Artificial Intelligence in Computed Tomography Radiology: A Systematic Review on Risk Reduction Potential

Abstract

Artificial intelligence (AI) has emerged as a transformative technology in radiology, offering enhanced diagnostic accuracy, improved workflow efficiency and potential risk mitigation. However, its effectiveness in reducing clinical and occupational risks in radiology departments remains underexplored. This systematic review aimed to evaluate the current literature on AI applications in computed tomography (CT) radiology and their contributions to risk reduction. Following the PRISMA 2020 guidelines, a systematic search was conducted in PubMed, Scopus and Web of Science for studies published between 2021 and 2025 (the databases were last accessed on 15 April 2025). Thirty-four studies were included based on their relevance to AI in radiology and reported outcomes. Extracted data included study type, geographic region, AI application and type, role in clinical workflow, use cases, sensitivity and specificity. The majority of studies addressed triage (61.8%) and computer-aided detection (32.4%). AI was most frequently applied in chest imaging (47.1%) and brain haemorrhage detection (29.4%). The mean reported sensitivity was 89.0% and specificity was 93.3%. AI tools demonstrated advantages in image interpretation, automated patient positioning, prioritisation and measurement standardisation. Reported benefits included reduced cognitive workload, improved triage efficiency, decreased manual annotation and shorter exposure times. AI systems in CT radiology show strong potential to enhance diagnostic consistency and reduce occupational risks. The evidence supports the integration of AI-based tools to assist diagnosis, lower human workload and improve overall safety in radiology departments.

1. Introduction

Radiology is at the forefront of healthcare innovation, incorporating technological advances to improve diagnostic accuracy. However, radiology departments continue to face significant occupational risks, including high diagnostic workloads, prolonged screen exposure, manual annotation tasks and potential radiation exposure during interventional procedures. These factors contribute to cognitive fatigue, repetitive strain injuries and increased psychological stress among radiologists and technicians. Artificial intelligence (AI) has recently emerged as a transformative tool in radiology, with the potential to support clinical decision making, streamline workflows and reduce both institutional and individual risk exposure [1]. The growing volume and complexity of imaging data, combined with workforce shortages and increasing diagnostic demands, highlight the need for intelligent automation to reduce human error and improve safety [2]. Machine learning and deep learning approaches, in particular, have shown high accuracy in pattern recognition within radiological data, supporting disease detection, enhancing image quality and assisting in image interpretation [1].

However, AI integration extends beyond diagnostic tasks. It is increasingly applied to workflow orchestration, scheduling optimisation, protocol standardisation and automation of administrative processes. These applications contribute to reducing occupational risks such as physical strain, cognitive overload and decision fatigue among radiology professionals [3,4,5]. Occupational safety in radiology has gained attention due to the cumulative effects of high workloads, extended reporting hours and ergonomic hazards related to prolonged workstation use [6,7]. AI systems can automate repetitive tasks, reduce manual reporting demands and support structured communication, decreasing the likelihood of stress-related errors and enhancing work–life balance among radiology staff [8]. Furthermore, AI tools have shown promise in standardising safety protocols, identifying missing or anomalous patient data and monitoring compliance with radiation exposure thresholds [2,4].

Exploring AI’s risk reduction potential lies in its capacity to address the root causes of occupational hazards in CT radiology. Automated systems can minimise repetitive manual processes, such as lesion measurement, image annotation and patient positioning, thereby reducing ergonomic strain. Dose optimisation algorithms help limit radiation exposure for both patients and staff, while workflow orchestration tools shorten reporting times and reduce prolonged static postures and visual fatigue. Decision support systems can alleviate cognitive overload by flagging urgent findings and supporting prioritisation, lowering the mental stress associated with high-stakes decision making. Together, these applications target ergonomic, cognitive and radiation-related risk factors, contributing to safer working conditions in radiology departments.

Despite the extensive literature on AI in medical imaging, there is a notable lack of systematic evidence regarding its contribution to occupational safety and risk mitigation in radiology departments. While previous reviews have focused on diagnostic performance and economic impact, few have addressed the occupational implications of AI integration—particularly its potential to prevent burnout, reduce ergonomic strain and improve procedural consistency across teams [7,9]. This systematic review aims to evaluate the current state-of-the-art AI applications in radiology departments, with a specific focus on their role in reducing occupational risks. Using the PRISMA 2020 methodology, we synthesised available evidence to classify AI tool types, their clinical applications, risk-reducing functions and reported outcomes. The findings are intended to inform future implementation strategies and promote broader adoption of AI solutions to enhance occupational safety in radiology.

