Autonomous chest x-ray image classification, capabilities and prospects: rapid evidence assessment

Abstract

Background:

Screening methods are essential for detection of numerous pathologies. Chest x-ray radiography (CXR) is the most widely used screening modality. During the screening, radiologists primarily examine normal radiographs, which results in a substantial workload and an increased risk of errors. There is an increasing necessity to automate radiological screening in order to facilitate the autonomous sorting of normal studies.

Objective:

We aimed to evaluate the capabilities of artificial intelligence (AI) techniques for the autonomous CXRs triage and to assess their potential for integration into routine clinical workflow.

Methods:

A rapid evidence assessment methodology was employed to conduct this review. Literature searches were performed using relevant keywords across PubMed, arXiv, medRxiv, Elibrary, and Google Scholar covering the period from 2019 to 2025. Inclusion criteria comprised large-scale studies addressing multiple pathologies and providing abstracts in English. Meta-analysis was conducted using confusion matrices derived from reported diagnostic performance metrics in the selected studies. Methodological quality and the overall quality of evidence were assessed using a combination of QUADAS-2, QUADAS-CAD, and GRADE frameworks.

Results:

Out of 327 records, 11 studies met the inclusion criteria. Among these, three studies analyzed datasets reflecting the real-world prevalence of pathologies. Three studies included very large cohorts exceeding 500,000 CXRs, whereas the remaining studies used considerably smaller samples. The proportion of autonomously triaged CXRs ranged from 15.0% to 99.8%, with a weighted average of 42.3% across all publications. Notably, in a study conducted under real-world clinical conditions on continuous data flow, this proportion was 54.8%. Sensitivity was 97.8% (95% CI: 94.8%–99.1%), and specificity was 94.8% (95% CI: 53.0%–99.7%). Fifty-five percent of the studies were classified as having a low risk of bias. Primarily, elevated risk of bias and heterogeneity of results were attributed to variability in sample selection criteria and reference standard evaluation.

Conclusions:

Modern AI systems for autonomous triage of CXRs are ready to be implemented in clinical practice. AI-driven screening can reduce radiologists' workload, decrease sorting errors and lower the costs associated with screening programs. However, implementation is often hindered by regulatory and legislative barriers. Consequently, comprehensive clinical trials conducted under real-world conditions remain scarce.

1 Introduction

1.1 Rationale

Chest radiography is one of the most frequently performed preventive radiological examinations. However, the ongoing shortage of radiologists (1–4) and the subsequent adoption of new x-ray equipment, makes the optimization of radiological workflows for screening studies increasingly critical. The vast majority of these preventive diagnostic imaging studies reveal an absence of pathological findings. The reported prevalence of pathologies in chest radiographs ranges from 1% to 20% (5–8). This low prevalence implies that radiologists spend most of their time interpreting “normal” examinations, which may inadvertently increase the risk of missing abnormalities. This phenomenon is an example of cognitive bias, where a radiologist, upon detecting an initial finding or confirming the absence of pathology, prematurely terminates the image analysis. Several factors contribute to such diagnostic errors, including «satisfaction of search», «localisation errors», «lack of thorough structural analysis», and «data shortage» (9).

In order to mitigate the risk of missed pathologies and reduce the workload of radiologists, clinical decision support systems—particularly those employing artificial intelligence (AI)—have gained widespread adoption. In the context of screening examinations, a notable approach is the autonomous triage of chest radiographs (CXRs), whereby AI algorithms differentiate normal from abnormal images without direct radiologist involvement. In practice, this means the automated exclusion of routine CXRs, enabling radiologists to concentrate on more complex cases. The primary objective of autonomous triage is to alleviate the burden on radiologists and reduce reporting delays while maintaining patient safety (10, 11). Over the past five years, numerous studies have investigated AI systems for autonomous chest radiograph triage.

1.2 Objectives

The present study aimed to evaluate the capabilities of contemporary AI methods for the autonomous triage of CXRs and assess their prospects for integration into routine clinical practice. Although numerous publications have explored AI for the detection and screening of individual diseases, such narrowly focused strategies have limited applicability in practical screening workflows.

In this study, we conducted a selective review of scientific literature to determine the performance and applicability of modern AI for autonomous chest radiograph triage. We employed a Rapid Evidence Assessment (REA) (12) methodology to enable a rapid yet comprehensive evaluation of the feasibility and clinical utility of autonomous triage systems in current real-world practice.

2 Methods

The prospective, unregistered study protocol is provided in the Supplementary Material.

2.1 Literature search

The article search was conducted for the period from 2019 to 2025. All articles and preprints with an English abstract were considered. Searches were performed across multiple databases: PubMed, arXiv, medRxiv, Elibrary. An additional literature search was conducted using the Google Scholar search engine.

