Artificial intelligence in healthcare: Technical advances, clinical integration, and future perspectives

Abstract

This narrative review examines the evolution of artificial intelligence (AI) in healthcare, with a focus on the transition from early rule-based systems to modern deep learning architectures and their integration into clinical practice. We examine foundational technologies, including convolutional neural networks for image interpretation, vision transformers for modeling long-range dependencies, and generative adversarial networks for image reconstruction and synthesis. The review further discusses the emergence of multimodal foundation models that integrate imaging with textual and genomic data to enhance diagnostic robustness. The application of these technologies is analyzed across three primary domains: Radiology (image enhancement and automated interpretation), cardiology (electrocardiographic and echocardiography analysis), and oncology (tumor classification and treatment planning). Specific attention is given to the national context in Türkiye, highlighting local initiatives such as TEKNOFEST and TÜBİTAK-supported projects that foster domestic AI development. While AI offers significant benefits in terms of diagnostic accuracy and treatment workflow optimization, challenges regarding data privacy, algorithmic bias, and interpretability (“black box” issues) persist. Future progress depends on the development of explainable AI, rigorous prospective validation, and the establishment of ethical regulatory frameworks.

Key Words: Artificial intelligence; Deep learning; Radiology; Cardiology; Oncology; Clinical decision support; Multimodal imaging; Explainability

Core Tip: Artificial intelligence (AI) is undergoing a paradigm shift from task-specific convolutional architectures to multimodal foundation models. While AI has demonstrated diagnostic performance comparable to experts in radiology, cardiology, and oncology-particularly in retrospective imaging cohorts-its seamless clinical integration remains contingent on overcoming barriers related to explainability, data privacy, and the transition from retrospective validation to prospective clinical trials. This review synthesizes current technical milestones and critically evaluates the path toward implementing human-centric AI.

INTRODUCTION

Artificial intelligence (AI), deep learning, has fundamentally changed medical diagnostics by enabling the automated analysis of high-dimensional datasets[1,2]. The trajectory of AI in medicine began with the development of rule-based expert systems in the 1970s[3]. Yet, it is the recent evolution of deep neural networks that has enabled expert-level precision in complex tasks, such as medical image interpretation[2,4]. For instance, convolutional neural networks (CNNs) have achieved robust performance benchmarks in diabetic retinopathy screening and multi-organ cancer detection[4].

Beyond task-specific models, modern large-scale “foundation models” (e.g., GPT-4, SAM), pre-trained on massive datasets, are now demonstrating zero-shot capabilities in medical imaging through the transfer of learned representations[1]. The primary objective of integrating AI into clinical workflows is to augment clinicians' abilities rather than replace them, promising accelerated diagnostics and individualized therapeutic strategies[2,3]. This review explores AI’s technical foundations, clinical applications in radiology, cardiology, and oncology, and discusses the emerging national AI landscape in Türkiye while addressing critical limitations in current evidence.

Technical background

CNNs: CNN remain the cornerstone of medical image analysis. These models automatically learn hierarchical spatial features through successive layers of convolution and pooling[4]. Architectures such as ResNet and DenseNet have facilitated breakthroughs in retinal and dermatologic diagnostics, often matching the performance of specialists in controlled experimental settings[4].

Transformers: Originally developed for natural language processing, vision transformers (ViTs) utilize self-attention mechanisms to process images as sequences of patches[3]. Unlike the local receptive fields of CNNs, ViTs can model long-range dependencies across an entire image. However, ViTs typically require larger datasets and more computational power to outperform traditional CNNs, which may still offer better generalization on the smaller, specialized datasets commonly used in clinical research[3].

Generative models: Generative adversarial networks (GANs) utilize a competitive framework between a generator and a discriminator to synthesize data. In medical imaging, GANs address data scarcity by generating synthetic images and enhancing image quality through denoising. Notably, CycleGANs have been utilized to strengthen low-dose computed tomography (CT) images and generate synthetic magnetic resonance imaging sequences, thereby reducing acquisition times[1,5]. Despite their utility, synthetic data must be carefully validated to ensure that no “hallucinated” artifacts are introduced, which could mislead diagnostics[5].

Multimodal foundation models: The current frontier involves models that fuse imaging with electronic health records, genomics, and pathology reports. Vision-language models learn joint representations from image-report pairs, enabling “explainable” outputs via natural language[1]. In oncology, combining radiology and genomics through multimodal fusion has shown promise in improving prognostic accuracy over single-modality approaches[6].

CLINICAL APPLICATIONS OF AI

AI in radiology

AI applications in radiology span the entire imaging chain, from acquisition to interpretation.

Image reconstruction: Deep learning-based denoising techniques, such as those utilizing CycleGANs, have demonstrated the ability to enhance signal-to-noise ratios in low-dose CT scans, potentially reducing radiation exposure while maintaining diagnostic accuracy[1].

