Artificial intelligence in breast imaging: potentials and challenges

Breast cancer, which is the most common type of malignant tumor among humans, is a leading cause of death in females. Standard treatment strategies, including neoadjuvant chemotherapy, surgery, postoperative chemotherapy, targeted therapy, endocrine therapy, and radiotherapy, are tailored for individual patients. Such personalized therapies have tremendously reduced the threat of breast cancer in females. Furthermore, early imaging screening plays an important role in reducing the treatment cycle and improving breast cancer prognosis. The recent innovative revolution in artificial intelligence (AI) has aided radiologists in the early and accurate diagnosis of breast cancer. In this review, we introduce the necessity of incorporating AI into breast imaging and the applications of AI in mammography, ultrasonography, magnetic resonance imaging, and positron emission tomography/computed tomography based on published articles since 1994. Moreover, the challenges of AI in breast imaging are discussed.


Introduction
Breast cancer, which was previously the most common type of malignant tumor among females, has become the top-ranking malignant tumor in humans overall (Ferlay et al 2021).However, the mortality rate of 15.5% is much lower than the incidence rate of 24.5% among females (Ferlay et al 2021).The promising prognosis is a result of the efforts of scientists and breast physicians, who have worked on understanding precise subtypes and personalized therapy in the past two decades.Specialists have found that breast cancer is a disease with the prospect of precise therapy.Breast cancer can be divided into unique groups using the immunohistochemical expression profile typically based on estrogen/progesterone receptor statuses, and expression of human epidermal growth factor 2 (HER2) (Goldhirsch et al 2013).Each subtype exhibits specific biological tumor characteristics.The significant advancements in surgery, radiotherapy, endocrine therapy, chemotherapy, and targeted treatment of patients based on molecular subtypes have laid the foundation for the precise classification and personalized treatment of breast cancer.Despite this progress in the treatment of breast cancer, tumor heterogeneity, the diversity within the tumor and/or among different patients (Turashvili and Brogi 2017) is a non-negligible factor in determining optimal treatment strategies and prognosis of breast cancer (Coates et al 2015, Prat et al 2015, Harbeck and Gnant 2017).The luminal subtype is characterized by a high incidence of long-term recurrence (Gao and Swain 2018), the HER2 subtype is likely to be resistant to targeted therapeutic medicine (Ocaña et al 2020), and the triple-negative subtype has a poor prognosis within five years (Hwang et al 2019).In addition, according to the authors' experience, tumor heterogeneity is an obstacle to the accurate and early diagnosis of certain atypical breast tumors (Li et al 2018).Delayed diagnosis may inevitably affect the timely treatment of breast cancer patients.
Imaging plays an important role in breast cancer screening and diagnosis.Mammography (MG) and ultrasonography (US) are the preferred techniques for breast cancer screening and diagnosis.Magnetic resonance imaging (MRI) provides high sensitivity for breast cancer detection and diagnosis.Positron emission tomography combined with computed tomography (PET/CT) is important in the management of breast cancer patients.With the recent advancements in imaging technologies, most breast cancers can be recognized with prompt diagnosis and efficient consecutive therapies.However, as mentioned above, tumor heterogeneity is a major obstacle to the precise diagnosis of certain atypical or very early breast cancers, such as triple-negative breast cancer (TNBC) (Li et al 2018), mucinous breast cancers (Ginter et al 2020, Pintican et al 2020), and ductal carcinoma in situ (Watanabe et al 2017).In these cases, physician experience is of limited value, as human eyes cannot differentiate the subtle differences in the images.
To assist radiologists, scientists have been trying to introduce the intelligent algorithms that can respond in a similar manner as human beings, which is known as artificial intelligence (AI).As a branch of computer science, AI is a technological process that simulates, extends, and expands human cognitive thinking to complete a task through extracting and synthesizing abstract information.The application of AI to the medical imaging field was first proposed in the 1990s (Shen et al 2021b).As breast cancer is the most common malignant tumor in females, early and precise diagnosis and treatment have a significant potential impact on patients, thereby motivating innovation in the application of AI techniques (Morgan and Mates 2021).AI has been proved to be valuable for breast cancer screening, detection, differential diagnosis, molecular subtyping, treatment response and prognosis prediction, etc, regarding to the commonly used breast imaging techniques.Recently, as summarized by Bahl, there have been about twenty AI applications that are approved by the Food and Drug Administration (FDA) for MG, breast US, and breast MRI (Bahl 2022).
In the literature, there have been a number of review articles about AI applications in breast imaging written from multiple perspectives (Le et al 2019, Mendelson 2019, Bahl 2020, Hickman et al 2021, Morgan and Mates 2021, Bitencourt et al 2021).Compared with previous reviews, in this article, we will address the issue of tumor heterogeneity which is one of the most important justifications for introducing AI in breast imaging.In addition, the applications of AI in breast PET/CT are summarized in the present article.