2. Materials and Methods

This systematic review followed a predefined protocol that was developed prior to study selection and data extraction. To ensure methodological transparency and facilitate reproducibility, the protocol has been registered and is publicly available since 11 August 2025 in the Open Science Framework (OSF) under the following DOI: https://doi.org/10.17605/OSF.IO/5RTUE. Full methodological details are transparently reported in accordance with the PRISMA 2020 guidelines. This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines [10,11].

2.1. Eligibility Criteria

The review focused on studies evaluating the use of AI in computed tomography (CT) within radiology departments, with particular attention paid to their clinical relevance and reported outcomes. Eligible studies met the following inclusion criteria: (i) original peer-reviewed articles; (ii) published in English between 2021 and 2025; (iii) involving adult human subjects undergoing CT imaging; and (iv) reporting an outcome accuracy > 70%. The exclusion criteria comprised unpublished articles, systematic reviews, book chapters and legislative or guideline documents. Studies were included if they described the use of AI in CT within radiology departments, regardless of the type of AI intervention. Only studies reporting outcomes related to the practical application of AI in daily radiological activities—such as workflow optimisation, diagnostic support or risk mitigation—were considered.

2.2. Information Sources and Search Strategy

A comprehensive electronic database search was conducted up to 15 April 2025 to identify relevant articles published between 2021 and 2025. Searches were performed in PubMed, Scopus and Web of Science using three predefined queries focused on AI applications in CT imaging. A preliminary Google search and author expertise guided the selection of candidate keywords by reviewing relevant article keywords. The keywords—“Artificial Intelligence”, “Radiology”, “Computed Tomography”, “Automated” and “Workflow”—were individually tested in each database to estimate relevant results. Subsequently, combinations of these keywords were evaluated for eligibility, with the final search strategy applying free-text terms to titles and abstracts as follows: (i) “artificial intelligence” AND (Radiology OR “emergency and trauma radiology”) AND (workflow OR “dose reduction” OR “automated”); (ii) “artificial intelligence” AND (Radiology OR “emergency and trauma radiology”) AND (deep learning OR machine learning OR neural networks); and (iii) Radiology AND (“automated process” OR “automated workflow”).

2.3. Study Selection

All retrieved articles were imported into Mendeley Reference Manager (Elsevier Ltd., London, UK) [12], where duplicate records were identified and removed to ensure data integrity. Two independent reviewers (S.C. and M.F.) conducted a blinded screening of titles and abstracts for 230 records using Rayyan software (Rayyan Systems Inc., Cambridge, MA, USA) [13] to assess initial relevance according to the predefined inclusion criteria. Full-text versions were obtained for all potentially eligible studies and final inclusion decisions were reached by consensus between the reviewers. Any disagreements were resolved through discussion or, when necessary, consultation with a third reviewer to ensure objectivity and minimise selection bias. The reviewers subsequently screened 40 full-text articles for eligibility using Rayyan, resulting in the inclusion of 34 studies in the final qualitative synthesis.

2.4. Data Management and Collection

The data charting process was conducted using a structured Excel data extraction form via SciSpace software (PubGenius Inc., Milpitas, CA, USA) [14]. This process involved the systematic extraction of predefined data items initially performed with the support of SciSpace and subsequently verified by two researchers (S.C. and M.F.) to ensure accuracy and consistency. Outcomes were broadly categorised into the following types: AI type, application type, role in the clinical workflow, use case category, AI-related outcomes, sensitivity and specificity. The researchers grouped outcomes by application type, use case category and clinical workflow role to enable comprehensive analysis. A formal critical appraisal of the included studies was not performed since this review aimed to identify AI applicability in CT radiology departments rather than assess methodological quality. For each selected study, data were extracted regarding the report (author, year, title and country), study characteristics (research focus and objectives), methodology (study type, population, target sample, control sample and measured variables) and research outcomes (results, main findings, limitations, identified biases and suggestions for future research).

3. Results

3.1. Selection of Sources of Evidence

The initial database search retrieved 17,284 articles. Subsequently, filters were applied, excluding 3443 articles by publication date, 8647 by document type, 416 by source type and 844 due to language criteria. In addition, 851 articles were excluded for being off topic and 2838 were removed for not meeting inclusion criteria (such as studies involving non-human or pediatric populations). This process resulted in 245 articles, from which duplicates were removed using Rayyan software, excluding 15 records. During title and abstract screening, 187 of the remaining 230 articles were excluded for irrelevance. Of the 43 full-text articles retrieved, 3 could not be obtained [15,16,17]. Six full-text articles were excluded because their subjects were unrelated to CT or involved out-of-scope populations. Ultimately, 34 studies were included in the qualitative synthesis, as seen in Figure 1.

