In recent years, AI has become a ubiquitous term that is being used in a wide spectrum of domains with remarkable betterment. However, to adopt AI (or GAI) in the clinical setting, the system (so-called AI system) must be first defined and characterized, depending on the intended purposes associated with different risk categories. Defining AI systems is important, as this outlines clear boundaries and expectations for what the system is designed to do, how it operates, what the expected outcomes are, and the potential risks associated with it. Risk assessment is crucial, allowing users to understand, anticipate, and mitigate potential negative outcomes that could arise from AI usage, involving critical domains, for example, ethical consideration, safety and security, trust and public confidence, prevention of harm and legal risks, data privacy and security, as well as robustness and reliability. With the introduction of the AI Act, manufacturers are now required to identify which AI systems fall within the regulation scope and are consequently subject to act compliance and obligations. Table 1 shows the list of definitions of AI systems with respect to different sources.

Fig. 3 shows the classification of models in GAI, highlighting the respective architecture component and training method in each model. The architecture component defines how a model processes information and subsequently generates outputs, while the training method shapes the performance and effectiveness of a model.

Briefly, the architecture of VAE is based on an encoder-decoder structure and uses variational inference in the training process [37,38]. VAEs are designed to learn compressed representations of input data by mapping it to a latent space. In the latent space, new samples can be generated by sampling and decoding from the learned distribution. A key highlight of VAEs is that the model incorporates a probabilistic framework, allowing them to generate diverse samples by drawing from a distribution rather than a fixed point, thus, making VAEs particularly suitable for tasks where a variety of output options is desirable, such as generating new images, reconstructing missing data, and/or performing anomaly detection.

GAN consists of two main components, namely, the generator and the discriminator. The primary role of the generator is to produce synthetic samples, for example, images, to deceive the discriminator, which evaluates whether a given sample is real or fake [39]. These two networks are trained adversarially, with the generator improving its output to fool the discriminator, and the discriminator improves capability in detecting fakes, formalizing a competition between the two leads (i.e., generator and discriminator), ultimately yielding highly realistic and diverse data. GANs have proven effective in areas, for example, image generation, video synthesis, and music composition. Despite their impressive results, GANs are known for being challenging to train, often requiring careful balancing between the generator and the discriminator to avoid issues like mode collapse.

Diffusion models involve a unique approach that starts with noisy data and progressively refines it through iterative denoising steps, which generate high-quality samples [41]. These models are trained by learning the dynamics of this noise-diffusion process, making them highly effective at tasks requiring precise, fine-grained detail, such as high-resolution image synthesis. The noising step inputs random noise into the data, and the denoising step gradually reconstructs the original signal through multiple iterations. Diffusion models have recently gained attention for their ability to outperform GANs in certain image generation tasks, specifically, in tasks in which the generation of detailed textures is required and successfully addresses some of the training instabilities seen in GANs.

Transformers employ an encoder-decoder architecture and make extensive use of self-attention mechanisms, which allow the model to capture long-range dependencies within data sequences [42]. Transformers are specifically effective in tasks that involve sequence data, for example, natural language processing, where the generation of coherent text or translation between languages is involved. The self-attention mechanism helps the model focus on relevant parts of the input sequence, making it highly scalable and efficient for tasks that involve handling large datasets with complex relationships. Notably, transformer models like Generative Pre-trained Transformers (GPTs) have demonstrated groundbreaking performance in text generation, translation, and summarization tasks.

Language models are typically built using RNNs or Long Short-Term Memory (LSTM) networks, which are designed to handle sequential data by predicting the next token in a sequence [43]. These models are trained through supervised learning and are specifically well-suited for generating natural language text, for example, completing sentences or writing paragraphs. The ability of language models to predict the next word in a sequence makes them crucial in applications like chatbots, translation systems, and automated content creation. With the rise of transformer-based models, language models have achieved even greater performance in generating coherent and contextually accurate text.

Normalizing Flow models use a sequence of invertible transformations, namely coupling layers, to transform data into a simpler distribution, for example, Gaussian distribution [44]. The key highlight of the normalizing flows is that the model preserves the exact probability density, allowing for accurate learning of complex distributions. By retaining density information in the data transformation process, this model is particularly suitable for tasks, for example, density estimation and probabilistic modeling. Normalizing Flow models are trained using maximum-likelihood estimation, making them useful in applications that require precise probability estimates, for example, anomaly detection or uncertainty quantification.

Hybrid models combine elements from multiple GAI models, allowing them to leverage the strengths of different models [40]. For example, a hybrid model can integrate the probabilistic sampling capabilities of VAEs with the adversarial training of GANs to achieve both diversity and realism in generated samples. By combining different architectures and training methods, hybrid models offer flexibility and can be tailored to meet specific generative goals, whether for image synthesis, text generation, or other creative tasks. These models are often designed to overcome the limitations of individual architectures, providing a more robust solution to GAI challenges.

Motivated by vigorous advancement in computation engineering and tremendous growth in world interest using GAI in breast cancer research, here, a systematic bibliographic survey focusing on GAI in breast cancer is presented. To ensure optimal search, no limiter is set on the publication year. The primary research questions of this study are: “What are the applications and impacts of generative AI in breast cancer diagnosis and treatment?” and “How has generative AI been utilized to enhance outcomes in breast cancer research, particularly in predictive modeling, diagnostics, and therapeutic development?”. The research question is framed as such to ensure the inclusion of all aspects of GAI models available within breast cancer. The present study is intended to focus on collating and synthesizing new insights and drawing constructive conclusions while highlighting the research gaps and challenges in the topic of interest. This study is organized as follows: Section 2 offers an overview of breast cancer diagnosis and treatment. Section 3 details the methodology employed in this systematic review. Section 4 presents synthesized findings from the included studies. Section 5 discusses self-assessment, limitations, challenges, and future directions. Finally, the conclusion is presented in Section 6.