Automated interpretation: Numerous retrospective studies have indicated that CNNs can achieve diagnostic accuracy comparable to senior radiologists in detecting intracranial hemorrhages, lung nodules, and fractures[4,6]. For example, the CheXNeXt algorithm demonstrated superior sensitivity for specific chest pathologies compared to human readers in a retrospective cohort study[6]. However, it is essential to note that these gains often diminish when applied to real-world datasets with high variability in hardware and patient demographics.

AI in cardiology

AI is transforming cardiology through the high-throughput analysis of electrocardiography (ECG) and echocardiography.

ECG analysis: In a significant retrospective study (EchoNext), deep learning models trained on over one million records demonstrated the ability to detect structural heart disease, with performance metrics exceeding those of traditional cardiologists[7]. These tools can identify subtle patterns of arrhythmias and electrolyte imbalances that may be imperceptible to the human eye.

Echocardiography: CNN frameworks, such as AIEchoDx, have achieved an area under the curve (AUC) greater than 0.98 for diagnosing conditions like atrial septal defects and cardiomyopathies[8]. While these results are promising, their utility in point-of-care settings by non-experts requires further prospective validation to ensure safety and inter-operator reliability.

AI in oncology

In oncology, AI facilitates early screening and precise tumor staging. Mammography-based CNNs have achieved over 96% accuracy in identifying suspicious lesions[6], and lung cancer screening tools have significantly improved sensitivity in detecting early-stage nodules[9].

However, a critical gap remains between academic performance and clinical translation. Most oncological AI models are developed on highly curated, retrospective datasets, which may not account for the biological heterogeneity of tumors across different populations. Furthermore, the “black box” nature of deep learning often hinders clinical trust in treatment planning. Current challenges include the need for multi-institutional validation and the standardization of radiomics features across different scanner manufacturers[8,9].

SUMMARY OF AI APPLICATIONS IN HEALTHCARE

Overview of AI architectures and clinical applications across medical domains. While radiology and cardiology rely on established CNN frameworks, oncology and clinical workflows are increasingly integrating advanced, multimodal, and foundational models (Table 1).

Table 1 Summary of artificial intelligence applications in healthcare¹.

Clinical domain	Representative AI architectures	Key clinical applications	Typical evidence level
Radiology	CNNs, GAN-based models	Nodule detection, image denoising, low-dose CT enhancement	Retrospective, in-silico validation
Cardiology	Deep CNNs, RNN-based models	Arrhythmia detection, echocardiographic segmentation, EF estimation	Large-scale retrospective studies
Oncology	Vision transformers, multimodal fusion models	Tumor grading, outcome prediction, treatment response monitoring	Retrospective cohorts, pilot clinical studies
Clinical Workflow	Large language models, multimodal foundation models	Report generation, study triage, clinical decision support	Emerging real-world deployments

¹Overview of artificial intelligence architectures and clinical applications across medical domains. While radiology and cardiology rely on established convolutional neural network frameworks, oncology and clinical workflows are increasingly integrating advanced multimodal and foundation models.

CNN: Convolutional neural network; AI: Artificial intelligence; GAN: Generative adversarial network; RNN: Recurrent neural network; EF: Ejection fraction; CT: Computed tomography.

NATIONAL CONTEXT AND CHALLENGES: AI IN TÜRKİYE

Türkiye has established a rapidly expanding ecosystem for AI in healthcare, driven by national initiatives such as TEKNOFEST and TÜBİTAK-supported research programs. A prominent example is the TEKNOFEST 2024 Health AI Competition, organized in collaboration with the Turkish Ministry of Health and the Turkish Institutes of Health which focused on clinically relevant challenges, including mammography-based breast cancer detection and Breast Imaging Reporting and Data System classification. The competition attracted unprecedented interest, with over 5000 teams applying nationwide and hundreds advancing through multi-stage evaluations using expert-annotated radiological datasets. Such initiatives not only foster innovation among students and early-career researchers but also contribute to the development of locally curated, clinically annotated datasets-an essential prerequisite for training robust AI models.

In parallel, academic medical centers and technology-focused universities in Türkiye (e.g., Hacettepe University, Bilkent University) have increasingly engaged in AI-driven imaging research, often in collaboration with local startups developing prototype systems integrated into Picture Archiving and Communication System environments. These efforts reflect a growing national capacity in computer vision, deep learning, and clinical data science, particularly within radiology-focused applications.