The search query used in PubMed was as follows:

(“Radiography, Thoracic”[Mesh] OR Thoracic Radiograph*[tiab] OR (x-ray[tiab] AND (chest[tiab] OR thora*[tiab])) OR CXR) AND (“Radiographic Image Interpretation, Computer-Assisted”[Mesh] OR Artificial Intelligence[tiab] OR Computer vision[tiab] OR (learning[tiab] AND (machine[tiab] OR deep[tiab]))) AND (Autonom*[tiab] OR standalone[tiab] OR unaided[tiab] OR unassisted[tiab]).

The search query for arXiv was: “Autonomous AI AND chest x-ray OR CXR AND triage”. The search query for medRxiv, Elibrary, and Google Scholar was: “Autonomous AI chest x-ray triage”.

2.2 Study selection

The present study incorporated all screening studies that used CXRs to analyze multiple pathologies. A two-stage approach was employed for study selection. In the initial phase, two independent reviewers screened titles and abstracts to identify the most relevant studies to our research question. In the subsequent stage, the full texts of articles that passed the initial screening were independently reviewed by two researchers, who made inclusion decisions.

The search and selection of articles were carried out by two researchers, each with over 10 years of experience in medical informatics.

2.3 Data extraction

The following information was extracted from the selected studies: bibliometric data, country of study, research objectives, study design (prospective/retrospective, multicenter/single-center, binary classification/prioritization), number of abnormalities, employed AI algorithm, inclusion and exclusion criteria, reference standard, assessment of time benefits and economic efficiency, all values of AI diagnostic accuracy, values of true positives (TP), false negatives (FN), false positives (FP), true negatives (TN), and comparison of AI diagnostic accuracy with that of physicians.

Information extraction was conducted independently by two researchers, each with over 10 years of experience in medical informatics.

2.4 Meta-analysis

The number of included and autonomously triaged CXRs for each study was reported.

Components of the confusion matrix that were not reported in the publications were independently calculated based on the reported diagnostic performance metrics, following the methodology suggested by (13). In studies employing prioritization, the category designated as “normal” was classified as normal, whereas all other categories (“non-urgent”, “urgent”, and “critical”) were classified as abnormal. Subsequently, sensitivities and specificities were calculated for each study with corresponding 95% confidence intervals (CIs). Paired forest plots displayed individual study estimates alongside pooled meta-analysis results for sensitivity and specificity. Additionally, a receiver operating characteristic (ROC) plot of sensitivity vs. 1-specificity summarized diagnostic performance based on meta-analysis data.

The meta-analysis was conducted using the MetaDTA tool (14).

The expected prevalence level of abnormalities was calculated per 10,000 cases based on minimal, maximal, and real-world findings in the selected studies.

A sensitivity analysis was conducted across two scenarios. In the first, two studies employing prioritization were excluded from the overall sample. The second scenario excluded the two studies with the highest sensitivity and specificity values (occupying the upper left corner on the ROC plot) from the overall sample.

Subgroup and random effects analyses were conducted to investigate potential sources of heterogeneity. The analyses were performed using the metafor package for R, employing the REML (Restricted Maximum Likelihood) model for small samples. The factors used were sample size (as a continuous variable), year of study publication (as a continuous variable), and population (as a discrete binary variable: general or hospital).

2.5 Assessment of methodological quality

Two review authors independently assessed the methodological quality of included studies using a combination of two instruments: QUADAS-2 (15) and the AI-adapted QUADAS-CAD (16, 17). We assessed each of the four domains (Patient Selection, Index Test, Reference Standard, Flow and Timing) in terms of risk of bias using QUADAS-CAD, and the first three domains in terms of concerns regarding applicability to the review question using QUADAS-2.

Any disagreements were resolved through discussion to reach consensus. If needed, a third author facilitated consensus.

In the Patient Selection domain, we assessed whether the screening samples reflected the distribution of normal and pathological cases in the real-world population. In the Index Test domain, we analyzed how the index test was applied. In the Reference Standard domain, we evaluated the types of reference standards used, taking into account the presence of multiple pathologies in the samples. In the Flow and Timing domain, we assessed whether the reference standards were consistently applied across all cases and examined the overall transparency of the results.

We visually examined forest plots and ROC plots for heterogeneity and discussed potential data heterogeneity sources. These estimations were further used in the assessment of certainty of evidence. We assessed the certainty of evidence using the GRADE methodology (18, 19).

3 Results

3.1 Results of the search and study selection

327 records were identified (Figure 1). Of these, 40 records were identified in PubMed, 1 in arXiv, 14 in medRxiv, 7 in Elibrary, and 265 in Google Scholar.