The proposed search strategies are the output of collaborative discussion from a team comprised of engineering and medical experts. As the present study focuses on generative AI in breast cancer, the search strings included two elements: “generative AI” and “breast cancer”. These elements are intentionally formulated in a broad manner to ensure optimal inclusion of data from all aspects (including diagnosis and treatment) that primarily adopt/integrate generative AI in breast cancer. The search strategies included limiters, for example, English language and journal articles. No limiter is set on publication years to maximize the data retrieved. The ‘‘AND” and ‘‘OR” operators were used to formalize the search string in the present study (see Table 2). The systematic literature searches were last performed in December 2024 in four core databases: Scopus, Web of Science Core Collection, and PubMed, in lieu of search engines provided by specific publishers, such as SpringerLink, ScienceDirect, Multidisciplinary Digital Publishing Institute (MDPI), and Frontiers. Mendeley reference management software was used for reference management and related purposes.

Qualitative and quantitative data from each included article were carefully distilled and systematized into a functional summary endpoint. Qualitative data, for example, type of modalities, dataset, methods, and findings were extracted, compiled, and tabulated using a table. Quantitative data, for example, statistical findings, co-occurrence analysis, and bibliometric information, were collated and systematized to offer a holistic view of the topic of interest. Charts, graphics, and tables were used as a medium to synthesize the mineable data. For all included articles, the affiliation and author name were manually disambiguated (for example, “Harvard Medical School” and “HMS” are both referring to the same affiliation) to avoid duplication in the analysis using software such as VOS Viewer [50].

Qualitative and quantitative findings from the included works were reported narratively via the descriptive approach, supported using charts, graphics, and tables. Statistical calculations and compilation of data were done using Microsoft Excel 2019. Co-occurrence analysis and bibliometric findings were generated using VOS Viewer 1.6.20 software for Windows [50]. Quantitative findings, for example, classification of included works was illustrated using a pie chart; geographical distribution was illustrated using a world map chart; distribution of most contributing journals and publishers using bar charts; temporal publication analysis using a line graph; subject profiling using a treemap; and keywords co-occurrence, co-authorship occurrence, and country-ship analysis using bibliometric networks.

The confusion matrix is one of the most commonly used performance metrics from the body of the literature, which is specifically useful in summarizing the classification performance of a classifier with respect to the testing/validating data. Typically, the confusion matrix consists of a two-dimensional matrix, where one axis represents the true class of an object, and the other represents the class predicted by the classifier. Table 3 provides an example of a confusion matrix for a three-class classification task with classes A, B, and C. The first row of the matrix indicates that 15 objects belong to class A, with 12 correctly classified as A, one misclassified as B, and two as C.

In binary classification tasks, a simplified version of the confusion matrix is often used, where one class is designated as positive and the other as negative. The four cells of the matrix represent true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN), as shown in Table 4. Here, TP refers to the correctly predicted positive events, TN refers to the correctly predicted negative events, FP indicates the incorrectly predicted positive events, and FN represents the incorrectly predicted negative events.

From the obtained TP, TN, FP, and FN, performance metrics, for example, Accuracy, Recall/Sensitivity, Precision, Specificity, F1-score, and Area under the curve (AUC) can be computed. High values denote a better performance and are preferable in the study. The equations for Accuracy, Recall/Sensitivity, Precision, Specificity, F1-score, and AUC are as follows (Eqs. (1)–(6)), where d in Eq. (6) denotes the differential element (the small change) in the false positive rate (i.e., 1-Specificity). (1)Accuracy=TP+TNTP+TN+FP+FN (2)Recall/Sensitivity=TPTP+FN (3)Precision=TPTP+FP (4)Specificity=TNTN+FP (5)F1−score=2∗Recall∗PrecisionRecall+Precision (6)AUC=∫01Recall(1−Specificity)d(1−Specificity)

The p-value is a metric used in hypothesis testing to assess the statistical significance of an observed effect or outcome. It represents the probability of obtaining a test statistic as extreme as the one observed, assuming that the null hypothesis is true. The formula for calculating the p-value depends on the statistical test being used (e.g., z-test, t-test, chi-squared test, etc.). Eqs. (7)–(9) show the formula for calculation of the two-tailed test (commonly known as the z-test or t-test), one-tailed test, and chi-squared test, respectively. (7)p=2×P(Z≥∣zobs∣)where zobs is the observed test statistic; P(Z≥∣zobs∣) is the probability that a standard normal random variable exceeds the absolute value of the observed z-score (area under the tail). (8)p=P(Z≥zobs)where zobs is the observed test statistic; P(Z≥zobs) is the probability corresponding to the test statistic under the null hypothesis. (9)p=P(x2≥xobs2)where xobs2 is the observed chi-squared statistic; P(x2≥xobs2) is the probability under the chi-squared distribution. The p-value test is commonly used to determine whether to reject the null hypothesis (i.e., H0) in favor of the alternative hypothesis (i.e., H1). A small p-value, typically ≤0.05, indicates that the observed data is unlikely under the null hypothesis, suggesting that the null hypothesis may be false, and the alternative hypothesis may be more plausible. Whereas, a large p-value, typically >0.05 suggests that the observed data is consistent with the null hypothesis, hence, lack of evidence to reject the hypothesis.