3.2. Study Descriptors

Figure 2 presents a graphical representation of the main findings extracted from the included studies, including publication year, country of origin, study type and anatomical focus, enabling visual comparison of the main trends identified in the included studies. The characteristics of the included studies provide valuable insights into methodological approaches, geographic distribution and temporal trends. Most of the studies employed a retrospective design (73.5%, n = 25), while others used mixed methods. Only one study was prospective [18], one study was experimental [19] and three studies were observational [20,21,22]. Geographically, the studies originated from diverse countries, with Germany contributing the largest number (n = 9), followed by the United States (n = 7) and Switzerland and China with three studies each. A notable increase in scientific output was observed over recent years, with 2022 accounting for the highest number of publications (n = 12, 35.3%), followed by 2021 and 2024 (n = 8 each, 23.5%), 2023 (n = 5, 14.7%) and one study published in 2025 at the time of this review. Among the studies, 82.4% (n = 28) explicitly reported benefits related to occupational risk reduction (such as decreased image review time, automated prioritisation and elimination of manual annotation).

3.3. Results of Individual Studies

A summary of the results of individual studies is presented in

Table 1. Key findings of the included studies listed by role in clinical workflow.

Study	Role in Clinical Workflow	Type of Application	Type of Use Cases	Anatomic Segment	Pathology	Sensitivity	Specificity	Main Findings
Postiglione T.J. et al. [23]	Addition	Segmentation	Separating normal versus not normal	Chest (aorta segments)	Dissections	85.5%	_____	Reliability of a fully automated artificial intelligence-driven solution, augmented radiology for vascular aneurysm, capable of quick aortic segmentation and analysis of diameter and volume for each segment, achieving a median absolute diameter difference of 1.6 mm compared to expert measurements.
Seyam M. et al. [24]	Addition	Detection	Computer-aided detection Workflow optimization	Head	Intracranial haemorrhage	87.2%	93.9%	AI-based tool for detecting intracranial haemorrhage on non-contrast CT images achieved a diagnostic accuracy of 93.0%, Additionally, the implementation of this tool positively impacted clinical workflow.
Brejnebol M.W. et al. [25]	Addition	Detection	Computer-aided detection	Abdominal and pelvis	Pneumoperitoneum	95.0%	99.0%	AI algorithm demonstrated high specificity and a sensitivity for pneumoperitoneum detection, achieving an area under the curve of 0.77 in the entire patient cohort. Additionally, when excluding cases with smaller amounts of free air, the area under the curve increased to 0.96.
Gang Y. et al. [20]	Replacement	Comparison	Workflow optimization	Chest	Ground glass opacification, consolidation opacification and interstitial thickening	-	-	AI-based automatic positioning method was successful for all patients, reducing total positioning time by 28% and achieving a higher proportion of positioning accuracy (99% vs. 92%) compared to the manual positioning method. Resulted in a 16% reduction in radiation dose and a 9% reduction in image noise in the erector spinae area.
Suri J.S. et al. [21]	Replacement	Segmentation	Computer-aided detection	Chest	Ground glass opacities	96.0%	_____	Development of the COVID Lung Image Analysis System (COVLIAS 1.0), utilises hybrid deep learning models for effective lung segmentation in CT scans of COVID-19 patients. Demonstrated high performance with an area under the curve of approximately 0.96 to 0.98 across different models.
Tekin H.O. et al. [26]	Replacement	Classification	Quality assurance Optimise radiation dose management	Abdominal	NA	82.2%	NA	Successful utilisation of AI approaches for the prediction of Dose Length Product values in abdominal CT scans, achieving high accuracy for risk assessment. Highlights the importance of various parameters affecting its values.
Fink M.A. et al. [27]	Replacement	Classification	Computer-aided detection	Chest (pulmonary arteries)	Pulmonary embolism	NA	NA	Developed a jointly optimised deep learning framework to generate synthetic monoenergetic images from dual-energy CT pulmonary angiography data, improving automatic pulmonary embolism detection in single-energy scans. The framework achieved high-quality visual SMI predictions with a structural similarity index of 0.984 ± 0.002 and a peak signal-to-noise ratio of 41.706 ± 0.547 dB.
Artzner C. et al. [28]	Replacement	Segmentation	Computer-aided detection Workflow optimisation	Chest (thoracic aorta)	Post-TEVAR status, intramural haematoma, mediastinal lymphadenopathy, pericardial effusion and dissections	-	-	The prototypical AI-based algorithm accurately measured thoracic aortic diameters at predefined anatomical locations, demonstrating substantial potential for rapid clinical evaluation of aortic pathology, independent of contrast utilisation or pathology.