Figure 1

311 records were excluded as irrelevant to the classification of CXRs. After the initial screening, 16 articles underwent full-text analysis: 5 from PubMed, 1 from arXiv, 0 from medRxiv, 1 from Elibrary, and 9 from Google Scholar.

During the full-text review, 5 articles were excluded because they focused on the detection of a single pathology (Appendix 1). Consequently, 11 articles were included in the systematic review: 3 from PubMed, 1 from arXiv, 0 from medRxiv, 1 from Elibrary, and 6 from Google Scholar.

In the study by (20), several distinct datasets were employed. For inclusion in our review, we selected two datasets relevant to our research focus (DS-1 and CXR-14), which involved multiple pathologies, and excluded model datasets containing only a single target pathology (e.g., tuberculosis or COVID-19).

Confusion matrices were reported in two publications (10, 21). For the remaining studies, we independently calculated confusion matrices based on the reported diagnostic performance metrics, following the methodology described in (13).

3.2 Main characteristics of the studies

The main characteristics of the selected publications are presented in Table 1 (general study information) and Table 2 (quantitative characteristics).

Only two of the 11 studies were prospective and conducted on streaming data (22, 23). Among these, only one large-scale study was performed under real-world conditions (22). Of the 11 publications, 10 were multicenter studies, and one was a single-center study (23).

Real-world population data were employed in three studies (22, 24, 25). In other cases, normal/abnormal distribution in employed samples could differ from general population: patients from hospitals (10, 11, 20, 21, 23, 26, 27) including those specializing in pulmonary pathologies (10), as well as migrant populations (28).

Most studies (9 out of 11) employed binary classification distinguishing between normal and abnormal categories. The abnormal category predominantly encompassed multiple pathologies (ranging from 9 to 75 categories) (10, 11, 22, 24–26, 28), while in some studies, the number of pathologies was unspecified (20, 27).

Two studies utilized prioritization schemes: normal, non-urgent, and urgent (23), or normal, non-urgent, urgent, and critical (21).

Most studies included CXRs of adult patients; however, two studies (23, 25) also included CXRs of patients under 18 years of age.

The largest datasets, each comprising over half a million samples, were employed in three publications (28): with 1,309,443 CXRs (accounting for 44.5% of the total review sample) (25), with 1,012,851 CXRs (34.4%), and (22) with 575,549 CXRs (19.5%).

The sample sizes of the remaining publications (10, 11, 20, 21, 23, 24, 26, 27) were considerably smaller, ranging from 810 to 20,944 CXRs (contributing from 0.03% to 0.71%, respectively).

3.3 Meta-analysis

The proportion of autonomously triaged CXRs varied substantially across studies, ranging from 15.0% to 99.8% (Table 2). In the largest screening studies, the number of autonomously triaged CXRs was notably high. For instance, in a real-world data stream study by (22) the proportion of autonomously triaged CXRs was 54.8%. In retrospective studies with sample sizes exceeding one million (25, 28), the proportion reached 99.8% and 71.5%, respectively.

Estimates of sensitivity and specificity are presented and summarized in Table 2. Sensitivity was consistently high across all studies, with a median value of 97.8% (95% CI: 94.8%–99.1%). Specificity was also generally high on average but exhibited considerable variability, with a median of 94.8% (95% CI: 53.0%–99.7%). In the first sensitivity analysis scenario, the median sensitivity was 97.8% (95% CI: 94.8–99.1%), and the median specificity was 94.8% (95% CI: 53.0%–99.7%). In the second scenario, the median sensitivity was 97.6% (95% CI: 93.5%–99.1%), and the median specificity was 97.5% (95% CI: 61.7%–99.9%).

Summary results regarding the variability of key diagnostic metrics—sensitivity and specificity—are illustrated in the forest plots (Figure 2) and the ROC curve plot (Figure 3). Notably, substantial heterogeneity was observed in sensitivity estimates and especially in specificity estimates.

Figure 2

Figure 3

In most cases, relatively narrow confidence intervals were reported, which is typical for studies with large sample sizes. However, wide confidence intervals for diagnostic performance were observed in (26) and (20).

The highest simultaneous sensitivity and specificity values were reported in the population-based real-time study by (22) and the hospital-based dataset by (27).

Additionally, sensitivity values exceeding 95.0% were found in the studies by (10, 11, 20, 23, 24, 28), although these were accompanied by a substantial reduction in specificity, ranging from 28.0% to 79.0%.

In (26), sensitivity estimates demonstrated a relatively wide confidence interval and low specificity (60%). Conversely, a high specificity of 100.0% was observed in (25), albeit with a relatively low sensitivity of 67.0%.

Analysis of variability without factors revealed substantial between-study heterogeneity: (estimated amount of residual heterogeneity) = 0.0005, (residual heterogeneity/unaccounted variability) = 100.0%. The Q test for heterogeneity also indicated significant between-study heterogeneity (p < 0.01).