Mean Squared Error (MSE) is a widely used metric for assessing the performance of regression models. It calculates the average of the squared differences between the actual values and the predicted values from the model. The MSE gauges how accurately the predicted values align with the true data points. It is always non-negative, with lower values indicating a better fit of the model to the data. Conversely, a higher MSE implies that the model’s predictions are further from the actual values. The formula for MSE is shown in Eq. (10). (10)MSE=1n∑i=1n(yi−y^i)2where n denotes the number of data points (sample size); yi denotes the actual/true value for the i-th data point; y^i denotes the predicted value for the i-th data point; ∑i=1n(yi−y^i)2 denotes the sum of the squared differences between the actual/true and predicted values. The difference calculation is computed for each data point, i, by calculating the difference between the actual value, yi, and the predicted value, y^i. The difference value is then squared to eliminate negative values and to give more weight to larger errors. The sum of these squared differences is averaged by dividing by the total number of data points, n, giving the mean squared error. Because the MSE equation involves a square of the differences between actual/true and predicted values, the MSE is then sensitive to outliers. MSE is often used to compare different models or algorithms by seeing which has the lowest error. A smaller MSE indicates better performance, with 0 being the ideal value, denoting no difference between predicted and actual/true values.

Intersection over Union (IoU) is a metric frequently used to assess the performance of object detection, segmentation, and other tasks that require evaluating the spatial overlap between predicted and ground truth regions. It quantifies the overlap between two areas, such as the predicted bounding box (or segmentation mask) and the ground truth bounding box (or mask). Eq. (11) presents the IoU formula, where the Area of Overlap represents the region where the predicted and ground truth areas intersect (i.e., the shared portion of both regions), and the Area of Union refers to the total area covered by both the predicted and ground truth regions, excluding the overlap (i.e., the combined area of both regions). (11)IoU=AreaofOverlapAreaofUnion

The ideal IoU value is 1, indicating that the predicted region exactly matches the ground truth. To the opposite extreme (IoU of 0), the equation implies no overlap between the predicted and ground truth regions. Higher IoU is preferable in all cases, indicating better alignment between the predicted and actual regions, with values closer to 1 being ideal.

Mean IoU is an extension of IoU, mainly used for multi-class segmentation tasks. The Mean IoU computes the IoU for each class in a dataset and then takes the average over all classes. This gives a more comprehensive evaluation metric when dealing with multiple objects or regions across several classes. Eq. (12) shows the formula of Mean IoU, such that C denotes the number of classes; Area of Overlap for Class i and Area of Union for Class i denote the overlap and union areas for class i across all images or regions in the dataset. (12)MeanIoU=1C∑i=1CAreaofOverlapforClassiAreaofUnionforClassi

The Mean IoU calculates the IoU for each class independently and then averages the results across all classes. The equation is particularly useful in semantic segmentation, where you need to evaluate how well the model predicts boundaries across multiple classes. Higher Mean IoU indicates better segmentation performance, with values closer to 1 indicating near-perfect segmentation.

The Intraclass Correlation Coefficient (ICC) is a statistic used to assess the reliability or agreement between measurements made by different observers, instruments, or measurement methods. ICC is especially useful when the measurements are made on the same subjects or items and can be used to evaluate consistency or conformity within groups of measurements. The ICC is commonly used in domains, for example, medicine, psychology, and medical image analysis, where multiple measurements or ratings are obtained for the same subjects under different conditions.

Eq. (13) shows the general form of the ICC formula, such that ICC (1,1) assesses absolute agreement between raters or measurements; ICC (2,1) assesses consistency across raters when raters are considered interchangeable; and ICC (3,1) assesses reliability when specific raters or instruments are used and generalization to other raters is not desired. (13)ICC=BetweensubjectvarianceBetweensubjectvariance+Withinsubjectvariancewhere “Between subject variance” denotes the variation between different subjects or items being measured, and “Within subject variance” denotes the variation between repeated measurements or raters for the same subject. ICC is mainly used in measuring the consistency or reliability of a measurement across different raters or instruments or to determine the level of agreement across two or more raters who provide ratings or scores for the same subjects.

The Peak Signal-to-Noise Ratio (PSNR) is a commonly used metric in image processing to assess the quality of a reconstructed or compressed image in relation to its original version. It measures the degree of distortion caused by lossy compression or other image processing methods by comparing the original and modified images. PSNR is particularly valuable for evaluating image compression algorithms, such as those used for image restoration through denoising or super-resolution techniques. The formula for PSNR is shown in Eq. (14). (14)PSNR=10⋅log10MAX2MSEwhere MAX denotes the maximum possible pixel value of the image. PSNR helps determine how much distortion or noise is introduced when compressing an image and is extremely useful in measuring the effectiveness of methods, for example, denoising, deblurring, and other image reconstruction techniques. A high PSNR value indicates that the processed image is closer to the original image, with less noise or distortion, whereas a low PSNR value implies that the processed image has higher distortion/noise, making it less similar to the original.

