In a study by Sahiner et al. [70], regions of interest are extracted for CNN’s input. The back-propagation technique is used to train the CNN. The area under the curve was 0.87, while 90% and 30% of TP and FP were reported in this study. The results were computed on the test dataset comprising 168 images. Moreover, MCP-CNN is classified based on mass features extracted from mammograms [71]. The area under the ROC curve of 0.89 was reported when taking 144 mammograms to evaluate the proposed model. Fonseca et al. [72] presented CNN with an SVM for the classification of breast cancer and reported a Kappa value of 0.58. Wu et al. [124] employed the convolution neural network for the classification of microcalcifications into benign and malignant categories. The pixel values of mammograms are considered global features and fed to CNN. After testing the methodology using 40 images, the sensitivity and specificity score was 75% for each. Furthermore, TL and CNN were used for the classification of the breast cancer dataset by Huynh et al. [125]. For ensemble and analytical extracted features, the AUC (ROC) was 0.86 and 0.81 respectively. The study used 219 mammograms for the evaluation of the methodology. Jiao et al. [74] combined intensity information and the deep features using original images for classification. The accuracy of 96.7% and 97.0% were achieved for CNN and VGG-based models using the DDSM dataset for the training and testing process. A hybrid approach was proposed by Arevalo et al. [75] to classify mammograms. Upon using 736 mammographic images, the area under the ROC curve of 0.89 was achieved.

Jadoon et al. [77] used the IRMA dataset for their experimentation. For testing, 2796 ROI patches from mammograms were used and achieved accuracy scores of 81.83% and 83.74% from CNN-DW and CNN-CT respectively. Another study used CNN to focus on semantic regions and the Recurrent Neural Network (RNN) was used to classify breast cancer. The study was performed by Wang et al. [80] and reported an accuracy of 89% using 763 images from the Breast Cancer Digital Repository. Gastounioti et al. [81] used Lattice-based techniques to extract texture feature maps and fed them to the multichannel CNN. Upon using 106 cases for evaluation, an accuracy score of 90% was reported.

Jaffar [78] used CNN for feature extraction and an SVM was applied for the classification of Mini-MIAS and DDSM datasets. An accuracy of 93.35% and sensitivity of 93.00% were achieved. In another research by Chakravarthy and Rajaguru [126], the ResNet-18 was used for feature extraction with the proposed improved crow-search optimized extreme learning machine (ICS-ELM). The proposed framework was experimented using DDMS, MIAS, and INbreast datasets. The accuracy scores of 97.193, 98.137, and 98.266% were obtained for DDSM, MIAS, and INbreast datasets respectively. Ueda et al. [127] used VGG-16 on the IDC dataset and reported a 61-70% accuracy score for the classification.. The DCNN system with modified GAN was proposed for mammogram images by Swiderski et al. [128]. The detection rate was reported at 89.71% for breast masses. Xu et al. [129] used a Multi-Scale Attention-Guided Network (MSANet) and achieved an AUC (ROC) as 0.94 when experimenting DDSM dataset. Furthermore, AlexNet, DenseNet, and ShuffleNet were experimented with using the INbreast dataset by Huang et al. [130]. The accuracies for test partitions using AlexNet, DenseNet, and ShuffleNet were 95.46%, 99.72%, and 97.84%, respectively.

Zebari et al. [131] worked on classification of benign and malignant tumors using mammograms. ROI was derived using hybrid thresholding and ML method. It was then separated into five blocks, and noise removal was carried out by applying wavelet transform. Multiple features were extracted using multifractal dimension approach and then features were reduced using genetic algorithm. They used five classifiers, and the final decision was made by fusing the results of all the blocks. The model was evaluated on four mammography datasets, and an accuracy of greater than 96% was achieved. Recently, Nomani et al. [132] presented a breast cancer detection model using ML based on particle swarm optimized wavelet neural network (PSOWNN). Dimension reduction was carried out through PCA, and each image was manually cropped to obtain ROI. Features were extracted using a gray-level co-occurrence matrix. The model was verified using a published dataset, and it correctly classified 95.2% of the 905 images. Maqsood et al. [133] worked on a DL breast cancer detection system for early diagnosis using an end-to-end training strategy. Initially, a contrast enhancement method was used to refine the edges of mammogram images. The classification performance was enhanced using transferable texture CNN (TTCNN) and energy layer employed to extract texture features. The model consisted of three layers of convolution and one energy layer. Deep features were extracted and fused, and the best ones were selected using entropy controlled firefly method. Three public datasets were used, and an average accuracy of 97.49% was achieved.

Further, Lin et al. [134] developed a DL model for breast cancer diagnosis. Data classification was carried out using ANN and SVM. TL-based DL architectures (ResNet101, AlexNet, and InceptionV3) were studied using adaptive moment estimation and stochastic gradient descent with momentum-based optimization algorithms. The model was verified using a breast cancer risk factor dataset, and an average accuracy of 94.2% was obtained. Mohapatra et al. [135] evaluated the performance of various CNN architectures by training through weights reset (from scratch) and others through TL (pretrained). A small public mammography dataset was used for this study, and data augmentation by rotation and zooming techniques was applied to overcome overfitting. AlexNet and VGG16 showed an accuracy of 65%, while ResNet50 showed an accuracy of 61%. Al-Tam et al. [136] developed a hybrid DL framework for breast lesion detection. The backbone residual DL network was utilized to create deep features, while the transformer was used to classify breast cancer according to the self-attention mechanism. Two publicly available datasets were used to verify the model, and an accuracy of 100% for binary and 95.8% for multiclass categorization was achieved using 5-fold cross-validation. In their work, Kumar et al. [137] employed CNNs with four distinct optimizers, utilizing input images sized at 299 × 299 × 3. Feature mapping was achieved through a pre-trained InceptionResNetV2 model. Evaluation of a curated subset of the DDSM mammogram dataset showcased superior performance in sensitivity, specificity, accuracy, and AUC (ROC) compared to previous DL and classical ML approaches.

