Machine learning in radiology: current applications, challenges, and future directions—a narrative review

Introduction

Radiology continues to be a crucial component of contemporary medicine, which uses the most recent technological advancements in the discipline to both diagnose and treat illnesses (1). The most recent technological advancements in radiology have been included in recent years, which has assisted radiologists in obtaining higher-quality images (2). Additionally, artificial intelligence (AI) has just entered the field of radiology, significantly influencing its development (3-6).

Machine learning (ML), a technique that enables machines to “learn” from data and generate predictions without the need for programming, is a fundamental component of AI. ML has been used in radiology to predict outcomes, identify abnormalities, classify diseases, and recreate images (7-9). Prior research has focused on image-based tasks, conventional ML methods, and computer-aided detection systems. Tumour detection, organ segmentation, and other aspects of radiology image analysis have advanced significantly with the introduction of deep learning (DL) and convolutional neural network (CNN) (10). Critical impact areas such as emergency radiology, cardiology, and cancer have shown promise in its implementation (11).

But in the last 3 to 5 years, the field has experienced substantial changes. Radiological data processing, interpretation, and integration have been significantly influenced by the introduction of foundation models, self-supervised learning, federated learning, generative AI, and large language models (LLMs) (12,13). These models are moving from task-specific ‘narrow AI’ to more general-purpose, multimodal models that learn from large-scale heterogeneous data. Generative learning, including synthetic image generation and translation across modalities, may provide solutions for the scarcity and privacy issues of radiological data. On the other hand, LLMs are being explored for report generation and decision support (14).

Despite the substantial progress in the development of algorithms, the clinical application of ML in radiology is inconsistent. Many studies published show inconsistent results based on the data set used, with many studies based on retrospective studies and highly curated environments. Reproducibility, dataset issues, lack of regulation, interpretability, and workflow issues continue to hinder the adoption of ML algorithms in radiology (15-18). The disparity between the results obtained from research environments and the real-world application of the algorithms is often termed the “AI chasm”. This disparity needs to be evaluated based on the existing literature.

Several previous reviews have examined AI and ML in radiology with varying scopes and levels of analysis. Earlier studies have provided general overviews of AI applications in radiology and discussed their integration into medical imaging workflows, with emphasis on clinical adoption and system-level implications (3,4). Other reviews have focused more specifically on ML techniques in medical imaging, particularly DL approaches for image classification, detection, and segmentation tasks (8,19). More recent literature has begun to highlight broader concerns, including validation strategies, reproducibility, and challenges in clinical implementation, reflecting a shift toward longitudinal evaluation of AI in real-world settings. However, existing reviews still provide limited discussion on emerging paradigms such as foundation models, generative AI, and LLMs, and often do not comprehensively address issues related to generalizability, regulatory considerations, and deployment in routine clinical practice. Therefore, the present review aims to provide a comprehensive and updated synthesis that integrates both established and emerging ML approaches while critically evaluating their clinical applicability and future potential. Therefore, the objectives of this narrative review are:

• To comprehensively investigate the current applications of ML in radiology, including image acquisition, disease detection, segmentation, prognosis, and optimization.
• To critically evaluate the quality, validation, and generalizability of the current evidence.
• To investigate the emerging trends, including foundation models, generative AI, and the incorporation of LLMs.
• To recognize the challenges, including technical, ethical, and regulatory issues, which affect the application of ML.

In this review, we have attempted to combine both existing and new paradigms to offer an inclusive and forward-looking perspective on the changing role of ML in radiology. Compared to earlier reviews, which primarily focused on task-specific DL models, this review provides a broader synthesis that incorporates emerging paradigms and emphasizes validation, reproducibility, and real-world implementation challenges. Recent studies have also emphasized the need for longitudinal evaluation of AI models in clinical workflows, highlighting unresolved issues in generalizability and deployment. We present this article in accordance with the Narrative Review reporting checklist (available at https://tro.amegroups.com/article/view/10.21037/tro-25-47/rc).

Methods

Study design and search strategy

To systematically synthesize the existing body of knowledge concerning the application of ML in the field of radiology, the study was undertaken as a structured narrative review. Every effort was made to reduce any form of selection bias in the methodology used. An exhaustive literature search was performed using the online databases of PubMed, Scopus, Web of Science, and IEEE Xplore from January 1, 2015, to July 31, 2025. The last search date was on July 31, 2025. The main search strategy utilized Boolean searches involving the use of the following key phrases: (‘machine learning’ OR ‘deep learning’ OR ‘artificial intelligence’ OR ‘foundation models’ OR ‘generative AI’ OR ‘large language models’) AND (‘radiology’ OR ‘medical imaging’ OR ‘diagnostic imaging’). In an attempt to identify additional literature that could be relevant for the study, the references provided in the selected articles were also manually reviewed.

Eligibility criteria and study selection

Only research papers discussing the role of ML in diagnostic radiology, such as imaging acquisition/processing, disease detection/segmentation, prognosis, workflow, or current topics like foundation models and generative AI, were selected. Editorials, reviews, research papers without AI-based imaging, and papers without clear methodology were not considered. Two reviewers independently screened titles, abstracts, and full texts. Disagreements were resolved through discussions. Only English-language peer-reviewed studies, including retrospective, prospective, and experimental designs, were included. In addition to the included studies, supplementary references were used to support background concepts and discussion.

Data extraction and synthesis

The data collected from the eligible studies included the imaging modality, ML methodology, dataset characteristics, validation approach, performance metrics [e.g., area under the curve (AUC), sensitivity, specificity] reported, and limitations acknowledged. Particular attention was paid to the external validation, reproducibility, and potential sources of bias. Considering the wide range of studies and their results, a quantitative meta-analysis was not possible; hence, the results were synthesized through a narrative approach based on a thematic framework that included the following themes: diagnostic, prognostic modelling, workflow, emerging technologies, and challenges. A summary of the search strategy is presented in Table 1.