Incorporating of sample size as a factor into the model left the unexplained between-study heterogeneity high: τ² = 0.0006, I² = 100.0%, and (amount of heterogeneity accounted for) = 0.0%. The intercept was statistically significant in the model (0.97, p < 0.01), while the coefficient for the factor was not (0.0041, p = 0.93).

Including year of publication as a factor yielded a nonsignificant effect: τ² = 0.0004, I² = 22.5%, R² = 20.2%. The intercept (−15.0617, p = 0.54) and factor coefficient (0.0079, p = 0.51) were both nonsignificant.

The population factor effectively explained the differences between studies: τ² = 0.0, I² = 0.0%, R² = 100.0%. In the model, the intercept (0.96, p < 0.01) and factor coefficient (0.04, p < 0.01) were both statistically significant. The subgroup effect estimates for population were as follows: “general” subgroup — sensitivity 0.999 ± 0.0, specificity 1.0 ± 0.0; “hospital” subgroup — sensitivity 0.962 ± 0.0, specificity 0.7 ± 0.06.

3.4 Quality assessment

The risk of bias assessment results according to QUADAS-CAD and QUADAS-2 is presented in Figure 4. A low risk of bias was assigned to six (55%) publications (11, 20, 22, 24, 26, 27). A high risk of bias was assigned to five (45%) publications (10, 21, 23, 25, 28).

Figure 4

The risk of bias in the Patient Selection domain was due only to a few screening studies using samples representative of the real-world population (22, 24, 25). In other cases, the sample composition was either biased (10, 26, 28) or there was a potential risk of bias (11, 20, 21, 23, 27).

A high risk of bias in the Flow and Timing domain was identified in three (27%) publications (21, 23, 25) due to concerns that not all cases employed the same reference standard and, correspondingly, due to unclear timing and transparency in obtaining results.

We calculated prevalence using the GRADE tool, based on the expected rate of CXRs observed in a real-world population (22) (45%), as well as minimum and maximum values.

The overall certainty of the evidence was rated as low owing to a high risk of bias affecting 45% of the included studies and the substantial heterogeneity of diagnostic estimates (Table 3), which was caused by significant differences in the analyzed populations.

3.5 Assessment of temporal and economic benefits, comparison of diagnostic accuracy results with physicians

The assessment of temporal benefits of automated triage compared to triage performed by physicians was conducted in one study (23). The study (23) reports that the mean difference between AI and radiologist turnaround time (defined as the interval from CXR upload to the system until approval of the radiology report by the radiologist) was 818.9 min, which was statistically significant (p < 0.05, Wilcoxon signed-rank test).

Economic efficiency was evaluated in a single study (22). The study (22) demonstrated economic viability, with costs reduced by 43.7% over a 5-month experimental period.

Diagnostic accuracy between AI and radiologists was compared in three studies (10, 20, 27). Overall, AI demonstrated superior diagnostic performance compared to radiologists.

In the (10), AI had significantly higher sensitivity for detecting abnormal CXRs than clinical radiology reports, with sensitivities of 99.1% vs. 72.3% respectively (p < 0.001).

The study (20) reported sensitivity value of 98.0% for AI and 48.0% for radiologists for the DS-1 dataset, with specificities of 38.0% and 96.0%, respectively. For the CXR-14 dataset, sensitivity was 95.0% for AI and 54.0% for radiologists, with specificity of 71.0% and 94.0%, respectively.

In the study by (27), the false negative (FN) rate for AI was 0.33%, which was substantially lower than the 13.5% FN rate observed for radiologists.

4 Discussion

The proportion of studies that are autonomously triaged varies widely across different publications (15.0%–99.8%) and is largely determined by whether the sample was obtained through screening a natural population or represents a biased cohort with an increased risk of pathology detection.

For studies conducted on samples reflecting the natural population (22, 24, 25), the proportion of autonomously triaged studies exceeds 40%, which closely aligns with the overall average of 42.3% across all publications. In the recent study by (22) involving real-world data and a large sample of over half a million cases, the proportion of autonomously triaged CXRs was found to be 54.8%. Consequently, this value can be considered a natural and practically expected benchmark. A subgroup analysis revealed that the differences between studies conducted in general and hospital populations accounted for a significant amount of the observed heterogeneity.