The Structural Similarity Index (SSIM) is a metric used to evaluate the similarity between two images. Unlike traditional methods, for example, MSE or PSNR, which focus on pixel-level differences, SSIM assesses image quality by considering structural information, resembling human visual perception. It compares the luminance, contrast, and structure of the images to determine their similarity. SSIM is especially beneficial in image processing tasks like compression, denoising, and enhancement, where preserving structural integrity is essential. The formula for SSIM is presented in Eq. (15). (15)SSIM(x,y)=(2μxμy+C1)(2σxy+C2)(μx2+μy2+C1)(σx2+σy2+C2)where μx and μy denote the means (average pixel intensity) of images x and y, representing luminance; σx2 and σy2 denote the variances of x and y, representing the contrast; σxy denotes the covariance between x and y, representing the correlation in structure between the images. When the SSIM equals to 1, the images are identical in terms of luminance, contrast, and structure; when the SSIM equals to 0, no structural similarity between the images. Thus, an SSIM value closer to 1 implies higher structural similarity, denoting that the processed image retains more of the original’s structural features.

The Kappa Value, also known as Cohen’s Kappa, is a statistical metric used to measure the degree of agreement between two raters or evaluators who categorize items into distinct groups. It is particularly valuable when assessing how well two different observers align in classifying the same items while accounting for the likelihood of agreement occurring by chance. Cohen’s Kappa quantifies the level of agreement between the two classifications, offering insight into whether the observed agreement surpasses what would be expected randomly. This metric is widely used in fields like psychology, medicine, and social sciences to assess the reliability of observers or raters. The formula for Cohen’s Kappa is shown in Eq. (16). (16)Cohen′sKappa=P0−Pe1−Pewhere P0 denotes the observed agreement, computed based on the proportion of times the raters agree (the total number of agreements divided by the total number of observations); Pe denotes the expected agreement, computed based on the proportion of times the raters would be expected to agree by chance. The Pe is calculated based on the relative probabilities of each category being assigned by the raters. Cohen’s Kappa commonly falls within the range of [0, 1], such that a value close to 1 is preferable, denoting perfect agreement between raters.

The Dice Coefficient, or Sorensen-Dice Index, is a statistical metric used to evaluate the similarity between two sets. It is widely used in image analysis and medical imaging to assess the accuracy of image segmentation and classification. The Dice Coefficient ranges from 0 to 1, where 0 signifies no overlap between the sets, and 1 denotes perfect overlap. The formula for the Dice Coefficient is shown in Eq. (17). (17)DiceCoefficient=2|A∩B||A|+|B|where A denotes the set of predicted positive samples; B denotes the set of ground truth positive samples; |A| denotes the number of elements (or pixels) in set A; |B| denotes the number of elements (or pixels) in set B; and |A∩B| denotes the number of elements (or pixels) that are common to both sets A and B.

Fréchet Inception Distance (FID) is a metric used to assess the quality of images produced by generative models, such as GANs. It calculates the distance between the distributions of real and generated images in a feature space, which is derived from a pre-trained Inception network, a specific type of convolutional neural network. The FID score evaluates both the quality and diversity of the generated images, making it a commonly used metric for evaluating generative models. The formula for FID is shown in Eq. (18). (18)FID=‖μr−μg‖2+Tr(Σr+Σg−2(ΣrΣg)1/2)where μr and μg denote the mean vectors of the features extracted from the real images and generated images, respectively; Σr and Σg denote the covariance matrices of the feature representations for the real and generated images, respectively; Tr denotes the trace of a matrix, which is the sum of the diagonal elements. The FID score can be ranged from 0 to infinity, where lower values indicate better performance. A score of 0 means that the generated images perfectly match the real images in the feature space. It is important to remark that the FID is sensitive to both the quality of generated images and the respective diversity. The FID can effectively capture mode collapse scenarios, where a model generates limited variations of images, leading to a higher FID score.

Based on Fig. 4, the search string, as formalized in Table 2, resulted in 31 included articles. These articles were then characterized into two generic categories, namely diagnosis and treatment. Notice that one included article reported findings using GAI in both the categories. Henceforth, a new category, namely Diagnosis & Treatment (i.e., hybrid category) is created. Fig. 5 shows the classification of the 31 included articles. According to Fig. 5, the diagnosis category formalized the main body of the literature, accounting for 90.3% (28 articles) of the included articles. This is followed by articles in treatment and hybrid categories, respectively accounted for 6.5% (2 articles) and 3.2% (1 article). This could be attributed to the superiority of GAI in accomplishing task-specific events [8], for example, data generation and augmentation [22], abnormalities detection [51,52], image-to-image translation [53,54], which formalized the core foundation in computer-aided diagnostics.

In the treatment domain, personalized regimens may be required most of the time. The capability of GAI in the detection of non-clinical significance can result in overtreatment [12], subsequently leading to unnecessary clinical procedures and associated to morbidity [55], leading to lower interest in using GAI for breast cancer treatment purposes. Table A1 in Appendix B summarizes the details of the 31 included articles, comprised of useful information, for example, the generic category, modalities, dataset, methods, and core findings.