Nadeem S.A. et al. [29]	Replacement	Classification Segmentation Detection	Computer-aided detection Workflow optimisation	Vertebra	Vertebral fractures	94.8%	98.5%	New automated methods for segmentation and labelling of individual vertebrae in chest CT images, detecting vertebral deformity fractures using computed vertebral height features and parametric computational modelling. Achieved high accuracy and sensitivity in vertebral fracture assessment.
Meng X.H. et al. [30]	Replacement	Classification Detection	Computer-aided detection	Chest	Rib fractures	92.2%	-	Proposed a fully automated detection method for rib fractures which effectively filtered false positives and improved detection performance by training three networks sequentially. The deep learning model achieved a recall rate of 0.922 and a classification accuracy of 0.863.
Schmuelling L. et al. [31]	Replacement	Detection	Computer aided detection Workflow optimisation	Chest	Pulmonary embolism	79.6%	95.0%	Deep learning-assisted detection of pulmonary embolism in CT pulmonary angiograms and the use of an electronic notification system for communication of results to referring physicians technically work, did not lead to significant improvements in clinical performance measures such as report reading times and patient turnaround times.
Freedman D. et al. [32]	Replacement	Post-processing	Workflow optimisation	Abdominal and pelvis	NA	-	-	Examinations utilising AI-generated multiplanar reformatted images were completed approximately 50% faster than those generated at the console, with no statistical difference in diagnostic confidence or image quality between the two methods.
Hu Q. et al. [18]	Triage	Detection	Computer-aided detection Workflow optimisation	Chest	Pulmonary nodules	78.0%(SDCT) 70.1%(LDCT)	-	Computer-aided detection system demonstrated a sensitivity of 78.03% for detecting pulmonary nodules in standard-dose CT images and 70.15% in low-dose CT images, indicating its effectiveness in identifying nodules under varying conditions.
Canayaz M. et al. [19]	Triage	Classification	Computer-aided detection	Chest	COVID-19	99.4%	-	Development of a new method using Bayesian optimisation-based MobilNetv2 and ResNet-50 models, along with machine learning algorithms, achieving hight accuracy in diagnosing COVID-19.
Villringer K. et al. [22]	Triage	Classification	Computer-aided detection	Head	Intracranial haemorrhage	90.0%	96.0%	AI algorithm successfully detected intracranial haemorrhage in 15% of the analysed cranial CT examinations, with a total of 947 out of 6284 cases identified as having it. The algorithm demonstrated high accuracy when compared to experts.
Wilder-Smith, A.J. et al. [33]	Triage	Detection Segmentation Classification	Grading and classification Workflow optimisation	Chest (heart chambers)	Pericardial effusions	97.0%	100%	Automatic tool for the detection, segmentation and classification of pericardial effusions on CT, achieving high sensitivity and specificity for diagnosing haemopericardium.
Wang D. et al. [34]	Triage	Detection	Computer-aided detection Workflow optimisation	Head	Intracranial haemorrhage	92.0%	96.0%	VeriScout™ detected haemorrhage with high sensitivity and specificity, effectively flagging cases for expedited clinical review within 10 min.
Palm V. et al. [35]	Triage	Detection Segmentation	Computer-aided detection Workflow optimisation	Chest	Pulmonary nodules	-	-	Holistic imaging diagnostics tool that utilises artificial intelligence to automate the detection, quantification and characterisation of common pulmonary, metabolic, cardiovascular and musculoskeletal comorbidities in chest computed tomography scans. Provides a comprehensive evaluation of patients and improve preventive care.
Zsarnoczay E. et al. [36]	Triage	Detection	Computer-aided detection Workflow optimisation	Chest (Pulmonary artery system)	Pulmonary embolism	84.6%	95.1%	Demonstrated that a deep neural network-based algorithm effectively detected pulmonary embolism on CT pulmonary angiogram scans, with the majority of false negatives attributed to small chronic pulmonary embolisms in subsegmental arteries.
Cotena M. et al. [37]	Triage	Detection Classification	Computer-aided detection Workflow optimisation	Chest (thoracic aorta)	Acute aortic dissection	94.3%	100%	Integrating a deep learning-based application for the automated detection and prioritisation of acute aortic dissection on chest CT angiographies significantly reduced the scan-to-assessment time from 15.84 min to 5.07 min, representing a 68% reduction, and the interpretation time from 21.22 s to 14.17 s.