Outline of review
A schematic outline of this review is presented in figure 1.In the Introduction section, the applications of in vivo patient imaging techniques for breast diseases are highlighted, followed by the introduction of AI into breast imaging.Based on published articles since 1994, the following sections of the article present the applications of AI in breast imaging, which include: (1) common AI methods that are used for breast imaging, (2) applications of AI in MG, (3) applications of AI in breast US, (4) applications of AI in breast MRI, and (5) applications of AI in breast PET/CT.Subsequently, the technical and clinical challenges of AI in breast imaging are discussed, followed by an outline of future work and conclusions.1.2.Application of imaging techniques for breast diseases MG and US have been recognized as the most common imaging techniques for the detection of breast cancer (Guo et al 2018a).MG is the primary diagnostic technique that is used in screening programs, particularly in Western countries.MG can detect microcalcifications that are invisible in US (Jackson 2004).Nevertheless, it exhibits certain limitations in the case of dense breasts, which are common in Asian women and young women.US is complementary to MG in the detection of breast cancer.In China, US is usually recommended as an effective combination with MG as the first choice for breast screening and the differential diagnosis of benign and malignant breast masses (CSCO 2021 ).As no radiation-related damage occurs, it is suitable for any age and physiological period of females, particularly for pregnant or lactating women.US also offers advantages in the detection of potentially malignant changes in axillary lymph nodes (Guo et al 2018a).
Among the available imaging modalities, MRI provides the highest sensitivity for breast cancer detection and diagnosis (Mann et al 2019).It is currently used as an adjunct to MG screening, especially for high-risk patients.As stratified by American Cancer Society, the risk factors for breast cancers mainly include genetic mutations, family history, and clinical risk factors such as thoracic radiotherapy, lobular neoplasia, ductal hyperplasia, and high mammographic density (Saslow et al 2007).Furthermore, it is an indispensable tool for assessing the preoperative stage, evaluating the treatment response, and diagnosing difficult and complicated cases (Mann et al 2019).MRI can provide more detailed information through a variety of scanning sequences compared to other imaging modalities.The dynamic contrast-enhanced (DCE) image and post-contrast T1-weighted image (T1WI), which provide the morphological and kinetic features of breast lesions observed after contrast material administration, form the basis for breast MRI protocols.The T2-weighted image (T2WI) enables the visualization of cysts, edema, and necrosis owing to their liquid nature, and such information is important for prognostic assessment.The diffusion-weighted image (DWI) quantifies the random movement of water molecules in tissues, which is associated with tissue microstructure and cell density (Mann et al 2019).However, owing to its high cost and long image acquisition time, MRI is not as popular as MG and US imaging.
PET/CT can provide three dimensional (3D) map of the activity distribution of a radioactive tracer, which in the case of 18 F-FDG can be used to estimate the glucose metabolism in tissues and standard metabolic parameters, including the maximum standardized uptake value (SUV max ), metabolic tumor volume, and total lesion glycolysis.However, compared with US, MG, and MRI, PET/CT is not suitable for breast cancer screening because of its high cost and radiation exposure.Its main applications include the staging and evaluation of the treatment response and suspected recurrence or metastasis (Fowler and Cho 2021, Kikano et al 2021, Sarikaya 2021).
In clinical practice, MG, US, MRI and PET-CT are applied selectively and complimentarily to aid in the screening, diagnosis, treatment response monitor and prognosis prediction (figure 2).With the advancement of imaging technology, the amount of data is increasing which is time consuming for radiologists.For example, with the introduction of digital breast tomosynthesis, a stack of 2D slices of the imaged breast, although the diagnostic performance is increased, the interpretation time for radiologists almost doubles compared with MG (Skaane et al 2013).The similar issue also arises for automated breast volume scanning (ABVS) (Ibraheem et al 2022) and multi-parametric MRI imaging comprising ∼5 imaging sequences (Mann et al 2019).Therefore, automated methods of interpreting these images are highly demanded to reach a balance between diagnostic performance and interpretation time.On the other hand, the performance of breast imaging is closely related to the heterogeneity of breast cancers especially for the prediction of treatment response and prognosis.Due to the heterogeneity, human interpretation without powerful computation is limited to achieve the best performance for differentiating different biological subtypes with a variety of possible treatments, predicting the response to therapy and overall survival.These noninvasive observations and predictions which are beyond traditional detection and diagnosis are important parts in the era of individualized and precision medicine.AI has been introduced in the context of advanced imaging and the demand of powerful computations.