Fig. 6 shows the geographical scientometric analysis of the included articles. The geographical distribution of the contributing organizations in GAI in breast cancer research demonstrates a wide global engagement, with notable concentrations in certain regions. The USA leads with 8 contributions, highlighting its dominant role in AI-driven breast cancer research. China and India each follow with 4 contributions, indicating strong research activity in Asia. Other Asian countries such as Japan and Korea also show significant participation, with 3 and 2 contributions, respectively. European countries like Germany, Spain, Switzerland, and France collectively contribute 7 works, showcasing the region’s involvement in integrating AI with medical research. While Australia, Bangladesh, and the Netherlands each contribute 1 study, this distribution indicates active participation from both developed and developing nations. This wide geographical spread underscore the global importance and collaboration in leveraging AI for breast cancer diagnosis and treatment across diverse healthcare systems. To better illustrate the collaboration between different countries, a co-country-ship network analysis is generated, as in Fig. 7. Because of small number of included articles in this study, the co-country-ship network may look scatted. The network provided however remain valid in illustrating the collaborative relationship between the countries as aforementioned. The network reveals three distinct clusters, dominated by the USA, China, and Italy, with the USA at the core of the contributions. This is followed by China, establishing a close network with the United Kingdom. The third cluster features Italy, collaborating with Switzerland, forming the third strong partnership in the network.

Fig. 8 shows the most contributing journals in reporting findings using GAI in breast cancer. Overall, a total of 29 journals were collated from the 31 included articles. Based on Fig. 8, insufficient evidences were found to substantiate or to support the argument on the most contributing journals herein. This is because the distribution of journals is found relatively scarce corresponding to the entire body of the literature. From the collected data, the journal, namely Journal of Medical Imaging and Dignostics, both ranked as the most contributing journals, but accounted only 6.5% (two articles), respectively, from the included articles. The scattered distribution may be due to the broad applicability of “diagnosis,” which aligns with the aims and scopes of various journals, for example, those shown in Fig. 8. It is important to note that diagnosis in breast cancer occurs at multiple stages, for example, during the examination phase using mammograms or during grading using histopathology images. These diverse diagnostic processes fit within the scope of numerous journals, leading to their widespread inclusion.

As mentioned in Section 3.1, general databases, for example, Scopus, Web of Science Core Collection, and PubMed, were used for the literature search process. From all the included articles, 14 publishers were retrieved, such that the top five most contributing publishers are Springer, Elsevier, MDPI, SPIE, and Wolters Kluwer. Fig. 9 summarises the contributing publishers retrieved from the included articles. From the figure, Springer appeared to be the most contributing publisher, accounted 22.6% (7 articles) in total. This is followed by Elsevier (19.3%, 6 articles), MDPI (12.9%, 4 articles), SPIE (9.6%, 3 articles), and Wolters Kluwer (6.5%, 2 articles). In general, publishers maintain an oligopolistic hold on the publishing industry, a trend consistent with previous findings from an analysis spanning four decades (1973–2013). This analysis identified Elsevier, Taylor & Francis, and Springer as leading publishers across various topics. The same trend as well can be observed from some of the recent reviews in various fields, for example, agriculture, engineering, applied mathematics, and medicine. The oligopolistic structure of the publishing industry is likely to persist, as building a reputable publishing company requires significant time and resources. Furthermore, researchers may hesitate to submit their work to newer publishers to avoid the risks associated with potential predatory publishers.

Fig. 10 shows the temporal scientometric analysis for the included articles. In order to maximize the data retrieved pertaining to the proposed search string, no limiter is implemented on the publication year. Based on the figure, the temporal distribution demonstrates a continual increment in the topic of interest, with the first relevant publication in the year 2007. Relevant publication in GAI, specifically in breast cancer was not evident from the years 2008 to 2015, with limited growth observed. This slow start can be attributed to the early stage of AI development, where machine learning and AI applications in healthcare were still nascent, and there was limited access to large datasets, computing power, and expertise. However, from 2018 to 2022, there was a gradual increase in publications, reflecting the rise of deep learning techniques, particularly with the introduction of GAI in 2014. These advancements allowed for greater exploration in medical imaging, including tasks like cancer diagnosis, image segmentation, and data augmentation. The availability of better computational resources, such as GPUs, and the development of pre-trained models further fueled research efforts, leading to an increase in publications. The most significant surge occurred in 2023 and 2024, with the number of publications rising sharply to 11 and 9, respectively. This dramatic increase is likely due to the growing adoption of advanced generative AI models, such as YOLO and transformer-based architectures, which have proven highly effective in medical imaging applications. Researchers are increasingly recognizing the potential of AI in enhancing breast cancer diagnosis, classification, and even generating synthetic datasets for training purposes. This surge also reflects the broader trend of AI integration in healthcare, as interdisciplinary collaborations between AI researchers and medical professionals become more common, leading to more impactful studies and publications. Overall, the trend indicates that GAI’s role in breast cancer research is rapidly expanding, with further growth expected as AI technologies continue to evolve and find new applications in the medical field.

Fig. 11 shows the subject areas profiling retrieved from the included articles in this study. Based on the figure, the subject area profiling for the 31 included works in GAI applied to breast cancer reveals a multidisciplinary research trend. Medicine is as the most dominant field, particularly in Radiology, Nuclear Medicine, and Imaging. This is then followed by Oncology, a sub-field of Medicine, reflecting the critical role of AI in medical imaging and cancer diagnosis. The presence of Computer Science areas, for example, AI Applications, Computer Vision, and Pattern Recognition highlights the use of computational methods in image analysis and diagnostics. Engineering fields, for example, Biomedical Engineering and Electrical Engineering underscore the integration of technology for enhanced healthcare solutions. Biochemistry, Genetics, and Molecular Biology, with emphasis on Cancer Research, and Materials Science also play crucial roles, pointing to advancements in the understanding and treatment of breast cancer through biomaterials and molecular studies. This multidisciplinary approach showcases the convergence of medicine, computer science, and engineering in leveraging AI to improve breast cancer diagnosis and treatment.