Zaazoue K.A. et al. [38]	Triage	Detection	Computer-aided detection	Chest	Pulmonary embolism	93.2%	99.6%	AI algorithm demonstrated high sensitivity and specificity for detecting pulmonary embolism in contrast-enhanced CT scans of patients with COVID-19, with optimal accuracy achieved at a pulmonary artery attenuation of more than 362 Hounsfield units.
Kim P.E. et al. [39]	Triage	Classification	Clinical outcome prediction	Head	Large vessel occlusion	80.1%	88.6%	Machine learning-based algorithm that utilises handcrafted features from non-contrast computed tomography to predict large vessel occlusion in patients with ischemic stroke, demonstrating reliable predictions and the potential to expedite stroke workflow.
Colasurdo M. et al. [40]	Triage	Detection Segmentation	Computer-aided detection Workflow optimisation	Head	Subdural haematoma	91.4%	96,40%	Potential of AI and deep learning to enhance its detection and analysis, facilitating earlier and more accurate diagnoses.
Dovrat A.Y. et al. [41]	Triage	Detection	Computer-aided detection Workflow optimisation	Head and neck	Large vessel occlusion	81%	94%	Viz LVO system demonstrated high sensitivity and accuracy in detecting large vessel occlusions in a comprehensive stroke centre, with 61 out of 75 identified by the system.
Kim T.M. et al. [42]	Triage	Segmentation Classification	Computer-aided detection	Abdominal (adrenal glands)	Adrenal hyperplasia	75.0–81.3%	97.3–100%	Fully automated deep learning model for adrenal segmentation, classification performance for adrenal hyperplasia.
Sato J. et al. [43]	Triage	Detection Segmentation	Computer-aided detection Workflow optimisation	Abdominal	Detecting organ abnormalities in abdominal images	-	-	Deep learning-based pipeline for detecting abnormalities in abdominal CT images using information from free-text radiology reports, which allows for accurate anomaly detection without the need for manual annotations. The model achieved an overall area under the curve of 0.886 in external validation.
Cheikh A.B. et al. [44]	Triage	Detection	Computer-aided detection Workflow optimisation	Chest (pulmonary arteries)	Pulmonary embolism	92.60	95.8%	The AI algorithm for detecting pulmonary embolism on CT pulmonary angiogram demonstrated high sensitivity, serving as a safety net in emergency radiology practice and enhancing the confidence of radiologists in their diagnoses.
Kundisch A. et al. [45]	Triage	Detection	Computer-aided detection	Head	Intracranial haemorrhages	87.6%	98.1%	AI algorithm detected an additional 29 instances of intracranial haemorrhages, resulting in a 12.2% increase in the number of detected cases compared to initial radiology reports. AI missed 12.4% of cases, while radiologists missed 10.9%.
Grenier P.A. et al. [46]	Triage	Detection Segmentation	Computer-aided detection Workflow optimisation	Chest	Pulmonary embolism	91.4%	91.5%	The deep learning-based application, CINA-PE, demonstrated a high degree of diagnostic accuracy. The algorithm correctly identified 170 of 186 exams positive for pulmonary embolism and 184 of 201 exams negative for pulmonary embolism.
Rava R.A. et al. [47]	Triage Workflow	Detection	Computer-aided detection	Head	Intracranial haemorrhages	93.0%	93.0%	Canon’s AUTO Stroke Solution algorithm accurately identified patients with intracranial haemorrhages and those without. Has the potential to significantly improve treatment times for intracranial haemorrhage patients.
Brendel J.M. et al. [48]	Triage Workflow	Detection	Computer-aided detection	Chest	Coronary artery disease	97.2%	81.7%	Automated deep learning demonstrated remarkable performance in detecting significant coronary artery disease on non-ultra-high-resolution photon-counting coronary CT angiography images.
Ruitenbeek H.C. et al. [49]	Triage workflow	Detection	Computer-aided detection Workflow optimisation	Vertebra	Cervical spine fractures	89.8%	95.3%	Demonstrated high diagnostic accuracy and a sensitivity for detecting cervical spine fractures on CT scans. Time gain of 16 min to diagnosis for fractured cases after its introduction.
Kiefer J. et al. [50]	Triage Workflow	Detection	Computer-aided detection Workflow optimisation	Head	Intracranial haemorrhages	98.1%	89.7%	Performance of a scanner-integrated artificial intelligence algorithm for detecting intracranial haemorrhages in a routine clinical setting, achieving high sensitivity and specificity. The algorithm successfully detected brain haemorrhages in 432 out of 435 cases, demonstrating its feasibility and robustness in emergency settings.
Seker F. et al. [51]	Workflow	Detection	Computer-aided detection Workflow optimisation	Head	Large vessel occlusions	84.0%	96.0%	e-CTA (Brainomix) demonstrates high diagnostic accuracy for the automatic detection of large vessel occlusions in anterior circulation stroke, with sensitivity and specificity values indicating effective performance in identifying proximal occlusions.