Necessity of introducing AI into breast imaging
In 2021, the Chinese Society of Clinical Oncology guidelines stressed the important role of AI in the field of tumor diagnosis and treatment with an increasing application in clinical practice (CSCO 2021).AI is gaining attention in the diagnosis and treatment of breast cancer, the prediction of survival prognosis, and the prediction of treatment response after neoadjuvant chemotherapy (NAC) owing to its powerful computing and learning ability (Bahl 2020).
In recent years, AI-assisted imaging diagnosis has flourished with the evolution of big data and computational frameworks.The role of conventional computer-aided diagnosis (CAD) is expanding beyond screening and differential diagnosis toward applications in therapy evaluation and risk assessment (Giger 2010, Mendelson 2019).Moreover, intelligent output can be obtained from automated analysis based on images that are collected by radiologists to assist in detecting lesion, determining the lesion malignancy rate, evaluating the response to therapy, and predicting the prognosis.
AI may reduce human efforts in terms of detecting suspicious nodules or masses in US or MG images in the screening field.At the authors' institute, an US AI equipment has been developed for the automatic location of suspicious areas during US scanning which is valuable for inexperienced physicians (Hou et al 2022).Meanwhile, the commercial S-Detect technology has been used clinically to increase the confidence of US physicians in diagnosing breast nodules (Kim et al 2017, Sun et al 2022a).
The role of imaging has expanded from screening and diagnosis to the prediction of treatment efficacy and prognosis, considering the vast information that is hidden in the images.However, human eyes have a limited ability to achieve advanced predictions without the aid of AI such as the prediction of response to chemotherapy and prognosis of breast cancer based on imaging information (Galati et al 2022).

Tumor heterogeneity
It is well known that breast cancer is a type of heterogeneous malignant tumor comprising multiple distinct subtypes that differ on the clinical, histopathological, and genetic levels (Roulot et al 2016).Tumor heterogeneity is characterized by inter-and intra-tumor heterogeneity.
Inter-tumor heterogeneity has been described as the variety among different tumors, and it was extensively characterized in the 2000s owing to the development of high-throughput analyses (Zardavas et al 2015, Roulot et al 2016).Inter-tumor heterogeneity has implications for guiding the treatment of the four breast cancer subtypes (Goldhirsch et al 2013).Each subtype has specific biological characteristics and clinical behaviors, which provide the foundations for the precise treatment of breast cancer with a promising prognosis.
Intra-tumor heterogeneity has been identified within the different regions of the tumor (spatial heterogeneity), along with tumor progression (temporal heterogeneity).Pathological and immunohistochemical results that are obtained from a biopsy or a small portion of tumor specimens may not represent the overall tumor composition owing to the spatial heterogeneity.Therefore, it is important to recognize the spatial heterogeneity as it may be indicative of treatment effectiveness, with evidence that increases heterogeneity corresponds to a reduced likelihood of pathological complete response (pCR) (Januškevičienė and Petrikaitė 2019).Furthermore, tumors with more prominent heterogeneity may be resistant to therapy as they adapt to new microenvironmental conditions more easily (Issa-Nummer et al 2013, Almendro et al 2014).
In the era of precision medicine, it is important to capture the heterogeneity of each specific molecular subtype, as this biological variance enables such heterogeneity to be anticipated and adaptive therapeutic strategies to be sought.The imaging appearance is an integral phenotype of all proteomics and genomics (Aerts 2016).Therefore, the imaging feature of the breast mass is an important supplement to the local pathological and immunohistochemical characteristics in the development of precision medicine and personalized treatment.AI is expected to facilitate this integration because of the vast amount of information hidden in the images that it conveys.

Precision medicine and personalized treatment
Evidence-based medicine that results from a precise subtype of breast cancer can provide sufficient details for precision medicine in breast cancer.However, inter-and intra-tumor heterogeneity is an obstacle to the efficient treatment of all breast cancers.Thus, personalized treatment is in high demand to improve the outcome of breast cancer further (Jiang et al 2021d).Advanced research in multi-omics analysis and intra-tumor interaction with the microenvironment is warranted to enrich the evidence for personalized treatment.
Precision tumor medicine refers to the use of various omics detection technologies, including proteomics, transcriptomics, genomics, epigenemics, and metabonomics, to obtain tumor-related biological information for guiding tumor screening, diagnosis, and treatment (Pinker et al 2018, Sachdev et al 2019).Multigene mutation detection offers significant value for molecular subtyping, breast cancer risk prediction, and the selection of precise treatment plans.For example, the Fudan University Shanghai Cancer Center Precision medicine should consider the static omics of individual breast cancers as well as the dynamic omics during treatment and follow-up (Pinker et al 2018).Breast imaging offers the advantage of the dynamic surveillance of breast tumors throughout the whole process of screening, diagnosis and post-operative follow up.Therefore, in addition to biological omics, Pinker et al proposed the combination of quantitative radiomics which can extract valuable quantifiable data from digital medical images with multiple biological omics to provide dynamic surveillance for breast cancer (Pinker et al 2018).This approach is known as radiogenomics, which may link the complete imaging appearance with genetic information.Radiogenomics can quantify lesion characteristics to stratify benign and malignant breast tumors more effectively, thereby enabling precise diagnosis.It can also reflect the genetic information of a heterogeneous tumor and guide tailored therapy.After the therapy, radiogenomics can also incorporate imaging biomarkers with phenomics and genomics to predict recurrence risk.