Fig. 12 shows the thematic landscape of the included articles, focusing on GAI in breast cancer. The co-occurrence mapping of keywords reveals three distinct clusters, categorized by objects of interest, methodologies, and relevant sub-field approaches. The objects of interest cluster appear to be the most prominent cluster, indicated by the prevalence of keywords, for example, “breast,” “breast neoplasms,” and “breast tumour.” This is followed by the methodology cluster, characterized by keywords, for example, “artificial intelligence” and “controlled study.” The third cluster comprises keywords, for example, “image processing” and “machine learning”. Notably, the keyword “artificial intelligence” appears as the largest dot on the map, highlighting its widespread adoption across the included articles. This prominence is justifiable, given the literature search focuses on GAI in breast cancer. Additionally, the sub-field approaches, including “deep learning”, “generative adversarial networks”, and advancements in algorithms, signify significant innovations within the realm of artificial intelligence. From the perspective of objects of interest, the mapping indicates that topics such as “diagnostic imaging”, “mammography”, “breast neoplasms”, and “breast tumours” have garnered considerable attention from researchers over the past decades.

The present study is a systematic review that focuses on the diagnosis and treatment of breast cancer. To maximize the data retrieved, no limiter is set on publication years. Here, a structured review methodology was used to collate, analyze, and synthesize the findings, highlighting patterns, trends, and the contents of included works in the topic of interest. A structured search strategy was proposed in compliance with the PRISMA guideline as detailed in Section 3.2. To affirm the quality of this study, an appraisal tool, namely A Measurement Tool to Assess Systematic Review (AMSTAR) is used to assess the quality of the present study in view of the content validity (see Table 5). AMSTAR is a self-appraisal tool that is commonly used to assess the quality of a systematic review, specifically to determine if the systematic reviews are comprehensive, have proper referencing, and are equipped with added value to the readers. The AMSTAR comprises 11 components for content validity. According to Table 5, the present study obtained one “no”, 10 “yes”, and one “NA”. Justifications were provided for each component to better support the evaluation outputs, as recommended by [8].

The findings of this study are subjected to several limitations. First, only full-text articles available in the English language are included in the synthesis process in accordance with the PRISMA guideline. Non-English material and/or articles without full text would be excluded in Phase 1 of the exclusion criteria. This may potentially introduce bias into the analysis, affecting the description of patterns and trends of the included articles in the topic of interest. Second, based on the proposed search string (Table 2), only articles utilizing GAI in breast cancer are retrieved. Therefore, emerging GAI models that are reported in other cancer types that potentially offer useful insight and utility in breast cancer may be excluded. Third, articles that were not indexed and populated by Scopus, Web of Science Core Collection, and PubMed databases were not included in the synthesis process. Lastly, grey literature is not included in the synthesis activities. This is mainly because the grey literature may lack rigorous peer review, subsequently resulting in inconsistent quality, associated with risk of bias, and challenging for reproducibility. However, considering the diversity and structured search methodology (in compliance with the PRISMA guideline) adopted here in retrieving articles within the topic of interest, the present study is confident with the findings synthesized from the included articles, such that the majority of the relevant literature has been included, and the findings accurately represent the current state of research.

While recognizing GAI as the potential game-changer to meet the ever-increasing demand for quality healthcare in breast cancer, typically in the era of precision medicine, the technology is inevitable to inherit challenges in various spectrums for widespread clinical adoption. Awareness of the constraints of the current GAI-based systems and concerted efforts (from researchers and industry experts) are essential in pushing GAI technology to impact the future direction of breast cancer research.

Similar to all the discriminative models, the nature of data-driven in GAI-based systems is inevitable to bias [56,57], as this serves as the foundation for the generative models. Here, two biases are identified: data bias and algorithm bias. Data bias could potentially occur in the training phase when the training data comprises information from a specific cohort, while under-represented cohorts may be afflicted by the GAI outputs, assuming the output solutions fulfill the input prompts. For example, a GAI falsely identifies black patients as being healthier than white patients who suffered from the same illness when the cost of healthcare is used as the proxy. This is inaccurate as black patients receive less medical cost attention. Thus, injecting faulty perception into the training data, which subsequently reflects in the output solutions [58]. Likewise, countries with poor data registries, especially low-resource countries, are at risk of being under-represented in GAI-based systems, leading to skewed or inaccurate diagnostic and treatment recommendations [59,60]. If the generative model training data disproportionately reflects healthcare practices and/or patient demographics from high-resource countries, the GAI-based system may fail to adequately address the unique needs and/or health profiles of populations in low-resource regions. This can exacerbate existing healthcare disparities, as the outputs generated by such GAI models might misinterpret conditions prevalent in these under-represented cohorts. Consequently, bias in training data not only compromises the fairness and accuracy of GAI outputs but also raises ethical concerns about its application in global healthcare systems, particularly for vulnerable populations. Algorithm bias or mode collapse is defined as a situation where a GAI fails to capture the full diversity of the training data, resulting in repetitive or limited variations in the generated outputs [61]. Often, this bias originates from overfitting, where the generative models fail to accurately learn the underlying distribution of the training data, producing identical outputs with no variation [62].