Background on ML in radiology

To improve the accuracy and effectiveness of the diagnostic procedure, radiologists have investigated the use of AI and ML. While ML is a subset of AI that uses algorithms to teach computers to perform tasks on their own without explicit programming, AI includes the methods that allow machines to carry out tasks that need human-like intelligence (20). Medical imaging has used ML to identify anomalies, categorize illnesses, and evaluate results. DL, which is a subset of ML, has been reported to have superior performance compared to ML for the detection and segmentation of tumours (8,10). The foundation models have the ability to distinguish themselves from conventional CNNs using a combination of transformer architecture and attention mechanisms that enable global feature modeling. As opposed to CNNs, which concentrate on local spatial features, foundation models are capable of incorporating images, clinical, and textual data into a single framework.

Distinction between AI, ML, and DL

While AI generally involves the replication of human intelligence, ML, which is a subset of AI, focuses on the creation of learning algorithms from data. DL, a subset of ML, is highly effective for tasks like segmentation and image identification because it uses deep neural networks to generate data representations (21,22). DL, unlike ML, which heavily relies on the extraction of data features, enables the direct learning of data features, enhancing the quality and efficiency of image analysis, which is very important in radiology, considering the complexity of the data (23).

Historical perspective on ML in radiology

Radiology first ventured into ML in the mid-1980s with computer-aided diagnosis (CAD) systems, which applied rule-based and statistical techniques to support diagnostics (24). ML algorithms developed for detecting various conditions, including lung nodules, formed the basis of more sophisticated algorithms developed later (8). With advancements in computing capabilities and digitized imaging in the 1990s–2000s, radiology was able to employ supervised ML techniques like decision trees and support vector machines, propelling ML applications in radiology (8).

Importance of data in ML applications for medical imaging

Two crucial elements that are crucial to ML for medical imaging are data quantity and quality. Data quality is necessary for making precise predictions, while poor quality or insufficient data may lead to overfitting, which is detrimental for ML algorithms (25,26). Medical imaging has its own challenges, including variability in image acquisition and the need for large datasets for better performance (27,28). For instance, in tasks related to tumour detection and classification, large datasets are necessary for better performance of DL algorithms (29). Additionally, multimodal data inputs, including imaging and clinical information, can enhance the robustness of ML algorithms.

Technological and computational advancements driving adoption

Significant advances in computing power and the advent of DL in the 2010s have transformed radiology, allowing for the automation of feature extraction and enhanced accuracy of diagnoses (6,30). For tasks like identifying diabetic retinopathy from retinal fundus images, CNNs have demonstrated near-human performance (31). Large datasets from various collaborations and health campaigns have enhanced the training of ML models (32), while the use of natural language processing (NLP) for the automation of data labelling has helped to increase the availability of data (33). These days, a number of ML applications could improve radiology accuracy (8,34).

Applications of ML in radiology: image acquisition and reconstruction

Image acquisition and reconstruction

In medical imaging, noise reduction is essential because it impacts how easily pictures can be interpreted. Wavelet transforms, total variation, DL, and other ML techniques are frequently employed (35). The identification of pulmonary nodules in chest X-rays is aided by the use of techniques like DL, total variation, and wavelet transforms to eliminate noise while maintaining the structural information in the pictures (36). Other ML techniques, such as the non-local means algorithm, utilize the self-similarity property to remove noise and edges in images (37). A combination of the wavelet transforms and the bilateral filter is also used to optimize the image denoising problem, especially in ultrasound images (38). Fast imaging protocols are also critical in improving the patient experience and emergency care. ML optimizes the imaging protocols to allow fast scanning without compromising the image quality. Iterative reconstruction using ML optimizes the quality of images in computed tomography (CT) scans, reducing the radiation exposure to the patient (39). ML also generates pseudo-CT images using magnetic resonance imaging (MRI) scans, which optimizes the radiotherapy planning using high-quality images (40). Overall, ML improves image quality, speeds up image acquisition, and improves diagnostic efficiency. Even though the ML-based reconstruction methods have been shown to have superior signal-to-noise ratio and contrast retention, the majority of the studies have been conducted retrospectively and have been evaluated with controlled phantom data and/or single-institution studies. External validation, especially with regard to cross-scan vendors and cross-protocols, has not been extensively conducted. Additionally, the comparative effectiveness studies to assess the long-term clinical impact, such as diagnostic confidence and patient-related outcomes, are not available. DL-based reconstruction techniques have demonstrated dose reductions of up to 30–40% while maintaining image quality.

Disease detection and diagnosis in radiology

These AI-based systems help radiologists interpret images, identify patterns, and offer diagnostic assistance (41,42). These systems continue to evolve through exposure to new data and improving through new data, making them useful for dealing with emerging diseases (43). AI is best at diagnosing diseases that are highly impactful. In the field of oncology, DL models have achieved AUC scores above 0.90 for detecting breast cancer using mammograms, approaching radiologist-level performance in controlled settings, and assist in the diagnosis of hepatocellular carcinoma (44). In neurology, ML has been applied in neuroimaging for the identification of imaging biomarkers in neurological disorders (45,46). In cardiovascular diseases, AI uses CT angiogram images for detecting coronary stenosis, thus improving the management and outcome of the disease. Although diagnostic AI models show promising accuracy, sensitivity, and specificity, most of the models are based on retrospective datasets, where the prevalence of disease is high, which may be artificially inflating the sensitivity and specificity of the models. Most models are not externally validated, and many models are not tested on diverse populations, making it difficult to generalize the results. Moreover, prospective clinical trial results for AI models are scarce, showing the huge gap between accuracy and practicality. Reported sensitivities range between 85% and 95% and specificities between 80% and 90% in controlled datasets.