Common AI methods in breast imaging
The most frequently concerned AI models are machine learning (ML), deep learning (DL) and convolutional neural networks (CNNs) (Castiglioni et al 2021).The relationships between AI, ML, DL, and CNNs are depicted in figure 3.They are different in terms of the capability, complexity, interpretability, and the types of problems they're best suited for.
ML models primarily include support vector machines and random forests.Support vector machines group data into two or more classes through a 'hyperplane' that separates the categories as far as possible by analyzing numerous features (Kohli et al 2017).Random forests employ a collection of decision trees based on a random subset of features that are extracted from the training data.When a new input appears, the model makes a prediction (e.g.'positive' or 'negative') for each tree and the voted result from all trees is considered as the best solution.These models are relatively straightforward and work well with structured, tabular data.They are less complex, easier to understand, and often provide good baseline models.However, they might not perform as well with extremely complex tasks or very large datasets.They also struggle with unstructured data, such as images or text.
DL, which is a subset of ML and offers the capability to cluster data and make predictions, uses neural networks to mimic the human brain (Tang et al 2018).The neural network consists of multiple layers of connected nodes, each of which receives input from other nodes with weights that are set randomly.DL models, can handle much more complex tasks and are particularly good at working with unstructured data like images, text, and audio.They can automatically learn and extract features from raw data, a process known as representation learning.However, they require large amount of data and computational resources.The complexity of DL models also makes them prone to overfitting if not properly regularized.Meanwhile, DL models are often referred to as 'black boxes' because it can be difficult to understand why they're making certain decisions.Some progress has been made in improving the interpretability of DL models (like attention mechanisms, feature visualization, etc), but it's still a significant challenge compared to traditional ML models.
As a subset of DL and the most common type of neural network, the CNN is suited to identify particular patterns in images which can occur at different locations because the convolution operation is spatially invariant (Burt et al 2018, Robertson et al 2018).In a systemic review, Nasser and Yusof found that CNN model has the most accurate performance with the most extensive application for breast cancer diagnosis (Nasser and Yusof 2023).A CNN consists of three layers: an input layer, a hidden layer (one or more hidden convolutional layers), and an output layer (Pesapane et al 2018).Although the performance of the CNN may improve with deeper architectures, this may result in network overfitting.An appropriate network design significantly contributes to the final performance (Abdelhafiz et al 2019).
Data are indispensable for training a sophisticated model.An AI model can be trained in three manners: supervised, unsupervised, and reinforcement learning (figure 3).Supervised learning creates a model to predict the outcomes based on labeled data.Unsupervised learning determines the patterns and associations in unlabeled data to create groups and clusters.Reinforcement learning takes advantage of the reward mechanism for training feedback to achieve a desirable or undesirable state.
Figure 4 presents the workflow of AI models.Most AI models in the breast imaging literature use supervised learning; for example, benign and malignant breast tumors are differentiated with breast images that are labeled as positive or negative.ML algorithms rely on hand-engineered (or hand-designed) features based on the knowledge and experience of the clinician (such as the density or shape), whereas DL algorithms learn the features automatically.Given a sufficiently large training dataset, DL-based AI systems may be able to classify data better than methods that use hand-designed features (Chartrand et al 2017).

Applications of AI in MG
Two main types of lesions appear on mammograms: calcification clusters and soft tissue findings (masses, distortions, and asymmetries).The significant advantage of MG is its high sensitivity in detecting calcifications, particularly microcalcifications, which are usually invisible in other imaging modalities such as US or MRI (Jackson 2004).
The earliest research on CAD in MG was conducted in 1967 by Winsberg et al with the motivation of liberating radiologists from the large volume of screening mammograms in asymptomatic women (Vyborny and Giger 1994).Subsequently, numerous trials have been conducted to develop CAD for MG.These trials can   The AI system has similar performance for detecting breast cancer in MG compared with an average of 101 radiologists.This finding was consistently validated in a large, heterogeneous, multi-center, multivendor, and cancer-enriched cohort.
The AI system had a higher AUC than the average of 101 radiologists (0.840 versus 0.814).The AI system had an AUC higher than 61.4% of the 101 radiologists Gu and Jiang 2022).In this section, we present a comprehensive review of these applications, with selected references indicated in table 2.

