A comprehensive literature search was conducted across multiple electronic databases, including PubMed, Embase, Web of Science, Science Direct, and Cochrane Library. The search strategy used a combination of Medical Subject Headings terms and keywords related to AI, deep learning, and CT imaging. The specific search string used was as follows: (Intelligence, Artificial [Title/Abstract]) OR (Computer Reasoning [Title/Abstract]) OR (Reasoning, Computer [Title/Abstract]) OR (AI (Artificial Intelligence [Title/Abstract]) OR (Machine Intelligence [Title/Abstract])) and so on (Multimedia Appendix 1).

The selected databases are renowned for their extensive coverage of medical and scientific literature, ensuring a thorough retrieval of studies pertinent to our research question. The choice of these databases was also influenced by their indexing of a wide array of medical journals and conference proceedings, which are critical for capturing the latest advancements in the field of AI and CT imaging.

The search was conducted without language restrictions and conducted from the inception of the databases until the search date in 2024, to ensure inclusion of the most current and relevant studies.

Studies were eligible for inclusion if they met the following criteria: (1) clinical validation studies comparing AI-based interventions with conventional CT imaging techniques; (2) reported outcomes related to image quality, radiation dose, or diagnostic performance; and (3) provided sufficient data for quantitative analysis. Exclusion criteria included (1) nonclinical studies or (2) those without a control group, (3) studies focusing solely on AI development without clinical validation, (4) review articles, (5) case reports, (6) conference abstracts, and (7) studies with insufficient data for meta-analysis.

Two independent reviewers extracted data from the included studies using a standardized form. The extracted information included study characteristics (author, year, and study design), sample size and patient demographics, AI intervention details, and outcome measures (image quality metrics, radiation dose indicators, and diagnostic performance).

The quality of the included studies was assessed using the Newcastle-Ottawa Scale for nonrandomized studies. This scale evaluates studies based on selection, comparability, and outcome domains, with a maximum score of 9 indicating the highest quality. The risk of bias for each included study was assessed using the Cochrane risk of bias tool, which provides a standardized evaluation of study quality.

Meta-analyses were performed using Review Manager (version 5.4; The Cochrane Collaboration). For continuous outcomes (eg, image quality scores and CT dose index), mean differences with 95% CIs were calculated. For binary outcomes (eg, diagnostic accuracy), odds ratios with 95% CIs were computed.

Random effect models were used to account for potential heterogeneity among studies. Heterogeneity was assessed using the I² statistic, with values of 25%, 50%, and 75% considered as low, moderate, and high heterogeneity, respectively.

Publication bias was assessed visually using funnel plots. Sensitivity analyses were conducted by sequentially removing each study to evaluate its impact on the overall effect size.

Results were considered statistically significant at P<.05. Forest plots were generated to visually represent the meta-analysis results, displaying individual study effects and the pooled effect size.

An ethical statement is not applicable because this study is based exclusively on published literature.

The initial database search yielded 835 records (PubMed: n=38, Embase: n=191, Web of Science: n=3, Science Direct: n=283, and Cochrane Library: n=320). After removing duplicates, 371 unique records remained. Following title and abstract screening, 11 full-text articles were assessed for eligibility. Ultimately, 5 studies met the inclusion criteria and were included in the meta-analysis (Figure 1).

The 5 included studies were all clinical validation studies published between 2022 and 2024. They encompassed a total of 929 participants (464 in the experimental AI group and 465 in the control group). The studies focused on various anatomical regions and clinical applications, including intracranial aneurysm detection [14], knee anterior cruciate ligament injury [15], breast cancer radiotherapy planning [16], head CT angiography [9], and liver CT imaging [17].

All included studies used deep learning–based AI algorithms for image reconstruction or analysis. The sample sizes ranged from 60 to 14,715 participants. The Newcastle-Ottawa Scale scores for the included studies ranged from 7 to 8, indicating good methodological quality (Table 1).

This meta-analysis synthesized evidence from 5 clinical validation studies to evaluate the role of AI in CT image quality control and radiation protection. The findings suggest that AI-based interventions can significantly improve CT image quality and efficiency in image analysis, with a potential trend toward radiation dose reduction.

The significant improvement in image quality observed with AI-based interventions (mean difference 0.70, 95% CI 0.43-0.96; P<.001) is a key finding of this meta-analysis. This improvement was consistent across different anatomical regions and clinical applications, suggesting that AI algorithms have broad applicability in enhancing CT image quality. The enhanced image quality achieved through AI could have substantial clinical implications. Improved image clarity and detail may lead to more accurate diagnoses, potentially reducing the need for repeat scans or additional imaging studies [19]. This, in turn, could contribute to reduced overall radiation exposure for patients and improved workflow efficiency in radiology departments.