On average, AI algorithms demonstrated high sensitivity across studies: 97.8% (95% CI: 94.8–99.1%). The sensitivity level of radiologists when analyzing CXRs, which serves as a benchmark, ranges from 53.6% to 95.5% and is highly dependent on the radiologists' experience and the type of pathology (29, 30). In contrast to radiologists, AI algorithms can be finely tuned by adjusting the threshold to balance sensitivity and specificity. Theoretical research on a limited dataset has demonstrated the feasibility of tuning AI services to achieve 100% sensitivity while maintaining 77.4% specificity (31). In a prospective experimental study of 209,497 screening CXRs, the false-negative rate of the AI was significantly lower (0.04%) than that of radiologists (8.6%) (32). The AI also autonomously triaged 55.9% (117,041) of CXRs in this study. Maximizing sensitivity inevitably results in decreased specificity. Nevertheless, across the majority of analyzed studies, AI also exhibited an acceptable specificity level of 94.8% (95% CI: 53.0–99.7%). Direct comparisons of the sensitivity and specificity of AI vs. radiologists in the analyzed studies (10, 20, 27) show that the performance of AI screening was not inferior to that of radiologists.

The sensitivity analysis revealed no distortion of the estimated sensitivity and specificity values when extreme values were excluded from the overall sample or when studies that used prioritization instead of binarization were excluded.

Most studies used either small datasets (72.7%) or retrospective designs (81.8%), which contributed to heterogeneity and limited the quality of the evidence. Consequently, considerable interest is directed towards studies conducted in real-world settings (22, 23) with large datasets (22, 25, 28), especially those that satisfy both criteria (22). The high complexity of large-scale screening experiments under real-world conditions stems from the stringent regulatory environments in many countries (33).

Currently, only AI systems that are registered as medical devices can be integrated into clinical practice. Prospective studies only include a limited number of AI solutions, such as the ChestEye algorithm by Oxipit (34), LUNIT INSIGHT CXR by Lunit (23), and QXR by Qure.ai (28). However, large-scale studies addressing financing models and the integration of autonomous triage into national healthcare systems are lacking for these solutions.

The temporary benefits of autonomous triage are often evaluated indirectly. Many researchers report a reduced workload for radiologists, primarily based on the theoretical savings of resources derived from the proportion of CXRs that are autonomously triaged. A prospective study involving 20,944 CXRs assessed time savings, approximated as potentially freed radiologist time (77%) (23).

The actual economic impact has only been evaluated in one study (22), largely due to the legislative and infrastructural environment established by the Moscow Experiment — the largest global, prospective, multicenter study on the applicability, safety and quality of AI in real clinical settings (35–40). Economic efficiency calculations incorporated the costs of reporting and interpreting each CXR with and without autonomous triage, as well as the expenses involved in acquiring and maintaining the AI system. The proposed autonomous triage model yielded a 43.7% reduction in costs. Additionally, a protocol was proposed to eliminate clinically significant errors by implementing dual automated reviews, whereby two independent AI medical devices analyze each diagnostic image simultaneously, with the final decision favoring the patient. While this approach would double the interpretation fee, it is projected to maintain a 40.4% economic benefit.

Selective literature review indicates that AI is highly ready for the autonomous triage of screening CXRs due to its high diagnostic performance. This would enable a reduction in radiologists' workload, optimization of workflows, and budgetary savings. The proposed methodology can be scaled up to various AI solutions for CXR interpretation. The real-world experiment conducted on large datasets was successful (22). However, more studies in real clinical settings are needed to robustly assess the potential benefits and challenges of implementing this technology into practice.

Our review focuses on a specific aspect: the ability of modern AI to autonomously identify normal examinations during screening. The next step in determining the readiness of AI for clinical implementation in CXR analysis is to evaluate its ability to classify chest pathologies (41), detect single pathology (42, 43), and foreign objects (44). AI has already demonstrated impressive diagnostic performance in these areas.

This review has several limitations. We selected only the most comprehensive studies that included analyses of multiple pathologies with English abstracts. The number of such studies is rather limited. Furthermore, there is a lack of research conducted in real-world conditions.

5 Conclusions and practical recommendations

A systematic review of studies has demonstrated that modern AI systems enable safe autonomous triage of CXRs, allowing the identification of on average over 40% of normal examinations, with a mean sensitivity of up to 97.8%. In conclusion, the diagnostic performance of AI in the preventive screening of thoracic diseases and the autonomous sorting of CXRs is highly promising.

However, there is a need for large-scale studies based on real-world data, which to date have been reported only in limited publications. This limitation is associated with multiple evident legislative and infrastructural barriers that have been overcome in only a few countries.

The implementation of AI for autonomous triage has been demonstrated to optimize workflow processes and reduce the routine workload of radiologists. The implementation of an autonomous triage approach within public healthcare systems has been demonstrated to result in a substantial reduction in financial costs, with a reported decrease of 43.7%. The proposed double parallel autonomous triage model is expected to completely mitigate clinically significant errors.

References

• 1.

  Rimmer A . Radiologist shortage leaves patient care at risk, warns royal college. Br Med J. (2017) 359:j4683. 10.1136/bmj.j4683

  • CrossRef
  • Google Scholar

• 2.