Automatic identification and segmentation
The hand-crafted outlining of breast lesion contours is time consuming and subjective, particularly for those without distinct margins.Therefore, the automated identification of lesions is desirable for efficient AI analysis.
Researchers have developed various automated AI detection models for breast lesions in US images (Marcomini et  Marcomini et al first developed an identification model for tissue-mimicking phantoms with nodules similar to breast lesions, and subsequently applied the algorithms to clinical images (Marcomini et al 2016).The neural multilayer perceptron classifier achieved an accuracy of 81% for breast lesion identification in clinical practice.The most suitable AI model for a specific clinical dataset needs to be selected among the vast number of available models.Cao et al evaluated the performance of four training protocols for object detection (Cao et al 2019).Shen et al developed an AI system to identify breast cancer in US images using the largest dataset to date, and the accuracy thereof could reach the level of radiologists (Shen et al 2021a).This demonstrates the potential of AI in future clinical practice.

Differential diagnosis
The detection of suspicious breast lesions in US images is the first step for US physicians.However, the most important aspect is the accurate diagnosis of the pathological properties; that is, whether the nodule is benign or malignant.Many AI models have been developed to assist US physicians in the differentiation of breast lesions In the authors' view, the differential diagnosis of most breast lesions with typical malignant sonographic features is not challenging for qualified US physicians.However, AI-assisted differential diagnosis is desirable for lesions with atypical US features.The authors' group previously evaluated the value of DL models in reducing the malignancy rate among breast imaging reporting and data system (BI-RADS) 4A lesions to achieve more accurate risk stratification (Zhao et al 2022).A further study is being conducted to evaluate the value of AI for diagnosing malignant breast tumors with atypical sonographic features using a large dataset from multiple centers.

Correlation with tumor invasive properties
Medical images were previously treated as gross anatomical images of tissues, organs, or lesions, in which the information of morphological changes was focused.However, in addition to displaying these conventional descriptive signs, medical images contain extremely large amounts of digital information that can be deeply excavated (Aerts et al 2014, Gillies et al 2016).The digital information is correlated with the molecular subtypes Breast cancer is a highly heterogeneous disease with four common molecular subtypes.Thus, the variety in imaging is expected to be a result of the heterogeneity of the biological properties.It has been found that sonographic radiomics can classify the molecular subtypes of both invasive breast cancer and ductal carcinoma in situ (Guo et al 2018b, Wu et al 2021, Zhou et al 2021a, Jiang et al 2021b).However, according to our experience, caution should be exercised in that the molecular subtypes overlap with one another in terms of the sonographic features.The classification potential of AI algorithms for molecular subtypes should be rationally examined (Shi et al 2021).
The axillary lymph node (ALN) status is crucial in determining the tumor stage and subsequent treatment strategy.The presence and load of ALN metastasis (ALNM) are dependent on the primary breast tumor.Therefore, the prediction of ALNM based on sonographic features using AI algorithms has been highlighted in various research articles (Yu et  Although the studies to date have been promising, robust results that have been verified using a large dataset remain lacking.Furthermore, it should be acknowledged that imaging information is not beyond clinicopathological factors, such as the molecular subtype, metastatic load in the axilla, and NAC regimen, which should be considered when designing similar studies.

Applications of AI in breast MRI
AI techniques that aid in MRI image analysis can facilitate radiologists in clinical decision-making with enhanced diagnostic efficiency and precision.Such techniques have mainly been applied to lesion detection, risk assessment, and treatment response prediction (Sheth andGiger 2020, Bitencourt et al , Satake et al 2022).Selected articles relating to MRI-based AI studies are summarized in table 3.   Fan et al found that the radiogenomic signature of the texture and morphological features was positively associated with the Oncotype DX RS, and a predicted RS that was greater than 29.9 was related to poor recurrence-free survival (Fan et al 2022a).In certain studies, MR images have been associated with other types of genetic testing, such as the 50-gene PAM50 and Curebest 95-gene assays, to identify radiogenomics signatures and provide alternatives for patients who did not undergo gene testing (Li et al 2016, Tokuda et al 2020).Ma et al developed a radiomics model using pre-and post-NAC DCE-MRI features to predict systemic recurrence in TNBC patients (Ma et al 2022).The radiomics achieved better predictive performance than the clinical model in predicting the recurrence risk within three years following NAC, with an AUC of 0.933.Thakran et al concluded that the radiomics features of parametric response maps that were derived from DCE-MRI kinetic maps achieved the best predictive performance for recurrence risk, with a C-statistic of 0.72 (Thakran et al 2022).