, with the characteristics of role in clinical workflow, type of applications, type of use case, anatomic segment, pathology, diagnostic performance metrics and main findings of the included studies, offering a consolidated view of the current evidence base. This structured synthesis allows for a rapid comparison across the studies and facilitates the identification of recurring patterns in AI deployment within CT radiology. A diverse range of AI methodologies reflects the interdisciplinary and rapidly evolving nature of AI in radiology. By integrating both technical performance metrics and contextual information about workflow roles, it provides a foundation for the subsequent thematic analysis presented in this section.

3.3.1. Anatomic Segment

The included studies were categorised by anatomical region, providing a thematic overview of AI application patterns across clinical contexts, as seen in

Table 2. Distribution of included studies by anatomical region, with corresponding number and main clinical applications.

Anatomical Region	Studies (n)	Main Application
Chest	17	Nodule detection, pulmonary lesion classification, pneumonia triage, pulmonary embolism detection, post-operative aortic evaluation, diameter segmentation
Vertebral bodies	2	Structural analysis and measurement
Brain	10	Haemorrhage detection, tumour classification, stroke evaluation, fracture detection, anatomical mapping
Abdomen and pelvis	5	Organ segmentation, pathology detection

. The chest was the most frequently investigated anatomical region, accounting for 50.0% of the studies (n = 17). These studies typically employed deep convolutional neural networks and other AI-based classification and detection algorithms to support tasks (such as nodule detection, pulmonary lesion classification and pneumonia triage). Approximately 8.8% of the studies (n = 3) focused on the pulmonary artery system, emphasising detection of conditions such as pulmonary embolism. A similar proportion addressed aortic imaging, including post-operative aortic evaluation and diameter segmentation. The studies involving heart chambers and vertebral bodies predominantly applied deep learning techniques and automated segmentation methods.

The studies targeting brain imaging represented 29.4% of the sample (n = 10). These investigations primarily employed machine learning algorithms and deep learning architectures, with a strong focus on haemorrhage detection, tumour classification and stroke evaluation. AI applications in this domain were used for the automated identification of intracranial haemorrhage, fracture detection and detailed anatomical mapping, thereby supporting timely clinical decision making and improving diagnostic accuracy. By contrast, the studies grouped under abdominal applications, including both the abdomen and pelvis, accounted for 14.7% of the included studies (n = 5). These studies utilised various AI models, such as artificial neural networks, deep recurrent networks and automated image-based diagnostic tools, primarily for organ segmentation and pathology detection tasks. These applications aim to enhance precision in identifying abnormalities and streamline radiological workflows in complex anatomical regions.

3.3.2. Role in Clinical Workflow

AI applications demonstrated considerable diversity in their clinical roles.

Table 3. Roles of AI applications in the clinical workflow of CT radiology, including number and description of their main functions.

Role in Clinical Workflow	Studies (n)	Description
Triage	21	AI pre-analyses imaging data and flags urgent cases for prioritised review, expediting reporting for time-sensitive conditions
Replacement	9	AI autonomously performs task of image segmentation, anatomical measurements and dose estimation.
Addition	4	AI supports clinicians with quantitative suggestions and pattern recognition, complementing human judgment.
Workflow	1	AI autodetection by algorithm training, improving response times

summarises this evidence. The predominant role was triage, reported in 61.8% of the studies (n = 21), where AI was deployed to pre-analyse imaging data and flag urgent cases for prioritised review [18,19,22,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50]. This approach expedited reporting and helped to reduce diagnostic delays in time-sensitive conditions such as stroke or haemorrhage detection [44]. In 26.5% of the studies (n = 9), AI replaced specific radiological tasks, particularly in image segmentation, anatomical measurements and dose estimation workflows [20,21,26,27,28,29,30,31,32]. These systems operated autonomously in discrete workflow components while maintaining interpretative oversight by radiologists [20]. A smaller proportion (11.8%) implemented AI alongside radiological interpretation, supporting clinicians with quantitative suggestions and pattern recognition without replacing human judgment [23,24,25].

3.3.3. Type of Use Cases

Regarding the types of use cases addressed, most applications fell into the computer-aided detection category, either independently or combined with workflow optimisation, as seen in

Table 4. Types of AI use cases in CT radiology, with number and description of their main functions.

Type of Use Case	Studies (n)	Description
Computer-aided detection	11	Systems trained to detect specific conditions
Computer-aided detection and workflow optimisation	17	Support image interpretation and logistical operations
Other use cases	6	Radiation dose management, outcome prediction and structured reporting

. Computer-aided detection was featured in 32.4% of the studies (n = 11). These systems were often trained to detect specific conditions, such as pulmonary embolism, large vessel occlusion or intracranial haemorrhage [17,18,19,20,21,22]. Combined computer-aided detection and workflow optimisation tools accounted for 50.0% of the studies (n = 17), supporting image interpretation and logistical operations such as exam routing, report prioritisation and radiographer positioning guidance [18,19,26,27]. Less frequent use cases included radiation dose management, outcome prediction and structured reporting enhancements [26].