Prognosis and outcome prediction in radiology

ML creates imaging biomarkers that are utilized to predict disease progression. Advanced imaging modalities such as CT and MRI have been used in prognostic modelling and treatment response assessment in oncology (47). These imaging biomarkers facilitate personalized treatment planning and therapeutic decision-making. MRI-CT fusion imaging can precisely delineate tumour volume for radiotherapy. CT imaging can calculate myocardial extracellular volume, which can detect cardiac pathology to plan the personalized treatment (48). Whole-body MRI can improve staging, as in Hodgkin lymphoma, which can improve treatment outcomes (49). There are, however, limitations to prognostic modelling, including limited sample sizes. Calibration performance and validation are also limited. There are also issues with radiomic feature stability due to imaging protocol variability.

Workflow optimization in radiology

For instance, AI systems help in the optimal scheduling of cases by prioritizing urgent cases and optimizing appointments based on patient demographics, urgency, and modality availability. This helps in reducing wait times while improving efficiency. This enables the radiologist to focus on the actual work of interpreting images (50). It is also important to note that radiologists are subjected to increased workload and burnout, which is exacerbated by the coronavirus disease 2019 (COVID-19) pandemic. However, AI systems help in reducing burnout by automating processes such as the detection of benign mammograms, reducing the workload and the risk of errors (51). Ergonomics and organizational factors also play a crucial role in enhancing the well-being of the radiologist. Although there is evidence that workflow optimization tools have the potential for reducing reading time and urgent cases, there is insufficient evidence on the longitudinal implementation of these tools (52). Moreover, there is a need for further studies on the effects of AI systems on the rates of diagnostic errors, medicolegal liability, and trust. It is also important to consider the possibility of automation bias, which might result from relying on AI systems for triage.

While collectively, the applications of ML techniques for acquisition, diagnosis, and prognosis, and workflow have significant technical potential, heterogeneity in study design, prospective validation, and reporting standards emphasizes the need for rigorous methodological harmonization.

Challenges in implementing ML in radiology: data-related challenges

Data-related challenges

The requirement for extensive and varied datasets is another obstacle to the use of ML in radiology. Large datasets are necessary for DL techniques; however, the majority of radiology datasets only include hundreds of cases. This is in contrast to the Deep Lesion dataset, which contains 32,000 annotated lesions (53). Small and non-diverse datasets result in increased cases of overfitting, leading to decreased accuracy in the diagnosis process (54). However, acquiring large datasets is a challenge in itself, as it requires collaboration among many institutions while adhering to regulations such as the Health Insurance Portability and Accountability Act (HIPAA) and General Data Protection Regulation (GDPR). Privacy and security are major barriers in the adoption of AI in the field of radiology (55). Data ownership, patient consent, and ethical issues are major concerns. Previous studies have shown that physicians support the use of patient data for AI development but express concerns regarding data privacy and misuse (56). Such issues affect the trust factor in adopting AI in the field of radiology. A strong security system is required in this context (57). Apart from that, there are also practical challenges that need to be addressed when it comes to building large, diverse, and well-annotated imaging datasets. First, radiologists’ annotation varies, leading to noisy labels that may affect the model’s performance. Second, when it comes to deploying the learned model in another environment, it can shift when it was learned in one environment only, especially when it comes to different scanners or patient populations. Federated learning can also be considered; however, challenges associated with it have yet to be addressed.

Model performance and validation in radiology

Overfitting is frequently encountered, especially in radiology, because of the heterogeneity of image-based data. Regularization, cross-validation, and data augmentation are techniques that help improve generalization. However, models may not generalize well when tested on other populations, which again emphasizes the importance of validation on diverse datasets (58). Training dataset bias may lead to lower diagnostic performance for certain populations. It is crucial to ensure that the datasets are representative of diverse populations and scenarios. Stratified sampling and evaluation are techniques that help mitigate bias and improve fairness (59). Although the strategies for regularization and cross-validation have been significantly improved, the performance of many radiology ML models has been observed to decrease with the use of external validation sets. Calibration metrics, which are critical for decision-making, have been less reported compared to discrimination metrics such as AUC. In addition, the reproducibility of the results is still a concern, given the variability in preprocessing, feature extraction, and training methodologies, which could impact the reproducibility of the results. Therefore, prospective validation experiments are necessary to evaluate the robustness of the ML models and deal with the radiomics reproducibility dilemma.

Integration into clinical practice

ML tools should be able to integrate smoothly into current radiology information technology (IT) infrastructure, such as Electronic Health Records (EHRs) and Picture Archiving and Communication Systems (PACS). The problem of interoperability arises from the variety of data formats and communication protocols (60,61). The creation of universal standards, such as the “LOINC/RSNA Radiology Playbook”, is essential for the adoption of ML tools (62). The radiologists’ trust in the ML tools is built on transparency, reliability, and validation (63). Education programs and radiologists’ engagement in AI tool development are important for the adoption of ML tools (64). Without clinician confidence, integration into clinical workflows will be limited. Apart from the technological integration, there is also the need for integration into clinical workflows and the current reimbursement systems. Studies on implementing AI systems have shown that poorly integrated systems can actually increase cognitive workload rather than reducing it. Moreover, there is also the issue of medico-legal liability, especially in cases of diagnostic errors. Economic evaluation on the cost-effectiveness of AI systems is also currently limited. It is important to develop governance systems and liability policies.