Applications of AI in breast PET/CT imaging
Breast imaging AI models based on PET/CT have also been studied in recent years (Romeo et al 2021, Sadaghiani et al 2021, Urso et al 2022).Applications of PET/CT include tumor staging, the evaluation of the treatment response, and suspected disease recurrence (Fowler and Cho 2021, Kikano et al 2021, Sarikaya 2021).Table 4 listed some elected AI studies in breast PET-CT or PET-MRI regarding to tumor detection, classification and prediction.
Krajnc et al established an ML model based on PET/CT to aid in the differentiation of benign and malignant tumors.Their method achieved an AUC of 0.81 for the differentiation and could identify TNBC with an AUC of 0.82 (Krajnc et al 2021).The PET and MRI-derived radiomic features were found to be associated with the tumor grade, overall stage, subtypes, prognosis (Huang et al 2018), and hormone receptors (Umutlu et al 2021).However, Araz et al found that all radiomics parameters from PET/CT failed to predict the hormone receptors (Araz et al 2022).PET-derived radiomics has also been applied to the prediction of other rare malignant breast cancers, such as breast lymphoma (Ou et al 2019).
As mentioned previously, ALNM is one of the most important clinical factors in determining treatment strategies and prognostic outcomes.PET/CT provides high specificity but relatively low sensitivity for ALNM evaluation.Advanced AI techniques have been applied to address this issue and promising results have been achieved (Li et al 2021, Song 2021, Chen et al 2022a).Chen et al used PET/CT radiomics to identify occult ALNM in clinically node-negative patients (Chen et al 2022a).The developed model improved the diagnostic performance of occult ALNM, with a mean AUC of 0.817 and mean accuracy of 0.812.With the prevalence of COVID-19 mRNA vaccinations in recent years, the correct differentiation between metastatic and reactive ALN has become a new challenge.Eifer et al found that the radiomics features that were extracted from PET/CT performed effectively in differentiating between breast-related ALNM and COVID-19 vaccine-related axillary lymphadenopathy (Eifer et al 2022).
The application of PET-based AI models for further identification of patients who may benefit from NAC at the early stage is an area of significant interest owing to the capability to quantify metabolic activity in breast tumors (Antunovic et

Challenges of AI in breast imaging
In addition to screening and detection, the ideal roles of AI in breast imaging include aiding radiologists in reaching the most appropriate diagnosis, assisting clinicians in creating the best treatment plan, and incorporating other clinical-pathological-immunohistochemical variables to predict the risk of recurrence or metastasis.Thus, breast imaging AI is expected to contribute to precision medicine and personalized treatment.However, various technical and clinical challenges exist in the sustainable development of breast imaging AI.

Technical challenges
First, big data forms the basis of AI in breast imaging.However, clinical breast images are not rich enough at one center.Multicenter studies are required to address this problem.Moreover, this challenge may be reinforced by the nonstandard nature of ultrasound images, e.g.deviations in the image collection, equipment, and image setting.This may be alleviated through accumulating enough data from various ultrasound equipment at different settings and developing sophisticated AI models to tolerate those interferences.
Second, the uninterpretability of current DL models that are applied to breast imaging makes it challenging to transfer the technique from research to real clinical practice despite of the applications commercial AI products.Such a challenge may be alleviated by the development of interpretable DL models in the future (Liu et al 2019a, Vellido 2020).
Third, as most current DL models are supervised, the model training process relies on well-defined training data.Thus, all regions of interest on the breast images should be well delineated, which requires substantial labor and is sensitive to subjective errors.The requirement of manually delineated labels in breast imaging may pose a significant challenge for a long time (Bi et al 2019).This challenge may be mitigated using unsupervised DL models which get rid of the delineation of labels (Chen et al 2023).

Clinical challenges
Although AI is a robust tool for dealing with complicated tasks, the integration of the computing resources that are required by AI necessitates human input, especially in the training stage.However, clinicians have limited time to collect massive amounts of data, which is why most related AI studies include a limited number of cases or focus on specific medical information (Nagendran et al 2020).Most studies evaluated the applications of AI based on one imaging modality.The combination of multiple imaging modalities is desired (Romeo et al 2021).Furthermore, the ethical issues relating to patient privacy and data security in breast imaging AI cannot be ignored.The protection of data security is critical when sharing data, especially in multicenter studies (Hickman et al 2021).
Tumor heterogeneity is a major obstacle for radiologists to give accuate diagnosis for each single case in the clinical circumstance as a result of variable imaging apperances, but also provides an opportunity for the continuous exploration of AI in breast imaging.It is difficult for radiologists to determine the pathological pCR prediction