  Golubev NА Ogryzko EV Tyurina EM Shelepova EA Shelekhov PV . Features of the development of the radiation diagnostics service in the Russian federation for 2014–2019. Curr Prob Health Care Med Stats. (2021) 2:356–76. 10.24412/2312-2935-2021-2-356-376

  • CrossRef
  • Google Scholar

• 3.

  Shelekhov PV . Personnel situation in radiative diagnostics. Curr Prob Health Care Med Stats. (2019) 1:265–75. 10.24411/2312-2935-2019-10018

  • CrossRef
  • Google Scholar

• 4.

  Chew C O'Dwyer PJ Young D . Radiology and the medical student: do increased hours of teaching translate to more radiologists?BJR Open. (2021) 3(1):20210074. 10.1259/bjro.20210074

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 5.

  Bouck Z Mecredy G Ivers NM Pendrith C Fine B Martin D et al Routine use of chest x-ray for low-risk patients undergoing a periodic health examination: a retrospective cohort study. CMAJ Open. (2018) 6(3):E322–9. 10.9778/cmajo.20170138

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 6.

  Twabi HH Semphere R Mukoka M Chiume L Nzawa R Feasey HRA et al Pattern of abnormalities amongst chest x-rays of adults undergoing computer-assisted digital chest x-ray screening for tuberculosis in Peri-Urban Blantyre, Malawi: a cross-sectional study. Trop Med Int Health. (2021) 26(11):1427–37. 10.1111/tmi.13658

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 7.

  Markelov Y Shchegоleva LV . Clinical and economic aspects of tuberculosis detection during mass fluorographic examinations of the population. J Radiol Nuclear Med. (2021) 102(3):148–54. 10.20862/0042-4676-2021-102-3-148-154

  • CrossRef
  • Google Scholar

• 8.

  Markelov Y Schegoleva LV . Evaluation of clinical and economic efficiency and impact of mass fluorography screening on tuberculosis epidemiological rates in four federal districts of the Russian federation with different levels of population coverage with mass fluorography screening. Tuberculosis Lung Dis. (2023) 101(1):8–16. 10.58838/2075-1230-2023-101-1-8-16

  • CrossRef
  • Google Scholar

• 9.

  Nechaev VA Vasilev A . Risk factors for development of perceptual errors in radiology. Vestnik SurGU Meditsina. (2024) 17(4):14–22. 10.35266/2949-3447-2024-4-2

  • CrossRef
  • Google Scholar

• 10.

  Plesner LL Müller FC Nybing JD Laustrup LC Rasmussen F Nielsen OW et al Autonomous chest radiograph reporting using ai: estimation of clinical impact. Radiology. (2023) 307(3):e222268. 10.1148/radiol.222268

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 11.

  Blake SR Das N Tadepalli M Reddy B Singh A Agrawal R et al Using artificial intelligence to stratify normal versus abnormal chest x-rays: external validation of a deep learning algorithm at east Kent hospitals university NHS foundation trust. Diagnostics. (2023) 3(22):3408. 10.3390/diagnostics13223408

  • CrossRef
  • Google Scholar

• 12.

  Barends E Rousseau DM Briner RB (Eds). CEBMA Guideline for Rapid Evidence Assessments in Management and Organizations, Version 1.0. Amsterdam: Center for Evidence Based Management (2017). p. 39. Available online at:https://cebma.org/resources/frequently-asked-questions/what-is-a-rapid-evidence-assessment-rea/(Accessed May 16, 2025).

  • Google Scholar

• 13.

  Vasilev YuA Vladzymyrskyy AV Omelyanskaya OV Reshetnikov RV Shumskaya YuF . Guidelines for writing a systematic review. Part 2: studies screening, data extraction. Moscow: State Budget-Funded Health Care Institution of the City of Moscow “Research and Practical Clinical Center for Diagnostics and Telemedicine Technologies of the Moscow Health Care Department”. (2023). p. 32. Available online at:https://telemedai.ru/biblioteka-dokumentov/metodicheskie-rekomendacii-po-podgotovke-sistematicheskogo-obzora-chast-2-skrining-issledovanij-izvlechenie-dannyh(Accessed May 16, 2025).

  • Google Scholar

• 14.

  Patel A Cooper N Freeman S Sutton A . Graphical enhancements to summary receiver operating characteristic plots to facilitate the analysis and reporting of meta-analysis of diagnostic test accuracy data. Res Syn Meth. (2021) 12:34–44. 10.1002/jrsm.143944

  • CrossRef
  • Google Scholar

• 15.

  Whiting PF Rutjes AW Westwood ME Mallett S Deeks JJ Reitsma JB et al QUADAS-2: a revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med. (2011) 155(8):529–36. 10.7326/0003-4819-155-8-201110180-0000

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 16.

