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

TL;DR

This paper reviews AI's ability to autonomously sort chest x-rays, showing it can reduce radiologists' workload and errors, but implementation is limited by regulatory issues.

Contribution

The study evaluates AI's potential for autonomous CXR triage using a rapid evidence assessment and meta-analysis of recent studies.

Findings

AI can autonomously triage 15.0% to 99.8% of chest x-rays, with a weighted average of 42.3%.

Sensitivity and specificity of AI systems were 97.8% and 94.8%, respectively, though with wide confidence intervals.

Only 55% of studies were classified as having low risk of bias, due to variability in sample selection and evaluation methods.

Abstract

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.

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.

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.

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.

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.

Figures