Regulatory and ethical issues in ML in radiology

Regulatory oversight is important to ensure the safe use of AI in the field of health care. The Food and Drug Administration (FDA) in the United States (US) has stressed the importance of validation, monitoring, and flexibility in the application of AI medical devices (65). The necessity of clinical validation and risk assessment has been highlighted in the European Union (EU) by European Conformity (CE) approval of medical devices under medical devices regulation (MDR)/in vitro diagnostic regulation (IVDR). Finding the ideal balance between safety and innovation is crucial, and cooperation is essential for this. Reduced human oversight, bias, and transparency are some of the ethical problems that require attention (66,67). It is important to ensure the data used to train the algorithm is diverse to prevent disparities in health outcomes (68,69). Transparency is another important factor; the decisions reached by the algorithm need to be interpretable by the clinician (70,71). There is a need to involve all the stakeholders to ensure equitable access and prevent health inequalities (71). A further layer of complexity for the regulatory oversight is the adaptive and learning nature of the algorithms. The conventional approval mechanisms are not completely geared towards accommodating changes to the models. The surveillance of the performance of the models in the post-market environment is a significant consideration. Ethical considerations also include issues beyond bias, such as the transparency of data sources for the models, the explainability of the decisions of the models, and the informed consent of the patients for the use of AI for diagnostic purposes. These considerations, in aggregate, demonstrate that the technical performance of the models is not sufficient for the clinical application of ML. The sustainability of ML in the field of radiology will necessitate a solution that addresses all the issues comprehensively.

Current trends and future directions in radiology

Emerging techniques

Among the potential developments in radiology are federated learning and self-supervised learning. When data is hard to come by and costly to label, self-supervised learning aids the model in learning from unlabelled data. It has shown encouraging results in enhancing ML models’ performance on tasks like segmentation and picture classification (12). Federated learning helps models learn collaboratively without sharing patient data, thus improving the generalizability of ML models and overcoming the problem of data sharing, which violates HIPAA and GDPR regulations (13).

Explainable AI (XAI) helps to improve the level of transparency for ML models, thereby addressing the concerns of interpretability and trust. By facilitating the provision of explanations for the predictions, XAI helps to ensure the acceptance of the clinicians for the use of ML for decision-making purposes (72). XAI methods such as class activation mapping (CAM) and local interpretable model-agnostic explanations (LIME) have been used to improve interpretability (73). Apart from the aspect of trust, XAI helps to address the concerns of bias, thereby becoming critical for the integration of AI into radiology.

Collaborative efforts in radiology

For ML in radiology to be successful, it is crucial to ensure a collaborative approach among radiologists, data scientists, and engineers. For instance, cancer response model design involves the integration of images, clinical data, and outcomes (74). Collaboration also ensures access to a larger set of annotated data, thus enhancing the reliability of ML tools in different populations. Finally, engineers play a crucial role in ensuring the ML tools are easily integrated into a clinical workflow (75).

Access to open datasets and standardized frameworks is important for reproducibility and interoperability. Public initiatives like the Medical Segmentation Decathlon enable benchmarking and training on different populations. Standards like HL7 and DICOM enable the free exchange of imaging information between different health systems, thus facilitating the adoption of AI in clinical practice. Overall, open data and international standards promote collaboration and innovation in the field of radiology.

Future possibilities

ML can be used to facilitate personalized medicine through the analysis of imaging studies as well as clinical data to predict the course of the disease. For cancer therapy, ML models can be used to identify various features of the tumour to develop personalized therapy. By combining this with omics data such as genomics and proteomics, a detailed patient profile can be established to make precise therapeutic decisions.

The integration of ML with multi-omics information has the potential to enhance the accuracy of precision medicine, where the mechanisms of disease can be understood, along with biomarkers for prediction. For example, the combination of genomics with images has the potential to enhance the knowledge of tumour heterogeneity, thereby guiding therapy. Multi-omics information helps in the stratification of patients, thereby minimizing costs while maximizing the benefits of therapy (76). As the science of radiology advances, the combination of ML, omics, open data, and interdisciplinarity will dictate the future of diagnostic accuracy and therapy.

Real-world implementation and lessons learned

Examples of clinical implementation

Several ML-based applications have moved beyond the proof-of-concept stage to the clinical or near-clinical environment. In breast cancer screening, for example, AI-based mammography systems have been reported to achieve sensitivity comparable to that of radiologists, with the potential for workload reduction for screening programs (50). Similarly, for lung nodule detection, there are reports of high sensitivity achieved by DL-based computer-aided detection systems, as demonstrated in the challenge datasets such as the LUNA16 for computerized detection of nodules.

In emergency radiology settings, AI-assisted detection of intracranial haemorrhage and acute ischemic strokes has been linked to quicker prioritization of cases and improved time-to-treatment outcomes (2,59). Cardiovascular imaging applications like the assessment of coronary stenosis and cardiac risk assessment with AI systems demonstrate the expanding footprint of ML systems (77). These examples demonstrate the potential viability of integrating AI systems with routine radiological practice under controlled validation settings. However, the extent of implementation for these tools is limited to particular institutions or controlled environments. Long-term outcome data, cost-effectiveness, and multicentre prospective validation are still scarce. Thus, though the potential for translational application is clear from the above case studies, generalizability is still to be determined.

Key lessons from implementation efforts

Several lessons can be drawn from the early experience with ML systems in radiology. First, diversity and external validation are important factors in the performance of ML systems in real-world settings. There is evidence that ML systems may perform poorly if they were trained on a homogeneous data set with limited diversity in terms of demographic or imaging conditions (78). Second, interdisciplinary collaboration among radiologists, data scientists, engineers, and institutional leaders is important to guarantee clinical relevance and workflow compatibility. Third, ML systems require constant monitoring and recalibration to guarantee their performance over time, especially as imaging conditions change (79).