While the meta-analysis showed a trend toward CT dose index reduction with AI-based interventions, the result was not statistically significant (mean difference 0.47, 95% CI –0.21 to 1.15; P=.18). The lack of statistical significance does not necessarily imply that AI has no effect on radiation dose reduction. Rather, it highlights the need for further research to elucidate the factors influencing AI’s impact on radiation dose across different CT applications. It is possible that some AI algorithms are more effective at dose reduction than others, or that dose reduction capabilities may be more pronounced in certain anatomical regions or clinical scenarios [20].

Moreover, the trend toward dose reduction, even if not statistically significant, is encouraging. When combined with the significant improvements in image quality, these findings suggest that AI may enable a favorable shift in the balance between image quality and radiation dose in CT imaging [21].

The significant improvement in the efficiency of image analysis with AI-based methods (odds ratio 1.57, 95% CI 1.08-2.27; P=.02) is another important finding of this meta-analysis. This improvement in efficiency could have far-reaching implications for radiology workflows and patient care. Faster image analysis times could lead to reduced reporting turnaround times, potentially enabling quicker clinical decision-making and treatment initiation [22]. Additionally, improved efficiency could help address the growing demand for CT imaging services and the increasing workload on radiologists [23]. The lack of heterogeneity (I²=0%) in this analysis suggests that the efficiency gains with AI are consistent across different applications. This consistency is particularly noteworthy given the diversity of clinical scenarios represented in the included studies.

While not all studies reported comprehensive diagnostic performance metrics, those that did showed promising results. The high accuracy, sensitivity, and specificity reported by Hu et al [14] for intracranial aneurysm detection, as well as the noninferiority of low-dose CT with AI reconstruction for liver lesion detection reported by Lee et al [17], suggest that AI can maintain or even enhance diagnostic performance while potentially reducing radiation dose. These findings align with previous studies that have demonstrated the potential of AI to improve diagnostic accuracy in various imaging modalities [24]. The combination of improved image quality, potential dose reduction, and maintained or enhanced diagnostic performance positions AI as a powerful tool for advancing CT imaging practices.

The principle of action by which AI, particularly deep learning algorithms, enhances CT image quality is multifaceted. First, AI-driven approaches use advanced computational techniques to reconstruct images from raw data, optimizing the balance between image sharpness and noise reduction. This is achieved through the use of convolutional neural networks that learn to identify and enhance features of interest while suppressing background noise, leading to clearer and more detailed images [24]. Second, AI algorithms can adaptively adjust image reconstruction parameters based on the specific characteristics of the scanned tissue, resulting in improved contrast resolution and better visualization of anatomical structures. Third, AI-based techniques offer the capability for automated image analysis, which can standardize the interpretation process and reduce interobserver variability, thereby contributing to more accurate and consistent diagnostic outcomes [22]. These mechanisms collectively contribute to the significant improvement in image quality observed in our meta-analysis, underscoring the potential of AI to revolutionize CT imaging practices by enhancing diagnostic accuracy and patient safety.

Several limitations of this meta-analysis should be noted. First, the small number of included studies limits the generalizability of the findings and precludes more detailed subgroup analyses. Second, the heterogeneity in AI algorithms, CT protocols, and outcome measures across studies made direct comparisons challenging. Third, the lack of long-term, follow-up data in the included studies means that the potential long-term impacts of AI-enhanced CT imaging on patient outcomes remain unclear. Future research should focus on larger, multicenter studies with standardized protocols and outcome measures to more definitively establish the impact of AI on CT image quality and radiation dose [25]. Long-term studies are needed to assess the clinical impact of AI-enhanced CT imaging on patient outcomes and radiation-induced cancer risk. Additionally, research into the cost-effectiveness of implementing AI in CT imaging workflows would be valuable for health care systems considering the adoption of these technologies.

This meta-analysis provides evidence that AI-based interventions can significantly improve CT image quality and efficiency in image analysis, with the potential for radiation dose reduction. These findings suggest a promising role for AI in enhancing CT imaging practices, potentially leading to improved diagnostic accuracy, reduced radiation exposure, and enhanced patient care. As AI technologies continue to evolve, their integration into CT imaging workflows may become increasingly important in addressing the ongoing challenges of balancing image quality, radiation safety, and diagnostic accuracy.