3.3.4. Diagnostic Accuracy

In terms of diagnostic accuracy, the performance of AI applications was generally strong. The average sensitivity across the studies was 89.0%, reflecting a high likelihood of accurately identifying pathological conditions. The average specificity was 93.3%, indicating a low rate of false positives and a high level of precision in classification tasks. Clinical outcomes and operational impacts were frequently reported [19,22,23,26,27,32,33,34,35,36,37,38,39,40,41,42,43,44,49,51]. The studies demonstrated substantial workflow efficiency benefits, including reduced scan-to-report times, improved patient positioning and more standardised measurements [16,19,29,30]. One notable example reported a 28.0% reduction in CT positioning time, a 16.0% dose reduction and a 9.0% reduction in image noise using AI-assisted automatic positioning tools [20]. In other cases, AI detected pathologies missed during initial reads or contributed to more consistent inter-reader interpretation [22,30,31].

4. Discussion

The geographical distribution of the studies (Figure 2) revealed a predominant contribution from high-income countries, particularly Germany (n = 9) and the United States n = 7), with well-established digital health infrastructures that facilitate the integration of AI into CT radiology. These findings align with earlier reports showing accelerated AI adoption in well-resourced healthcare systems [4,6]. Publication trends reflect a growing academic and clinical interest in the integration of AI into CT workflows, particularly in the post-pandemic era, during which healthcare digitisation has become a strategic priority, with a peak in 2022 [3].

High sensitivity and specificity values indicate that, when adequately trained and validated, AI models can perform on par with radiologists in well-defined diagnostic tasks. The sensitivity and specificity in our dataset presented high values (Table 1). For example, CINA-PE [44] achieved a sensitivity of 91.4% and specificity of 91.5% for pulmonary embolism detection, while Zaazoue et al. [38] reported 93.2% sensitivity and 99.6% specificity in a similar context. In neurological imaging, Kiefer et al. [50] reached 98.1% sensitivity and 89.7% specificity for intracranial haemorrhage detection, and Rava et al. [47] achieved balanced values of 93% for both metrics. Although thoracic and neurological imaging currently dominate AI research in radiology, applications are progressively expanding into abdominal, cardiovascular and musculoskeletal domains. Overall, AI applications in radiology are becoming increasingly robust and multifaceted, demonstrating tangible value in triage, clinical decision support and workflow optimisation, especially when implemented as adjunct tools that enhance—rather than replace—the role of the radiologist.

The results suggest that AI tools are becoming increasingly capable and impactful, particularly when thoughtfully integrated into clinical workflows. One of the most consistent findings across the studies was improvements in diagnostic accuracy and efficiency. As shown in Table 3, 61.8% of the studies used AI for triage purposes. Dovrat et al. [41] achieved 94% accuracy in detecting large vessel occlusions, while Kundisch et al. [45] increased intracranial haemorrhage detection by 12.2% compared to initial radiology reports. An electronic “round-trip” model has been described, in which machine learning algorithms automate the entire process—from examination request to report communication—thereby reducing workflow friction and protocol variability [3]. Deep learning models, including convolutional neural networks, demonstrated substantial potential in detecting critical pathologies (such as pulmonary embolism, subdural haematoma and intracranial haemorrhage). For example, Cotena et al. [37] reduced scan-to-assessment time for acute aortic dissection by 68% through automated prioritisation, underscoring the potential of such systems to significantly enhance radiologists diagnostic capabilities [1].

Another key benefit consistently reported was workflow optimisation. AI has shown potential to accelerate triage, prioritise urgent findings and support early clinical decision making. These operational benefits were also evident in a replacement role in 26.5% of the studies (n = 9, Table 3). Gang et al. [20] demonstrated that AI-based automatic positioning reduced setup time by 28%, radiation dose by 16% and image noise by 9%, directly contributing to occupational safety by lowering physical strain and exposure risk. Postiglione et al. [23] and Nadeem et al. [29] similarly reduced manual measurement workloads through segmentation automation.

In stroke care, the Viz LVO system achieved an accuracy of 94% in detecting large vessel occlusions [41], with authors highlighting its promise in improving early stroke identification. Several of the studies also demonstrated the capacity of AI to augment human performance and, rather than replacing radiologists, AI frequently acted as a second reader or safety net. For instance, one algorithm identified an additional 12.2% of intracranial haemorrhages compared to standard radiologist interpretations [45]. However, the limitations of retrospective study designs and single-centre cohorts were noted [40]. In their work on subdural haematoma detection, although the Viz.ai system performed well in haematoma quantification, factors such as false positives from pericerebral spaces and reliance on axial-only measurements constrained its generalisability, aligning with observations from previous studies [7].