Lastly, ethical and regulatory forms of governance need to be integrated with technical innovation. Accountability frameworks, transparency of model behaviour, and education of clinicians are key to creating trust and ensuring safe practice (80). Combined, all these instances demonstrate that for successful implementation of ML in the field of radiology, it is important to consider the success of both the algorithms and the organization in addition to human monitoring and proper validation.

Discussion

Significant enthusiasm has been raised by the rapid growth in the application of ML techniques in radiology, with several studies showing promising diagnostic accuracy in a variety of clinical settings. In particular, CNNs and transformer models often achieve performance comparable to radiologists in controlled settings using performance metrics in retrospective analyses (8,10,79). Nevertheless, an in-depth look at the literature reveals important limitations in the methodology that restricts the clinical application.

However, the dominant trend among studies is the use of retrospective data from a single centre with pre-defined disease prevalence (78). Although this allows model optimization and development, overestimation of performance may occur when applied to real-world settings. External validation studies, especially when conducted on populations with distinct geographical locations and imaging modalities, remain underreported (59,78). In addition, there is a trend towards underreporting studies that report performance on model calibration and estimation of uncertainties, which is critical to model deployment.

Another challenge that arises with the reproducibility of radiomics and DL models is that differences in the techniques used during the preprocessing stage, segmentation process, and feature extraction may produce significantly different outcomes compared to what has been obtained in other studies (34,78). This has been characterized as the ‘reproducibility crisis’ of medical AI, which necessitates the need to adopt standardized methodological approaches and reporting guidelines (15,71).

Besides algorithmic validation, barriers to implementation are also a contributor to the “AI chasm”—the gap between research prototype potential and clinical practice adoption (79). Alignment with the financial model, legal accountability, and compatibility with electronic health records and photo archiving communication systems are all necessary for integration into the therapeutic route (60).

New paradigms like foundation models, self-supervised learning, federated learning, and generative AI provide promise in resolving some of the long-standing issues with data availability and variability. Federated learning models provide frameworks that enable collaborative research among institutions without sharing data, thus addressing privacy and regulatory concerns (12,13). Generative models provide promise in addressing issues with data balance and annotation availability; however, the fidelity of generative models is an open research question. These emerging models may provide better generalization capabilities but also bring new complexities in regulatory control, especially with adaptive models that change post-implementation (65).

LLMs are another emerging technology, especially with regard to report summarization and decision support systems (76). In their early stages, they promise efficiency benefits, but concerns regarding hallucinations, factual inconsistencies, and a lack of grounding in specific domains require thorough evaluation and human oversight. LLMs in radiology should be implemented with a focus on explainability and transparency in their training data and against clinical benchmarks (67,68).

The integration of AI technology with sustainability is still heavily influenced by ethical and legal issues. In addition, bias may occur as a result of the underrepresentation of certain populations within the population under consideration and may worsen health disparities among these populations (69). There is a need to document the model to build trust and to conduct post-market surveillance to ensure accountability (66,75). There is a need to balance innovation with patient safety as adaptive AI becomes more prevalent (65).

The combined results imply that the application of ML in radiology is shifting from experimental proof-of-concept research to early clinical deployment. However, for the technology to be adopted, there is a need for a paradigm shift from performance-based research to validation-based deployment. In order to close the AI gap, prospective multicentre trials, assessment metrics, calibration, and multidisciplinary frameworks will be essential. ML will only deliver its full potential in radiography with the aid of scientific integrity, transparency, and collaboration. In order to ensure that ML algorithms are used safely and effectively, future research needs to focus on prospective multicentre studies, standard evaluations, and practical implementation in clinical settings.

Strengths and limitations of this review

Even though this is a narrative review, the paper provides a comprehensive and up-to-date overview of ML applications in radiology, underpinned by a systematic review process and rigorous inclusion criteria. Focusing on validation quality, reproducibility, and ethical aspects, the paper advances the scope of earlier evaluations of task-based DL applications by including innovative concepts such as foundation models, federated learning, generative intelligence, and LLMs.

Nevertheless, risk-of-bias analysis was not conducted since the study is a narrative review. Additionally, focusing only on studies written in the English language could give rise to selection bias, and some results might not be applicable anymore owing to the fast development of AI technology. Nevertheless, despite its flaws, the paper tries to provide an unbiased overview of recent progress in applying ML in radiology.

Conclusions

The application of ML in the sphere of radiology is gaining importance when it comes to imaging, diagnosis, prediction, and process improvement. Algorithms designed for this purpose have been developed from being a narrow approach to using sophisticated technology, including ML, foundation models, and LLMs. Nevertheless, some issues remain pertinent, including data bias, reproducibility, regulation, and lack of external validation. Addressing these issues is crucial in order to establish the connection between AI theories and practicality through multicenter validation and standardized reporting, among other factors. The future of ML in radiology hinges on striking the right balance.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://tro.amegroups.com/article/view/10.21037/tro-25-47/rc