Generalization and adaptation of GAI denote the model capability to produce a new and diverse output that is not directly replicated from the training data, implying the superiority of GAI in capturing the underpinning patterns and distributions of the training data and reflected in the output solution [63–66]. For example, a robust GAI shall demonstrate promising performance even when the input prompts are pertaining to breast histopathology images that are not presented in the training data. A well-generalized GAI can generate realistic and meaningful data across various scenarios, even when presented with inputs or conditions that have not been explicitly encountered during training. Generalization is crucial to avoid overfitting, ensuring that the model can generalize the patterns (from the training data) and adapt to unseen data while maintaining high performance in real-world applications. Considering the diversity of incoming patients (i.e., wide variety of races, ethnicity, nationality, and dietary backgrounds), GAIs that lack generalization are not reliable, increasing the risk associated, afflicting the diagnostic and treatment recommendations, ultimately, fragmenting the quality of healthcare service and exacerbate trust in GAI for clinical adoption. Adapting the GAI models to a new domain is now an ongoing challenge, where techniques, for example, fine-tuning, domain adaptation, and transfer learning are continued to be explored to address the present challenge.

Because of the generative nature of GAI, the model is susceptible to hallucination or confabulation, where the model itself expects a plausible output solution. In fact, the solution generated is unreasonable with respect to the training data [67,68]. The erroneous and/or misleading output solutions are not a result of bias, but the output solutions are factually incorrect, unrealistic, and/or completely unrelated to the training data. To date, the core reasons underpinning hallucination or confabulation are still mysterious, with early inference that training data embedded with fictional or contradictory content besides factual information may further promote the risk of hallucination [69]. This may suggest the importance of data cleaning (i.e., removal of outliers) in the training phase, to avoid skewed models due to disproportionate influence, distorted learning, and random noise resulting from the outliers. Considering the rich amount of training data in GAI, data cleaning is tedious and cumbersome however, especially for medical experts with mediocre knowledge of computational engineering and data analytics. Concerted efforts in data cleaning, data transparency, and algorithm error-checking mechanisms may be required to mitigate the aforementioned problem.

Transparency, explainability, and interpretability of an AI system in clinical settings are crucial, as medical decisions cannot be adopted solely from a “black box” model, which lacks sophisticated reasonings and justifications, with output solutions that are challenging to verify [70]. Unlike the discriminative model, the explainable model, for example, Explainable Artificial Intelligence (XAI) [71] in the generative model is still in its infancy with robust models yet to surface. Additionally, a recent study argues that the GAI (i.e., typically LLM) that gained interest across multiple spectrums due to its formal reasoning capabilities, particularly in mathematics may not be accurate. The findings argue the hypothesis with evidence proving the LLMs may not be capable of genuine local reasoning but, in fact, replicating the reasoning steps based on the observation in the training phase [72]. Transparency in GAI holds an important role, especially when multiple output solutions could be produced toward the same input prompts. Thus, sensemaking in fulfilling the input prompts is essential. To date, most of the GAIs are trained on a pre-trained model with respective fine-tuning performed in accordance with different circumstances or applications for local adaptation.

From a regulatory standpoint, the use of GAI-generated medical data introduces complex issues related to data privacy, intellectual property, and legal responsibilities, specifically, in the context of cross-border healthcare collaboration. Data privacy is a primary concern, as aforementioned, the data-driven GAI models often rely on large datasets, which may comprise sensitive patient information. Ensuring compliance with data protection regulations such as the General Data Protection Regulation (GDPR) in the European Union or the Health Insurance Portability and Accountability Act (HIPAA) in the United States becomes challenging in cross-border settings, where differing privacy standards may conflict. Additionally, the origins and handling of data used to train pre-trained models may lack transparency (i.e., black-box), raising ethical concerns about patient consent and data provenance. Intellectual property issues further complicate the landscape, as questions arise over the ownership of GAI-generated content, the rights to the original training data, and liability for errors or misuse of model outputs [73]. Events involving cross-border collaborations further exacerbate the situation, where intellectual property laws vary significantly between countries [74,75]. Legal responsibilities also present challenges, as determining accountability for adverse outcomes or misdiagnoses stemming from GAI-generated recommendations can be difficult, especially when models are developed in one jurisdiction and deployed in another. Differing healthcare regulations and liability frameworks across borders further complicate the establishment of clear accountability. Addressing these issues requires robust international frameworks to harmonize data privacy standards, clarify intellectual property rights, and establish legal accountability, ensuring the ethical and responsible use of GAI in global healthcare collaborations.

The adoption of GAI in clinical settings requires comprehensive technical support, planned preventive maintenance, and robust infrastructure readiness. This includes ensuring the availability of high-performance workstations and sufficient energy supply to handle the intensive computational demands of GAI models. Additionally, having skilled experts on hand is crucial to address daily technical difficulties and to provide manpower for routine maintenance and system checks. Cybersecurity is another critical aspect, as the integrity of GAI models must be safeguarded against potential hacking attempts that could compromise GAI accuracy and reliability [76]. Mitigation measures must also be in place to prevent malicious interference, for example, the introduction of noise or unauthorized alterations to the models, which could jeopardize healthcare services, afflicting the generative recommendations. Ensuring both technical and cybersecurity readiness is key to the safe and effective implementation of GAI in clinical environments. Considering all these requirements, ranging from infrastructure and energy to expert manpower and cybersecurity, the associated cost becomes a significant factor in the adoption process [77]. To keep healthcare affordable, GAI should not solely aim to eliminate occasional errors made by medical experts but should instead focus on fully automating specific procedures currently performed by experts. The primary goal of GAI in breast cancer research shall focus on enhancing clinical efficiency while avoiding unnecessary healthcare costs.

