AI Chest X-Ray Prioritization Fails to Speed Lung Cancer Diagnosis

A large, multi-site randomized study has revealed that using artificial intelligence (AI) to prioritize chest X-rays (CXRs) did not improve the speed of the lung cancer diagnostic pathway. This finding suggests that this specific application of AI – prioritization – adds complexity and cost without delivering a tangible benefit in accelerating diagnosis. The study also quantified the number of false positives and false negatives, providing insights into the potential impact of AI on cancer detection rates.

The Current Landscape of Lung Cancer Diagnosis

Chest X-rays are the most frequently performed imaging investigation, with over 7 million performed annually in England, including approximately 2.2 million referred from primary care alone ¹. Following an abnormal CXR or identification of a high-risk patient, the National Lung Cancer Screening Programme (NOLCP) mandates rapid progression to a computed tomography (CT) scan. However, current median reporting times for CXRs referred from primary care (2 days) and waiting times for CT scans (15 days) often do not meet the NOLCP standards of 72 hours from abnormal CXR to CT ¹. Limited CT scanner availability and shortages of radiologists and radiographers contribute to these delays ².

The LungIMPACT Study: Assessing AI Prioritization

The LungIMPACT study was designed to evaluate whether AI prioritization of CXRs could optimize the diagnostic pathway. The study involved 93,326 CXRs, randomized into two groups: one with AI prioritization ‘on’ and one with it ‘off’. The primary outcomes were time to CT scan and time to lung cancer diagnosis. Secondary outcomes included the number of urgent suspected lung cancer referrals, incidence and stage of lung cancer, and concordance between AI and radiology reports.

Key Findings

• No Impact on Time to CT or Diagnosis: The study conclusively demonstrated that AI prioritization had no significant impact on either the time to CT scan or the time to lung cancer diagnosis. Median times to CT were 53 days in both groups (P = 0.31), and median times to diagnosis were 44 and 46 days, respectively (P = 0.84).
• Reduced CXR Reporting Time: A significant reduction in the median time from CXR acquisition to report was observed, decreasing from 47 hours to 34.1 hours. However, this reduction did not translate into faster overall diagnosis times.
• Discordance Between AI and Radiologists: Discordance between AI findings and radiology reports occurred in 28,261 CXRs (30%). AI false positive findings comprised 11.6% of the total across all categories.
• Potential for Missed Cancers: Analysis of cases where AI identified a potential issue that was dismissed by the radiologist revealed a potential association with longer times to diagnosis.

Limitations and Considerations

The study acknowledged several limitations:

• Single AI Product: The study used a single AI product, and results may vary with different algorithms.
• Focus on Prioritization Only: The study evaluated only the impact of AI prioritization, not the overall use of AI in the clinical workflow.
• Difficulty Measuring Time to CT: Many patients undergo CT scans for reasons unrelated to the initial CXR, potentially diluting the primary outcome.

Implications for Clinical Practice

The findings suggest that, as tested in this study, AI prioritization of CXRs does not currently require potentially time-consuming and costly efforts to include in clinical services. The study highlights the importance of considering the entire clinical pathway and the potential for AI to impact workflow without necessarily improving patient outcomes. Further research is needed to explore the potential benefits of AI in combination with pathway changes, such as immediate radiologist review of AI-flagged abnormalities.

Future Directions

Ongoing research, including a prospective study in Glasgow, Scotland ³, is investigating the impact of AI prioritization on time to CT. Additional studies are needed to evaluate the accuracy of AI in detecting lung nodules and pneumonia, and to assess the potential for AI to improve radiologist performance as a second reader ⁴. AI algorithms have demonstrated high diagnostic accuracy for pneumonia on CXR and can markedly increase the detection of occult lung nodules when used as a second reader.

References:

1. ¹ Diagnostic Imaging Dataset Annual Statistical Release 2023/24 (NHS England, 2024)
2. ² Life Sciences Competitiveness Indicators 2024: Summary (Department of Science, Innovation and Technology & Department of Health and Social Care, 2024)
3. ³ Duncan, S. F. Et al. Radiograph accelerated detection and identification of cancer in the lung (RADICAL): a mixed methods study to assess the clinical effectiveness and acceptability of Qure.ai artificial intelligence software to prioritise chest X-ray (CXR) interpretation. BMJ Open 14, e081062 (2024).
4. ⁴ Diagnostic accuracy of AI in chest radiography for pneumonia and lung cancer: A meta-analysis