Peer Review File: Available at https://tro.amegroups.com/article/view/10.21037/tro-25-47/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tro.amegroups.com/article/view/10.21037/tro-25-47/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Abimbola A. A Review of the Historical Development and Future Significance of Artificial Intelligence in Radiology. medtigo J Med 2024;2:e3062253.
2. Zare N, Loor MM, Badakhshan S, et al. Influence of progress in artificial intelligence on radiology’s future: a two-fold view on advantages and challenge. World J Biol Pharm Health Sci 2024;17:215-9.
3. Gampala S, Vankeshwaram V, Gadula SSP. Is Artificial Intelligence the New Friend for Radiologists? A Review Article. Cureus 2020;12:e11137.
4. Najjar R. Redefining Radiology: A Review of Artificial Intelligence Integration in Medical Imaging. Diagnostics (Basel) 2023;13:2760. [Crossref] [PubMed]
5. Pesapane F, Codari M, Sardanelli F. Artificial intelligence in medical imaging: threat or opportunity? Radiologists again at the forefront of innovation in medicine. Eur Radiol Exp 2018;2:35.
6. Hameed BMZ, Prerepa G, Patil V, et al. Engineering and clinical use of artificial intelligence (AI) with machine learning and data science advancements: radiology leading the way for future. Ther Adv Urol 2021;13:17562872211044880. [Crossref] [PubMed]
7. Barragán-Montero A, Javaid U, Valdés G, et al. Artificial intelligence and machine learning for medical imaging: A technology review. Phys Med 2021;83:242-56. [Crossref] [PubMed]
8. Erickson BJ, Korfiatis P, Akkus Z, et al. Machine Learning for Medical Imaging. Radiographics 2017;37:505-15. [Crossref] [PubMed]
9. Suzuki K. Overview of deep learning in medical imaging. Radiol Phys Technol 2017;10:257-73. [Crossref] [PubMed]
10. Rana M, Bhushan M. Machine learning and deep learning approach for medical image analysis: diagnosis to detection. Multimed Tools Appl 2022. [Epub ahead of print]. doi: 10.1007/s11042-022-14305-w.
11. Niu K, Chen S, Fan J, et al. The diagnostic value of a coronary computed tomography angiography scan-based radiomics model for coronary stenosis. In: Piccaluga PP, Baloch Z, editors. Third International Conference on Biomedical and Intelligent Systems (IC-BIS 2024). Proc SPIE 2024;13208:132081G.
12. Lee D. The Application of Foundation Models in Radiology: Bridging Images, Reports, and Beyond. J Korean Soc Radiol 2025;86:919-37. [Crossref] [PubMed]
13. Xu J, Glicksberg BS, Su C, et al. Federated Learning for Healthcare Informatics. J Healthc Inform Res 2021;5:1-19. [Crossref] [PubMed]
14. Mahajan D, Tsou CH, Liang JJ. IBMResearch at MEDIQA 2021: toward improving factual correctness of radiology report abstractive summarization. In: Proceedings of the 20th Workshop on Biomedical Language Processing. 2021:302-10.
15. Castro DC, Walker I, Glocker B. Causality matters in medical imaging. Nat Commun 2020;11:3673. [Crossref] [PubMed]
16. Varoquaux G, Cheplygina V. Machine learning for medical imaging: methodological failures and recommendations for the future. NPJ Digit Med 2022;5:48. [Crossref] [PubMed]
17. Haq IU, Mhamed M, Al-Harbi M, et al. Advancements in Medical Radiology Through Multimodal Machine Learning: A Comprehensive Overview. Bioengineering (Basel) 2025;12:477. [Crossref] [PubMed]
18. Xu Y, Li Y, Wang F, et al. Addressing the current challenges in the clinical application of AI-based Radiomics for cancer imaging. Front Med (Lausanne) 2025;12:1674397. [Crossref] [PubMed]
19. Litjens G, Kooi T, Bejnordi BE, et al. A survey on deep learning in medical image analysis. Med Image Anal 2017;42:60-88. [Crossref] [PubMed]
20. Arasi MA, Babu S. Survey of Machine Learning Techniques in Medical Imaging. Int J Adv Trends Comput Sci Eng 2019;8:2107-16.
21. Militello C, Ranieri A, Rundo L, et al. On Unsupervised Methods for Medical Image Segmentation: Investigating Classic Approaches in Breast Cancer DCE-MRI. Applied Sciences 2021;12:162.
22. Alwiyah A, Setyowati W. A comprehensive survey of machine learning applications in medical image analysis for artificial vision. International Transactions on Artificial Intelligence 2023;2:90-8.
23. Li M, Jiang Y, Zhang Y, et al. Medical image analysis using deep learning algorithms. Front Public Health 2023;11:1273253. [Crossref] [PubMed]
24. Mun SK, Wong KH, Lo SB, et al. Artificial Intelligence for the Future Radiology Diagnostic Service. Front Mol Biosci 2020;7:614258. [Crossref] [PubMed]
25. Habehh H, Gohel S. Machine Learning in Healthcare. Curr Genomics 2021;22:291-300. [Crossref] [PubMed]
26. Alhumaidi NH, Dermawan D, Kamaruzaman HF, et al. The use of machine learning for analyzing real-world data in disease prediction and management: systematic review. JMIR Med Inform 2025;13:e68898. [Crossref] [PubMed]
27. Hagenah J, Leymann S, Ernst F. Integrating label uncertainty in ultrasound image classification using weighted support vector machines. Curr Dir Biomed Eng 2019;5:285-7.
28. Puttagunta M, Ravi S. Medical image analysis based on deep learning approach. Multimed Tools Appl 2021;80:24365-98. [Crossref] [PubMed]