Despite this momentum, several barriers continue to limit large-scale clinical translation in the national context. First, fragmentation of healthcare data across public and private institutions restricts access to large, standardized, multi-center datasets, thereby constraining model generalizability. Second, regulatory pathways for transitioning research prototypes into certified medical devices (e.g., Medical Device Regulation/Conformité Européenne approval) remain lengthy and resource-intensive, slowing real-world deployment. Finally, while Türkiye’s data protection framework, as outlined in the law on the protection of personal data, is essential for patient privacy, the associated ethical and legal approval processes can delay multi-institutional collaborative research. Consequently, although many national AI initiatives demonstrate strong technical promise, most remain at the prototype or pre-clinical validation stage, underscoring the need for coordinated regulatory, infrastructural, and data-sharing strategies to enable sustainable clinical impact.

BENEFITS AND CHALLENGES

AI offers substantial clinical benefits, including improved diagnostic accuracy, accelerated workflows, optimized resource utilization, and progress toward personalized medicine. By rapidly analyzing large and complex datasets, AI systems can reduce diagnostic errors and augment clinician decision-making[2,7]. For example, AI-assisted colonoscopy has been shown to increase polyp detection sensitivity from approximately 80% to nearly 97%[7], while AI-driven ECG screening can identify subclinical cardiac disease in asymptomatic individuals[8]. In addition, AI-based triage systems, such as Aidoc and Viz.ai, enable the prioritization of urgent imaging findings, with the potential to shorten the time-to-treatment in acute care settings.

Despite these advantages, significant challenges and limitations remain. From a methodological perspective, this review represents a narrative synthesis and is therefore subject to potential selection bias in the included literature. Moreover, most reported performance metrics (e.g., AUC, sensitivity, specificity) are derived from retrospective studies conducted on curated or balanced datasets, which may overestimate real-world performance in settings with lower disease prevalence and greater population heterogeneity.

At the model level, the limited transparency of many deep learning systems (“black box” behavior) continues to hinder clinician trust and widespread adoption. Although explainable AI approaches such as attention or saliency maps are increasingly employed, these techniques may themselves be misleading by highlighting non-diagnostic image features. Ethical considerations-including data privacy, informed consent, and algorithmic bias-remain central concerns, especially when models are trained on non-representative datasets that may underperform in underserved populations[2].

Regulatory and operational barriers further constrain clinical deployment. While regulatory frameworks established by agencies such as the Food and Drug Administration and European Medicines Agency are evolving, only a limited number of AI tools (e.g., IDx-DR for diabetic retinopathy screening and Aidoc for stroke triage) have achieved formal clinical approval. Effective integration into healthcare systems requires clinician training, human-in-the-loop validation, and robust interoperability standards (e.g., Health Level Seven, Digital Imaging and Communications in Medicine). Finally, the predominance of retrospective evidence underscores the need for prospective, multi-center clinical trials to demonstrate real-world effectiveness and to avoid exaggerated claims regarding AI performance.

FUTURE DIRECTIONS

Emerging trends are poised to shape the future of AI in healthcare. Explainable AI will continue to grow, as methods that make model reasoning transparent are essential for clinical trust. For instance, integrating saliency maps or counterfactual explanations may help clinicians better interpret AI-generated suggestions. Multimodal and foundation models are expected to become increasingly prominent. Large language and vision models, such as Med-PaLM, BioGPT, and specialized systems (e.g., LLaVA-Med), are being adapted for healthcare applications, enabling reasoning across text, imaging, and genomic data. Ongoing research is exploring the use of multimodal large language models for clinical decision support. Human–AI collaboration (“human-in-the-loop”) will remain central, positioning AI as an assistive rather than a replacement technology. For example, radiologist-AI feedback loops may allow for the continuous refinement of model performance[3]. AI applications in population health (large-scale screening) and portable devices (e.g., point-of-care ultrasound with AI) are likely to expand access to care. In oncology, AI-guided drug discovery using generative models and integration with theranostic approaches represent important future frontiers.

Nevertheless, many of these approaches remain in early or exploratory stages, with limited prospective and real-world clinical validation, underscoring the need for cautious interpretation and rigorous evaluation before widespread deployment. It is therefore crucial that future developments adhere to principles of trustworthy AI, including robust validation, bias mitigation, and compliance with data protection regulations such as General Data Protection Regulation and Health Insurance Portability and Accountability Act.

CONCLUSION

AI has transitioned from an experimental tool to a pivotal component of modern medical imaging, with architectures such as CNNs, ViTs, and GANs driving advances in diagnostic accuracy across radiology, cardiology, and oncology. While national initiatives in Türkiye are fostering local innovation, a global challenge remains the translation of strong retrospective performance into robust prospective clinical reliability. It should be noted that this work represents a narrative review and is therefore subject to potential selection bias in the included literature.

Future research must prioritize explainability, robust cross-institutional validation, and human-in-the-loop design principles to ensure that AI systems function as safe, equitable, and clinically meaningful augmentative tools rather than replacements for healthcare professionals.