ML radiomics model
The PET/CT-based radiomics analysis might provide efficient predictors of pCR in patients with breast cancer with the AUC ranging from 0.819 to 0.849 in the validation cohort nature of certain atypical breast lesions owing to the high heterogeneity of cancer.Thus, AI is desirable for aiding radiologists in reaching the most appropriate diagnosis for such breast lesions.Moreover, AI is expected to aid in evaluating the prospects, success, and failure of treatment outcomes based on learning from the successful treatment of clinical cases (Maddox et al 2019).However, the challenge from tumor heterogeneity also exists for AI to reach perfect computation results and agreement among different studies.An integrated model to incorporate imaging data with clinical-pathological-immunohistochemical-genetic information is desired to overcome the effec of tumor heterogeneity.
The role of AI systems in diagnosis applications has been disputed (Giger 2010).Should AI be used as the second reader or replace human readers if its standalone performance is comparable or superior to that of radiologists?Furthermore, when a controversy arises between AI and human readers, which diagnostic conclusion should be the final one?These questions need to be answered before AI can be applied extensively in clinical practice.
Breast imaging is useful for preoperative diagnosis, and its significance in guiding treatment strategies and prognosis prediction should also be explored.However, it is difficult to integrate breast imaging AI with clinical datasets without the support of clinicians.Fortunately, an increasing number of breast clinicians are focusing on the integration of imaging data and other related information into AI models to cater to personalized treatment and precision medicine (Jiang et al 2022a).Moreover, multi-omics studies have become a hot topic for characterizing the molecular biology of tumors, including the genomics, transcriptomics, proteomics, and metabolomics (Ponzi et al 2021).Current evidence suggests that the clinical transformation of most developed high-performance AI algorithms remains in the initial stages (Nagendran et al 2020).It is expected that all information of each specific patient will be consolidated to build a large data archive for training robust AI models at all institutions in the near future.Personalized treatment and prognosis prediction for subsequent breast cancer patients can hopefully be realized using such models.

Future work
Breast imaging AI is not expected to exceed radiologists for lesions with typical benign or malignant imaging features in the diagnosis of breast cancer; however, it can offer significant advantages for lesions that are difficult for radiologists to differentiate.Therefore, further studies to evaluate the performance of AI in the diagnosis of atypical breast lesions are warranted.Furthermore, the combination of multiple imaging modalities may provide beneficial reference resources for clinical decisions.
AI has also undergone rapid development in medical fields other than medical imaging in recent years.Imaging data, pathological sections, and gene sequencing of patients have become important prerequisites for the accurate diagnosis and treatment of tumors.However, most AI models that have been proposed by researchers to date are based on a single imaging system and lack model training in combination with information from other imaging methods as well as information from electronic medical records.Therefore, the integration of this information with AI models is of great importance for the development of individualized treatment strategies.To this end, AI is expected to be incorporated into clinical practice and to become routinely used by clinical workers.

Conclusions
In this article, we have justified the necessity of introducing AI techniques into breast imaging, reviewed the applications of AI in breast imaging modalities, and presented technical and clinical challenges in this area.The key conclusions can be summarized as follows: (1) Breast imaging AI is clinically necessary and practically feasible in the era of precision medicine and personalized treatment.(2) The expectation should be for AI to aid radiologists in dealing with difficult cases, rather than to replace radiologists, in the diagnosis of breast cancer.(3) The future integration of multiple imaging modalities as well as radiomics with clinical data and multi-omics is warranted.

Figure 1 .
Figure 1.Schematic outline of the article.

Figure 2 .
Figure 2. Digital MG, US, MRI and PET-CT scans of a 42 y old woman with an infiltrating ductal carcinoma in the left breast, indicated with the white arrows.(a): MG, (b): US, (c): MRI, (d): PET-CT.
(FUSCC) subtype of TNBC has been established based on proteomics (Zhao et al 2020, Gong et al 2022), transcriptomics (Liu et al 2016), genomics (Jiang et al 2019), and metabolomics (Gong et al 2021).It has been proven that the combination of AI imaging and multiple omics may achieve an FUSCC subtype more rapidly and easily (Jiang et al 2022a).

Figure 3 .
Figure 3.The relationships between AI, ML, DL and CNN.They can be used in CAD with three common methods: supervised learning, unsupervised learning and reinforcement learning.

Figure 4 .
Figure 4. Workflow of AI methods.Outputs of the AI model are finally used to provide information regarding the detection, diagnosis, and therapy response, etc.

Figure 5 .
Figure 5.The architecture of the classification model based on MG to predict pCR in breast cancer patients.Reproduced fromfrom Skarping et al (2022).CC BY 4.0.

(
Han et al 2017, Xiao et al 2018, Byra et al 2019, Choi et al 2019, Ciritsis et al 2019, Fujioka et al 2019, Hejduk et al 2022).The performance of AI models varies significantly as a result of the different dataset sources and algorithms that are adopted.Most studies have affirmed the auxiliary diagnostic value of AI models for US physicians.A commercial US system that incorporated an AI module was launched and exhibited promising results, which further confirmed the clinical potential of AI technology(Kim et al 2017, Di Segni et al 2018).