29. Faes L, Wagner SK, Fu DJ, et al. Automated deep learning design for medical image classification by health-care professionals with no coding experience: a feasibility study. Lancet Digit Health 2019;1:e232-42. [Crossref] [PubMed]
30. Tahmasebi AM, Zhu H, Mankovich G, et al. Automatic Normalization of Anatomical Phrases in Radiology Reports Using Unsupervised Learning. J Digit Imaging 2019;32:6-18. [Crossref] [PubMed]
31. Gulshan V, Peng L, Coram M, et al. Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs. JAMA 2016;316:2402-10. [Crossref] [PubMed]
32. Chng SY, Tern PJW, Kan MRX, et al. Automated labelling of radiology reports using natural language processing: Comparison of traditional and newer methods. Health Care Sci 2023;2:120-8. [Crossref] [PubMed]
33. Hassanpour S, Langlotz CP. Information extraction from multi-institutional radiology reports. Artif Intell Med 2016;66:29-39. [Crossref] [PubMed]
34. Parekh VS, Jacobs MA. Deep learning and radiomics in precision medicine. Expert Rev Precis Med Drug Dev 2019;4:59-72. [Crossref] [PubMed]
35. Song Z, Guo D, Tang Z, et al. Noncontrast Computed Tomography-Based Radiomics Analysis in Discriminating Early Hematoma Expansion after Spontaneous Intracerebral Hemorrhage. Korean J Radiol 2021;22:415-24. [Crossref] [PubMed]
36. Kwak HJ, Lee SJ, Lee Y, et al. Quantitative study of total variation (TV) noise reduction algorithm with chest X-ray imaging. J Instrum 2018;13:T01006.
37. Kim K, Choi J, Lee Y. Effectiveness of Non-Local Means Algorithm with an Industrial 3 MeV LINAC High-Energy X-Ray System for Non-Destructive Testing. Sensors (Basel) 2020;20:2634. [Crossref] [PubMed]
38. Raj VNP, Venkateswarlu T. Ultrasound medical image denoising using hybrid bilateral filtering. Int J Comput Appl 2012;56:44-51. [Crossref]
39. Leng S, Yu L, Fletcher JG, et al. Maximizing Iodine Contrast-to-Noise Ratios in Abdominal CT Imaging through Use of Energy Domain Noise Reduction and Virtual Monoenergetic Dual-Energy CT. Radiology 2015;276:562-70. [Crossref] [PubMed]
40. Wang T, Manohar N, Lei Y, et al. MRI-based treatment planning for brain stereotactic radiosurgery: Dosimetric validation of a learning-based pseudo-CT generation method. Med Dosim 2019;44:199-204. [Crossref] [PubMed]
41. Strubchevska O, Kozyk M, Kozyk A, et al. The Role of Artificial Intelligence in Diagnostic Radiology. Cureus 2024;16:e72173. [Crossref] [PubMed]
42. Fowler GE, Blencowe NS, Hardacre C, et al. Artificial intelligence as a diagnostic aid in cross-sectional radiological imaging of surgical pathology in the abdominopelvic cavity: a systematic review. BMJ Open 2023;13:e064739. [Crossref] [PubMed]
43. Nguyen NC, Beriwal S, Moon CH, et al. (18)F-FDG PET/MRI Primary Staging of Cervical Cancer: A Pilot Study with PET/CT Comparison. J Nucl Med Technol 2020;48:331-5. [Crossref] [PubMed]
44. Silva HECD, Santos GNM, Leite AF, et al. The use of artificial intelligence tools in cancer detection compared to the traditional diagnostic imaging methods: An overview of the systematic reviews. PLoS One 2023;18:e0292063. [Crossref] [PubMed]
45. Zhu G, Jiang B, Tong L, et al. Applications of Deep Learning to Neuro-Imaging Techniques. Front Neurol 2019;10:869. [Crossref] [PubMed]
46. Damer A, Chaudry E, Eftekhari D, et al. Neuroimaging Scoring Tools to Differentiate Inflammatory Central Nervous System Small-Vessel Vasculitis: A Need for Artificial Intelligence/Machine Learning?-A Scoping Review. Tomography 2023;9:1811-28. [Crossref] [PubMed]
47. Yamamichi J, Kawaguchi Y, Otsuka T, et al. Assessment of tumor volume and density as a measure of the response of advanced hepatocellular carcinoma to sorafenib: Application of automated measurements on computed tomography scans. JGH Open 2020;4:145-52. [Crossref] [PubMed]
48. Cundari G, Galea N, Mergen V, et al. Myocardial extracellular volume quantification with computed tomography—current status and future outlook. Insights Imaging 2023;14:156. [Crossref] [PubMed]
49. Spijkers S, Littooij AS, Kwee TC, et al. Whole-body MRI versus an FDG-PET/CT-based reference standard for staging of paediatric Hodgkin lymphoma: a prospective multicentre study. Eur Radiol 2021;31:1494-504. [Crossref] [PubMed]
50. Rodriguez-Ruiz A, Lång K, Gubern-Merida A, et al. Can we reduce the workload of mammographic screening by automatic identification of normal exams with artificial intelligence? A feasibility study. Eur Radiol 2019;29:4825-32.
51. Al Rekabi A, Chen M, Patel N, et al. Well-being and Burnout Amongst Interventional Radiologists in the United Kingdom. Cardiovasc Intervent Radiol 2023;46:1053-63. [Crossref] [PubMed]
52. Pankhania M. Ergonomics in radiology: Preventing radiologist burnout. Int J Radiol Diagn Imaging 2020;3:9-15.
53. Yan K, Wang X, Lu L, et al. DeepLesion: automated mining of large-scale lesion annotations and universal lesion detection with deep learning. J Med Imaging 2018;5:036501.
54. Mutasa S, Sun S, Ha R. Understanding artificial intelligence based radiology studies: What is overfitting? Clin Imaging 2020;65:96-9. [Crossref] [PubMed]
