Artificial intelligence in chest X-Ray reporting

Artificial intelligence (AI) is becoming an increasingly prominent feature of modern healthcare, with one of its most visible applications being the integration of AI software into chest X-ray reporting pathways. AI-powered chest X-ray (CXR) analysis has emerged as a promising tool in the early detection of lung cancer, which remains the leading cause of cancer-related mortality worldwide. In the United Kingdom, lung cancer survival rates continue to lag behind those of comparable countries, with delayed diagnosis considered a significant factor. [1] Against this backdrop, AI is increasingly being promoted as a practical tool to improve reporting efficiency and support earlier cancer detection.

Recognising this potential, the UK Government has committed more than £20 million through its AI Diagnostic Fund to support the rollout of AI-powered CXR technology across National Health Service (NHS) Trusts by 2029. With more than seven million chest X-rays performed annually across the NHS, the initiative is intended to address increasing demand on radiology services and reporting backlogs.

One of the most widely adopted systems is ‘Harrison.ai's Annalise CXR’ platform,[2] which can identify up to 124 findings on a chest radiograph. The technology is now used in more than 130 hospitals and 40 NHS trusts across the UK. The Annalise AI model is also utilised for non-contrast CT brain scans to detect intracranial pathology and is available to more than half of Australian radiologists reflecting its growth and expanding international adoption.[3]

Emerging data suggests that AI-assisted reporting can substantially reduce reporting turnaround times, analysing CXR’s and flagging abnormal features that may warrant closer review such as pulmonary nodules or lung masses. Acting as a "second reader", AI may assist radiologists in identifying subtle findings that could otherwise be overlooked in high-volume reporting environments like trauma centres. These platforms also prioritise potential ‘abnormal’ studies, enabling urgent examinations to be reviewed sooner.

Despite considerable positives, evidence highlights that this use of AI is not without functional limitation. The LungIMPACT randomised controlled trial found that while AI-based prioritisation reduced median reporting times from approximately 47 hours to 34 hours, this did not translate into earlier lung cancer diagnosis or treatment.[4] “Bottlenecks” elsewhere in the diagnostic pathway, including patient notification, access to CT imaging, specialist review and multidisciplinary discussion, continued to delay progress in the cancer diagnosis pathway. Furthermore, concerns exist regarding the variability in performance between commercially available AI systems and their potential for automation bias. Studies have demonstrated significant differences in diagnostic accuracy between AI products.[5] While some CXR AI programs improved cancer detection, others performed no better than conventional reporting. Excessive reliance on AI can create a false sense of confidence, which is particularly risky when rare or complex presentations are involved. Further, given the limited legal frameworks governing AI decision-making tools, they must be integrated cautiously with appropriate clinical validation, ethical governance, and ongoing human oversight.

Ultimately, AI should be viewed as a clinical decision-support tool rather than an autonomous diagnostic solution. AI may assist, but cannot replace, clinical judgment and expertise and responsibility rests with the clinician. . Its greatest value may lie in supporting clinical expertise and perhaps helping radiologists manage their increasing workload.