  Kodenko MR Vasilev Y Vladzymyrskyy AV Omelyanskaya OV Leonov DV Blokhin IA et al Diagnostic accuracy of AI for opportunistic screening of abdominal aortic aneurysm in CT: a systematic review and narrative synthesis. Diagnostics. (2022) 12:3197. 10.3390/diagnostics12123197

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 17.

  Kodenko MR Reshetnikov RV Makarova TA . Modification of quality assessment tool for artificial intelligence diagnostic test accuracy studies (QUADAS-CAD). Digit Diagnostics. (2022) 3(S1):4–5. 10.17816/DD105567

  • CrossRef
  • Google Scholar

• 18.

  Schünemann HJ Mustafa RA Brozek J Steingart KR Leeflang M Murad MH et al GRADE Guidelines: 21 part 1. Study design, risk of bias, and indirectness in rating the certainty across a body of evidence for test accuracy. J Clin Epidemiol. (2020) 122:129–41. 10.1016/j.jclinepi.2019.12.020

  • CrossRef
  • Google Scholar

• 19.

  Schünemann HJ Mustafa RA Brozek J Steingart KR Leeflang M Murad MH et al GRADE Guidelines: 21 part 2. Test accuracy: inconsistency, imprecision, publication bias, and other domains for rating the certainty of evidence and presenting it in evidence profiles and summary of findings tables. J Clin Epidemiol. (2020) 122:142–52. 10.1016/j.jclinepi.2019.12.021

  • CrossRef
  • Google Scholar

• 20.

  Nabulsi Z Sellergren A Jamshy S Lau C Santos E Kiraly AP et al Deep learning for distinguishing normal versus abnormal chest radiographs and generalization to two unseen diseases tuberculosis and COVID-19. Sci Rep. (2021) 11:15523. 10.1038/s41598-021-93967-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 21.

  Annarumma M Withey SJ Bakewell RJ Pesce E Goh V Montana G . Automated triaging of adult chest radiographs with deep artificial neural networks. Radiology. (2019) 291(1):196–202. 10.1148/radiol.2018180921

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 22.

  Vasilev Y Sychev DA Bazhin AV Shulkin IM Vladzymyrskyy AV Golikova A et al Autonomous artificial intelligence for sorting results of preventive radiological examinations of chest organs: medical and economic efficiency. Digit Diagnostics. (2025) 6(1):5–22. 10.17816/DD641703

  • CrossRef
  • Google Scholar

• 23.

  Sridharan S Xin Hui AS Venkataraman N Tirukonda PS Jeyaratnam RP John S et al Real-world evaluation of an AI triaging system for chest x-rays: a prospective clinical study. Eur J Radiol. (2024) 181:111783. 10.1016/j.ejrad.2024.111783

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 24.

  Yoo H Kim EY Kim H Choi YR Kim MY Hwang SH et al Artificial intelligence-based identification of normal chest radiographs: a simulation study in a multicenter health screening cohort. Korean J Radiol. (2022) 23(10):1009–18. 10.3348/kjr.2022.0189

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 25.

  Subramanian B Jaikumar S Kumarasami PSN Sivasailam K Keerthana ADR Mounigasri M et al Autonomous AI for multi-pathology detection in chest x-rays: a multi-site study in Indian healthcare system. arXiv [preprint]. (2025). Available online at:https://arxiv.org/html/2504.00022v1(Accessed August 04, 2025)

  • Google Scholar

• 26.

  Schalekamp S van Leeuwen K Calli E Murphy K Rutten M Geurts B et al Performance of AI to exclude normal chest radiographs to reduce radiologists' Workload. Eur Radiol. (2024) 34:7255–63. 10.1007/s00330-024-10794-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 27.

  Dyer T Dillard L Harrison M Morgan TN Tappouni R Malik Q et al Diagnosis of normal chest radiographs using an autonomous deep-learning algorithm. Clin Radiol. (2021) 76(6):473.e9–e15. 10.1016/j.crad.2021.01.015

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 28.

  AlJasmi AAM Ghonim H Fahmy ME Nair A Kumar S Robert D et al Post-deployment performance of a deep learning algorithm for normal and abnormal chest x-ray classification: a study at visa screening centers in the United Arab Emirates. Eur J Radiol Open. (2024) 13:100606. 10.1016/j.ejro.2024.100606

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 29.

  Bradley SH Bhartia BSK Callister MEJ Hamilton WT Hatton NLF Kennedy MPT et al Chest x-ray sensitivity and lung cancer outcomes: a retrospective observational study. Br J Gen Pract. (2021) 71(712):e862–8. 10.3399/BJGP.2020.1099

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 30.