Misuse of GAI is not new, especially in domains involving entertainment and social media using techniques, for example, deepfakes. Considering the ease and robustness of the deepfake techniques in this era, the low cost of implementation, as well as sub-optimal legal boundaries, it is challenging to affirm a GAI is free from malicious actions if a comprehensive adversarial monitoring system is not in place [78]. Factors contributing to misuse of GAI in breast cancer, especially in diagnosis and treatment can be multi-factorial, especially involving enormous personal interests that may arise from conflicts, for example, malicious incidences involving insurance claims and tension between opting for optimal healthcare or medical profit. These underscore the vulnerability of GAI to malicious actions and exploitation, with potential threats that collectively hinder trust in GAI in clinical settings.

While GAI poses significant advancement and revolutionizes breast cancer research, typically in diagnosis and treatment stages, offering enhanced efficiency and productivity in clinical operations, awareness of the challenges of GAI adoption is important. The far-reaching risks of GAI, typically in bias, hallucination or confabulation, and misuse required concerted and sustained efforts from researchers and industry experts in developing comprehensive mitigation mechanisms and solutions to ensure responsible AI usage. The adoption of GAI can be performed in multiple stages, where continuous temporal assessment and quality improvement are implemented at different checkpoints [25], and allows interrogation and feedback from experts at different levels as well as consumers and the public, aiming for robust and reliable GAI adoption.

As GAI continues to evolve, its integration into clinical practice requires a multidisciplinary approach, addressing technical, ethical, legal, and practical considerations. Ensuring the successful deployment of GAI in breast cancer diagnosis and treatment will depend on collaborative efforts among researchers, policymakers, healthcare institutions, and clinical practitioners. This section outlines key focus areas and strategies for advancing GAI in healthcare.

From a researcher’s standpoint, the primary focus should be on developing GAI models that are robust against bias and capable of generalizing across diverse patient populations. This necessitates the creation of inclusive and representative training datasets, particularly from underrepresented cohorts and low-resource regions. Techniques such as domain adaptation, transfer learning, and federated learning should be explored to ensure GAI models perform reliably across varied demographics and healthcare settings. Additionally, XAI for generative models must be advanced to enhance clinicians’ trust in GAI-generated outputs. Developing interpretable models that provide clear justifications for their recommendations, along with visualization tools for decision-making processes, will be crucial for bridging the gap between AI and clinical practice. Furthermore, reducing hallucinations in GAI models is essential for improving reliability. This can be achieved through robust data-cleaning pipelines, outlier detection mechanisms, and error-checking algorithms that ensure the accuracy and consistency of generated solutions. Collaboration between AI developers and medical experts is critical to refining these processes and maintaining clinical relevance.

From a policymaker’s standpoint, establishing legal and ethical frameworks is essential for facilitating cross-border collaboration in GAI research and deployment. Regional and international alliances are necessary to create standardized regulations for data privacy, intellectual property, and legal accountability, ensuring the global adoption of GAI in healthcare. Open-access initiatives, such as DeepSeek, can serve as models for promoting transparency and equitable access to GAI technologies. Policymakers should also develop guidelines for data sharing, model ownership, and liability, ensuring that GAI systems are deployed responsibly and ethically. Effective collaboration among policymakers, researchers, and industry stakeholders is necessary to establish harmonized frameworks that balance technological innovation with patient rights and ethical considerations.

From a healthcare institution’s standpoint, a phased approach to GAI adoption should be considered for seamless integration into clinical workflows. Initially, GAI can be implemented in low-risk applications such as data augmentation, image synthesis, and preliminary diagnostic support, allowing healthcare providers to build confidence in the technology while minimizing risks to patient care. Establishing scalable and secure computational infrastructure is crucial for supporting the demands of GAI in clinical settings. This includes high-performance workstations, energy-efficient systems, and robust cybersecurity measures to protect sensitive patient data and maintain the integrity of GAI models. Additionally, comprehensive training programs and workshops should be provided to clinical practitioners to equip them with the necessary knowledge for effectively utilizing GAI technologies in medical practice.

From a clinical practitioner’s standpoint, collaboration with healthcare institutions is crucial in establishing strong data governance frameworks that comply with data privacy regulations such as GDPR and HIPAA. This includes verifying the provenance of training data used in GAI models and obtaining informed patient consent for data usage. Before fully integrating GAI into clinical workflows, clinicians should treat GAI-generated recommendations as supplementary tools, rather than definitive solutions. Rigorous validation of AI-generated outputs against clinical expertise and established guidelines is necessary to ensure patient safety and prevent over-reliance on AI systems. Clinical practitioners also play a vital role in advocating for the ethical use of AI by participating in policy discussions, addressing biases in AI models, and ensuring that GAI technologies prioritize patient welfare and equity. Continuous research and collaboration between clinicians and AI researchers are essential to refine GAI models and ensure their practical applicability in real-world healthcare settings.

Achieving full integration of GAI in clinical practice will require a multidisciplinary effort. Experts from various fields must exchange knowledge, break down research silos, and complement each other’s expertise to foster the development of innovative solutions. By addressing the technical, legal, and practical challenges, GAI has the potential to revolutionize breast cancer diagnosis and treatment, ultimately improving patient outcomes and advancing precision medicine.