AI-based triage systems assist in identifying and prioritising urgent cases by providing an additional layer of review and enabling rapid alerts to radiologists in emergencies such as bleeding or embolism, reinforcing previous findings [3,4]. These systems contribute to faster clinical responses and help to create a more manageable workload by alleviating the cognitive burden associated with high-stakes decision making. In addition to triage support, other AI tools have taken on a replacement role by automating repetitive or time-consuming tasks (e.g., image measurements and reformatting). Although these systems do not replace radiologists’ clinical judgement, they help to reduce the physical demands of routine work, potentially mitigating fatigue and lowering the risk of work-related musculoskeletal issues over time. When used as second readers or assistant tools, AI systems offer valuable cognitive support by acting as a safety net in complex or borderline cases, ultimately reducing decision-making anxiety and enhancing diagnostic confidence.

The relationship between these findings and Table 1 is evident. Tekin et al. [26] developed predictive models for the dose length product in abdominal CT scans, facilitating personalised protocol optimisation and minimising unnecessary radiation exposure for both patients and staff. This enhances patient safety while reducing scatter radiation risks for radiology personnel. Gang et al. [20] improved ergonomics through AI-based automatic positioning, decreasing setup time by 28% and boosting accuracy to 99%. These advancements alleviate physical strain from repetitive patient handling. Freedman et al. [32] reported a 50% reduction in image processing times with AI-generated multiplanar reformatted images, maintaining diagnostic quality and resulting in less screen time and cognitive load. Together, these examples illustrate how AI contributes to reducing occupational risks, enhancing ergonomics, improving radiological safety and effectively managing workloads in accordance with international health recommendations.

A significant insight from this review is the strong association between the mode of AI integration into radiology workflows and its potential to reduce occupational risks. In other words, the way AI is implemented affects not only patient outcomes but also the well-being of healthcare professionals who interact with these systems daily. AI technologies in radiology are evolving from experimental innovations to practical, supportive systems. Their key strengths include enhancing diagnostic consistency, reducing interpretation time and expanding diagnostic reach (particularly in high-volume or high-risk settings). Nonetheless, to fully realise their potential and mitigate new risks introduced by automation, robust external validation, clear regulatory frameworks and collaborative integration with radiologists remain essential, in alignment with findings from other studies [5,6].

4.1. Limitations

Despite the growing body of evidence supporting AI applications in radiology, it is important to acknowledge that several of the included studies presented methodological limitations. Many relied on non-random retrospective datasets from single institutions, which limited sample diversity and may have led to performance overestimation (particularly when AI models were trained and tested on similar populations). In addition, some of the studies lacked independent ground truth verification by multiple experts, relying instead on single-reader validation or consensus among system developers. Several AI tools were also evaluated under ideal non-real-time conditions, which limits the generalisability of their reported performance to routine clinical practice. Finally, a number of publications were authored or co-authored by system developers or industry collaborators, without fully disclosing the extent of commercial involvement, raising concerns about potential conflicts of interest.

4.2. Future Research Directions

Future research should prioritise prospective multicentre trials to rigorously assess AI tools within real-world clinical environments. Emphasis must be placed on achieving robust external validation and seamless integration of AI applications into existing radiology workflows. Large-scale studies are needed to improve the diversity and volume of training data, enhancing the fairness, generalisability and overall effectiveness of AI systems across different populations and clinical contexts. Moreover, future investigations should explicitly quantify occupational health benefits by measuring relevant outcomes such as clinician fatigue, musculoskeletal complaints and cognitive overload. Incorporating these endpoints would provide a more comprehensive understanding of AI’s potential impact, supporting evidence-based justification for investments in AI technologies from both human resources and well-being perspectives.

Although a meta-analysis was not performed in the current systematic review, future investigations may consider incorporating quantitative synthesis methods to better characterise the diagnostic accuracy of AI tools within radiology workflows. Such approaches could provide more robust estimates of performance and help clarify the potential of AI to reduce diagnostic workload, support clinical decision making and ultimately contribute to occupational risk mitigation. As research in this area progresses, meta-analytical evidence may play a valuable role in guiding implementation strategies and informing policy decisions regarding AI integration in radiology departments.

5. Conclusions

AI has the potential to fundamentally reshape radiology workflows by enhancing diagnostic accuracy, reducing reporting times and mitigating both occupational and clinical risks. Beyond improving diagnostic quality, AI can simultaneously streamline administrative processes and enhance the overall patient experience. However, the successful implementation of AI technologies depends on rigorous validation, adherence to ethical safeguards and thoughtful integration within existing healthcare systems. Although recent studies demonstrate promising outcomes, they also expose critical areas for improvement, particularly regarding methodological transparency, sample diversity and the optimisation of human–machine interaction models. The future of AI in radiology lies in amplifying the capabilities of healthcare professionals within a safe, data-driven environment. Importantly, AI can contribute to improved occupational health by reducing the mental stress associated with urgent decision making, alleviating the physical strain caused by repetitive tasks and easing cognitive load.