  Hossain R Wu CC de Groot PM Carter BW Gilman MD Abbott GF . Missed lung cancer. Radiol Clin North Am. (2018) 56(3):365–75. 10.1016/j.rcl.2018.01.004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 31.

  Vasilev Y Tyrov IA Vladzymyrskyy AV Arzamasov KM Pestrenin LD Shulkin IM . A new model of organizing mass screening based on stand-alone artificial intelligence used for fluorography image triage. Public Health Life Environ. (2023) 31(11):23–32. 10.35627/2219-5238/2023-31-11-23-32

  • CrossRef
  • Google Scholar

• 32.

  Vasilev Y Tyrov IA Vladzymirskyy AV Shulkin IM Arzamasov KM . Autonomous artificial intelligence for sorting the preventive imaging studies’ results. Russian J Prev Med. (2024) 27(7):23–9. 10.17116/profmed20242707123

  • CrossRef
  • Google Scholar

• 33.

  Mennella C Maniscalco U De Pietro G Esposito M . Ethical and regulatory challenges of AI technologies in healthcare: a narrative review. Heliyon. (2024) 10(4):e26297. 10.1016/j.heliyon.2024.e26297

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34.

  Miró Catalina Q Vidal-Alaball J Fuster-Casanovas A Escalé-Besa A Ruiz Comellas A Solé-Casals J . Real-world testing of an artificial intelligence algorithm for the analysis of chest x-rays in primary care settings. Sci Rep. (2024) 4(1):5199. 10.1038/s41598-024-55792-1

  • CrossRef
  • Google Scholar

• 35.

  Bobrovskaya TM Vasilev Y Nikitin N Arzamasov KM . Approaches to building radiology datasets. Med Doctor IT. (2023) 4:14–23. 10.25881/18110193_2023_4_14

  • CrossRef
  • Google Scholar

• 36.

  Vasilev Y Vladzymyrskyy AV Omelyanskaya OV Arzamasov KM Chetverikov SF Rumyantsev DA et al Methodology for testing and monitoring artificial intelligence-based software for medical diagnostics. Digit Diagnostics. (2023) 4(3):252–67. 10.17816/DD321971

  • CrossRef
  • Google Scholar

• 37.

  Vasilev Y Vladzymyrskyy AV Arzamasov KM Shulkin IM Aksenova LE Pestrenin LD et al The first 10,000 mammography exams performed as part of the “description and interpretation of mammography data using artificial intelligence” service. Manager Zdravookhranenia. (2023) 8:54–67. 10.21045/1811-0185-2023-8-54-67

  • CrossRef
  • Google Scholar

• 38.

  Vasilev Y Arzamasov KM Kolsanov AV Vladzymyrskyy AV Omelyanskaya OV Pestrenin LD et al Experience of application artificial intelligence software on 800 thousand fluorographic studies. Medi Doctor Info Technol. (2023) 4:54–65. 10.25881/18110193_2023_4_54

  • CrossRef
  • Google Scholar

• 39.

  Vasilev Y Vladzimirsky AV Arzamasov KM Andreichenko AE Gombolevskiy VA Kulberg NS et al Computer Vision in Radiology: Stage one of the Moscow Experiment. 2nd ed.Moscow: Publishing solutions (2023). p. 376. (In Russ).

  • Google Scholar

• 40.

  Vasilev Y Vladzimirsky AV . Artificial intelligence in radiology: per Aspera ad astra. Moscow: Publishing Solutions (2025). p. 491. (In Russ).

  • Google Scholar

• 41.

  Kufel J Bielówka M Rojek M Mitręga A Lewandowski P Cebula M et al Multi-label classification of chest x-ray abnormalities using transfer learning techniques. J Pers Med. (2023) 13:1426. 10.3390/jpm13101426

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 42.

  DuPont M Castro R Kik SV Palmer M Seddon JA Jaganath D . Computer-aided reading of chest radiographs for paediatric tuberculosis: current status and future directions. Lancet Digit Health. (2025) 7:100884. 10.1016/j.landig.2025.100884

  • CrossRef
  • Google Scholar

• 43.

  Linsen S Kamoun A Gunda A Mwenifumbo T Chavula C Nchimunya L et al A comparison of CXR-CAD software to radiologists in identifying COVID-19 in individuals evaluated for Sars CoV-2 infection in Malawi and Zambia. PLoS Digit Health. (2025) 4:e0000535. 10.1371/journal.pdig.0000535

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 44.

  Kufel J Bargieł-Łączek K Koźlik M Czogalik Ł Dudek P Magiera M et al Chest x-ray foreign objects detection using artificial intelligence. J Clin Med. (2023) 12:5841. 10.3390/jcm12185841

  • Pubmed Abstract
  • CrossRef
  • Google Scholar