Another study introduced a fully automated system for diagnosing breast cancer, utilizing an AlexNet and multiple classifiers to attain heightened accuracy level. Validation through testing on three Kaggle datasets confirmed its superior performance, indicating its potential utility in aiding medical professionals with precise diagnosis [138]. Marathe et al. [139] introduced a quantitative method to differentiate between benign and actionable amorphous calcifications on mammograms. This method utilizes local textures, global spatial relationships, and expert features. The approach, trained and validated on 168 digital mammography exams, demonstrated high sensitivity and positive predictive value on a test set. Liu et al. [140] examined the effectiveness of DL-based on full-field digital mammography in predicting the malignancy of BI-RADS 4 microcalcifications. Their combined DL model surpassed clinical and DL image models, achieving high diagnostic accuracy. Moreover, it assisted junior radiologists in enhancing their performance. This study underscores the potential of DL in improving malignancy prediction in screening mammography and supporting clinical decision-making. A study explored the use of CNNs to predict whether patients diagnosed with pure atypical ductal hyperplasia (ADH) could be safely monitored rather than undergoing surgery. By analyzing 298 mammographic images from 149 patients, the AUC (ROC) was found to be 0.86. The findings suggest the feasibility of distinguishing ADH from DCIS [141].

Recently, Prodan et al. [142] explored DL methods, incorporating CNN and ViT architectures, to analyze mammograms with the goal of enhancing classification accuracy. They suggested an alternative data enhancement method using simulations to increase model accuracy and emphasized the importance of careful data preprocessing. Furthermore, their study combined interpretable AI techniques such as class function mapping and centralized bounding boxes to gain insight into the decision-making process of the model. Beuke et al. [143] developed a DL algorithm to detect and classify breast lesions in contrast-enhanced mammography (CEM) images. Collected retrospective disease data, trained DL, and handcrafted radiomic features to distinguish between benign and malignant lesions were developed. The results showed remarkable sensitivity in lesion detection and good diagnostic accuracy. Notably, the combined DL and manual radiomics models achieved a remarkable AUC (ROC), indicating an encouraging probability for accurately detecting malignant lesions depicted in CEM images. Pesapane et al. [144] developed an AI model to detect and characterize microcalcifications in mammograms and facilitate breast cancer screening. The model used three networks (AlexNet, ResNet18, and ResNet34) to find impressive sensitivity, specificity, and AUC for ROC curves in microcalcification detection and classification tasks. The results confirmed the exceptional performance of AlexNet, while ResNet18 and ResNet34 also showed encouraging results.

Similarly, Kumar et al. [145] focused on a DL-based model for breast cancer detection. When model performance was evaluated using metrics such as sensitivity, specificity, accuracy, and F-measure on benchmark datasets. It was found that the proposed models performed better than current methods, assisted radiologists with diagnostic and classification accuracy, and highlighted potential improvements. Tsai et al. [146] proposed a multitask DL algorithm to predict extensive intraductal component (EIC) in breast cancer development. Using mammograms from (2010-2019), the three-stage DL model showed impressive performance, with AUC (ROC) values of 0.758 and 0.775 for EIC prediction in imaging and breast, respectively. This multitask paradigm facilitated the simultaneous classification of imaging findings and prediction of EIC, thereby highlighting the potential of DL algorithms in enhancing breast cancer diagnosis via mammography. A comprehensive comparison of evaluations of some more DL-based techniques is illustrated in Table 5.

3.5. Future Applications of AI in Mammography

There are several potential future applications of AI in mammography. AI can be used as a triage tool to flag highly suspicious mammograms for expedient interpretation. Improved efficiency of radiologists using AI can translate to decreased turnaround times. AI can be used for breast density assessment, which can be associated with breast cancer risk. Changes in breast density over time can be assessed with AI and utilized as a biomarker for treatment response. AI-based breast density algorithms may also be incorporated with other mammographic features as well as age, genetic data, and other clinical factors to produce a risk prediction mode for risk stratification that can be used as a clinical decision support system. AI algorithm for the detection of arterial calcifications can identify patients at risk for coronary artery disease [151]. DL models that accurately exclude cancer on mammograms can diminish radiologists’ workload and can be useful in underserved areas where breast imaging expertise is less available [152]. Widespread acceptance of AI models can be promoted where they generalize well and are validated on large, diverse populations, varied vendors, and imaging acquisition techniques in large retrospective studies. The higher specificity of AI-based CAD compared to the older non-AI CAD models and the potential to decrease recall rates, increase cancer detection rates, and increase overall performance by enhancing these metrics, in addition to the ability of AI to improve efficiency will be of great benefit [153].

4. Discussion

Breast cancer continues to rank as the second-leading cause of cancer-related deaths among women. This study examines breast cancer detection via mammography over a specific period, aligning with multiple UN Sustainable Development Goals. Timely detection through screening has decreased mortality rates and potentially saved lives. Despite numerous review articles on mammographic breast cancer screening, there remains a gap for a focused review covering breast cancer detection from 1970 to 2023, particularly emphasizing image processing, machine learning (ML), and deep learning (DL). Utilizing DL architectures with multi-scale feature extraction that transforms contextual information into feature data has yielded exceptional outcomes. The article also delves into the significance of extensive mammography datasets for thoroughly exploring innovative early detection methods. While DL algorithms for breast cancer detection are increasingly being applied clinically, broader integration into mammogram-based cancer detection in clinical settings calls for additional clinical validation, enhanced reliability, generalizability, and interpretability, among other considerations.