55. Sheller MJ, Edwards B, Reina GA, et al. Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data. Sci Rep 2020;10:12598. [Crossref] [PubMed]
56. Castagno S, Khalifa M. Perceptions of Artificial Intelligence Among Healthcare Staff: A Qualitative Survey Study. Front Artif Intell 2020;3:578983. [Crossref] [PubMed]
57. Jayaneththi B, McCaffery F, Regan G. Data security challenges in AI-enabled medical device software. In: 2023 31st Irish Conference on Artificial Intelligence and Cognitive Science (AICS). IEEE; 2023:1-6.
58. Tran AT, Zeevi T, Payabvash S. Strategies to Improve the Robustness and Generalizability of Deep Learning Segmentation and Classification in Neuroimaging. BioMedInformatics 2025;5:20. [Crossref] [PubMed]
59. Kotovich D, Twig G, Itsekson-Hayosh Z, et al. The impact on clinical outcomes after 1 year of implementation of an artificial intelligence solution for the detection of intracranial hemorrhage. Int J Emerg Med 2023;16:50. [Crossref] [PubMed]
60. Wang X, Yang G, Zhang Y, et al. Automated delineation of nasopharynx gross tumor volume for nasopharyngeal carcinoma by plain CT combining contrast-enhanced CT using deep learning. J Radiat Res Appl Sci 2020;13:568-77.
61. Fink A, Tran H, Reisert M, et al. A deep learning approach for projection and body-side classification in musculoskeletal radiographs. Eur Radiol Exp 2024;8:23. [Crossref] [PubMed]
62. Gomes R, Kamrowski C, Mohan PD, et al. Application of Deep Learning to IVC Filter Detection from CT Scans. Diagnostics (Basel) 2022;12:2475. [Crossref] [PubMed]
63. Bergquist M, Rolandsson B, Gryska E, et al. Trust and stakeholder perspectives on the implementation of AI tools in clinical radiology. Eur Radiol 2024;34:338-47. [Crossref] [PubMed]
64. Lee YH. Efficiency Improvement in a Busy Radiology Practice: Determination of Musculoskeletal Magnetic Resonance Imaging Protocol Using Deep-Learning Convolutional Neural Networks. J Digit Imaging 2018;31:604-10. [Crossref] [PubMed]
65. Ahmad R. Reviewing the relationship between machines and radiology: the application of artificial intelligence. Acta Radiol Open 2021;10:2058460121990296. [Crossref] [PubMed]
66. Rademakers FE, Biasin E, Bruining N, et al. CORE-MD clinical risk score for regulatory evaluation of artificial intelligence-based medical device software. NPJ Digit Med 2025;8:90. [Crossref] [PubMed]
67. Maresova P. Impact of Regulatory Changes on Innovations in the Medical Device Industry Comment on "Clinical Decision Support and New Regulatory Frameworks for Medical Devices: Are We Ready for It? - A Viewpoint Paper". Int J Health Policy Manag 2023;12:7262.
68. Green BL, Murphy A, Robinson E. Accelerating health disparities research with artificial intelligence. Front Digit Health 2024;6:1330160. [Crossref] [PubMed]
69. Arora A, Alderman JE, Palmer J, et al. The value of standards for health datasets in artificial intelligence-based applications. Nat Med 2023;29:2929-38. [Crossref] [PubMed]
70. Yu Y, Gomez-Cabello CA, Haider SA, et al. Enhancing Clinician Trust in AI Diagnostics: A Dynamic Framework for Confidence Calibration and Transparency. Diagnostics (Basel) 2025;15:2204. [Crossref] [PubMed]
71. Abgrall G, Holder AL, Chelly Dagdia Z, et al. Should AI models be explainable to clinicians? Crit Care 2024;28:301. [Crossref] [PubMed]
72. Alkhanbouli R, Matar Abdulla Almadhaani H, Alhosani F, et al. The role of explainable artificial intelligence in disease prediction: a systematic literature review and future research directions. BMC Med Inform Decis Mak 2025;25:110. [Crossref] [PubMed]
73. Aasem M, Javed Iqbal M. Toward explainable AI in radiology: Ensemble-CAM for effective thoracic disease localization in chest X-ray images using weak supervised learning. Front Big Data 2024;7:1366415. [Crossref] [PubMed]
74. Giuste F, Shi W, Zhu Y, et al. Explainable Artificial Intelligence Methods in Combating Pandemics: A Systematic Review. IEEE Rev Biomed Eng 2023;16:5-21. [Crossref] [PubMed]
75. Brereton TA, Malik MM, Lifson M, et al. The Role of Artificial Intelligence Model Documentation in Translational Science: Scoping Review. Interact J Med Res 2023;12:e45903. [Crossref] [PubMed]
76. Shih JY, Mohanty V, Katsis Y, et al. Leveraging large language models to enhance domain expert inclusion in data science workflows. In: Extended abstracts of the CHI Conference on Human Factors in Computing Systems. 2024:1-11.
77. Weikert T, Francone M, Abbara S, et al. Machine learning in cardiovascular radiology: ESCR position statement on design requirements, quality assessment, current applications, opportunities, and challenges. Eur Radiol 2021;31:3909-22. [Crossref] [PubMed]
78. Hosny A, Parmar C, Quackenbush J, et al. Artificial intelligence in radiology. Nat Rev Cancer 2018;18:500-10. [Crossref] [PubMed]
79. Choy G, Khalilzadeh O, Michalski M, et al. Current Applications and Future Impact of Machine Learning in Radiology. Radiology 2018;288:318-28. [Crossref] [PubMed]
80. Mennella C, Maniscalco U, De Pietro G, et al. Ethical and regulatory challenges of AI technologies in healthcare: A narrative review. Heliyon 2024;10:e26297. [Crossref] [PubMed]