(
Guo et al 2018b, Wu et al 2021, Zhou et al 2021a, Jiang et al 2021b), and histopathological variables (Cui et al 2021, Li et al 2022a) of breast cancer, as well as axillary lymph node metastasis (Sun et al 2020, Zheng et al 2020, Zhou et al 2020).
al 2019, Guo et al 2020, Zheng et al 2020, Lee et al 2021a, 2021b, Zhou et al 2021c, Jiang et al 2022b).Although the performance of these models is acceptable, it has been suggested that the prediction should also consider clinicopathological features to achieve satisfactory performance (Guo et al 2020, Zheng et al 2020, Lee et al 2021a).Meanwhile, the perineural region (Moon et al 2017, Sun et al 2020) and elastography (Jiang et al 2022b) are also valuable for predicting the ALNM status.2.3.4.Prediction of treatment response and recurrence US has traditionally been the major screening and diagnostic tool for breast cancer.However, an increasing number of studies have demonstrated that US features, particularly radiomics features, are potential imaging biomarkers for predicting the treatment response to NAC (Byra et al 2021, Jiang et al 2021a, Gu et al 2022) and the risk of postoperative recurrence of breast cancer (Xiong et al 2021, Yu et al 2021a, Sheng et al 2022).
2.4.1.Detection and classification AI based on breast MRI has mainly been used to aid in classifying breast lesions as benign or malignant (Zhang et al 2020, Potsch et al 2021, Sun et al 2021, Jiang et al 2021c, Altabella et al 2022, Daimiel Naranjo et al 2022, Militello et al 2022).Most of these studies achieved comparable classification efficacy to radiologists using an ML model (Daimiel Naranjo et al 2022) or a DL CNN model (Truhn et al 2019, Chung et al 2022, Witowski et al 2022).Ultrafast MRI, which reduces the image acquisition and interpretation time, has attracted increasing attention in recent years (Jing et al 2022).The DL model can be used for the automatic identification of normal scans in ultrafast breast tissue, thereby greatly decreasing the MRI screening time and costs (Ayatollahi et al 2021, Jing et al 2022).Figure 6 shows the process of using DL to exclude lesions with ultrafast breast MRI to shorten acquisition and reading time.AI has also been used to diagnose clinically challenging lesions, such as non-masslike lesions, sub-centimeter lesions, and lesions in patients with dense breasts (Lo Gullo et al 2020, Verburg et al 2022, Wang et al 2022a).Another application of AI is the classification of the pathological or molecular subtypes of breast cancer.Multiclass molecular subtype differentiation is a substantially more challenging task than diagnosis.Many studies have explored the potential of radiomics or DL models to classify breast cancer subtypes (Zhang et al 2021b, Zhou et al 2021b, Lee et al 2022, Tsuchiya et al 2022, Yin et al 2022, Sun et al 2022b, Lafcı et al 2023).In addition to the molecular subtype, MRI AI can classify the Ki-67 expression and histological grade, which are

Figure 6 .
Figure 6.(A) schematic illustration of the ultrafast breast DCE-MRI classification system, which includes three main stages: breast region segmentation, maximum intensity projection generation and abnormality prediction.Reproduced from Jing et al (2022).CC BY 4.0.

Table 1 .
CNN framework-based AI studies in MG for screening and detection of breast cancers.

Table 2 .
Selected AI studies in breast US regarding to tumor differentiation, biological property evaluation and prognosis prediction.

Table 3 .
Selected AI studies in breast MRI regarding to tumor detection, classification and prediction.factors in estimating the biological behavior and treatment sensitivity (Liu et al 2021, Song et al 2021, Zhang et al 2022, Fan et al 2022b).2.4.2.Prediction Similar to the application of AI in breast US, many investigators have evaluated AI techniques for breast MRI in predicting the ALNM status (Yu et al 2020, Zhang et al 2021a, Yu et al 2021b, Gao et al 2022, Zhan et al 2022, Li et al 2022b, Wang et al 2022b, Li et al 20223) as well the response to NAC (Braman et al 2017, Banerjee et al 2018, Braman et al 2019, Liu et al 2019b, Bitencourt et al 2020, Sutton et al 2020, Choudhery et al 2022, Massafra et al 2022, Caballo et al 2023).A recently published study explored the potential of four-dimensional (4D, 3D + time) ML radiomics based on spatiotemporal information from pretreatment DCE-MRI to identify patients who achieved pCR following NAC (Caballo et al 2023).Although AI techniques are unlikely to replace invasive biopsies, they offer the advantage of providing prognostic information that is derived from the entire tumor, whereas biopsy sampling only represents a small part of the tumor.This may be particularly useful for monitoring biological changes during treatment.AI-enhanced MRI has been investigated as a noninvasive predictor of breast cancer prognosis (Eun et al 2021, Ma et al 2022, Thakran et al 2022, Fan et al 2022a, Chen et al 2022c). important

Table 4 .
Selected AI studies in breast PET-CT or PET-MRI regarding to tumor detection, classification and prediction.