Prevalence of AI findings on Chest X-ray in patients with lung cancer: a cross-sectional cohort study

Background

Chest X-ray Radiography (CXR) is the primary investigation for patients with potential symptoms of lung cancer in the UK. Artificial intelligence (AI) can detect abnormalities on CXR and prioritise cases for reporting. We describe a method to determine which AI findings are associated with lung cancer to inform and validate prioritisation strategies.

Methods

This multicentre study compared the prevalence of AI findings on CXR in a retrospective cohort of patients diagnosed with lung cancer (4408 CXR) to the prevalence of AI findings in a prospective cohort of CXR from the referral population (107,065 CXR). Nineteen AI findings were assessed individually and in combination.

Results

The most common AI findings in patients with lung cancer compared to the referral population were ‘Abnormal’ (92.6% vs 60.9%), Opacity (83.4% vs 45.4%), Consolidation (36.9% vs 12.9%), Atelectasis (33.5% vs 20.9%) and Nodule (32.7% vs 10.9%). The finding most associated with cancer based on the prevalence ratio in the cancer and referral cohorts were ‘Lung nodule malignancy’ (13.3), Cavity (4.0), Tracheal deviation (3.1), Nodule (3.0) and Consolidation (2.9). The percentage of CXR classed as ‘AI Abnormal’ varied by the referral cohort, 63.5% from Accident & Emergency vs 43.0% from General Practice. This suggests significant variation in the complexity of cases across referral pathways.

Conclusion

Individual AI findings had limited sensitivity in detecting lung cancer. Using combinations of AI findings significantly improved cancer detection rates but required prioritising a larger proportion of CXR from the referral population.

Key messages

What is already known on this topic

The National Institute for Clinical Excellence (NICE) early value assessment of “Artificial intelligence-derived software to analyse CXR for suspected lung cancer in primary care referrals” has identified the need for more research on the ability of AI to identify normal and abnormal findings on CXR to enable prioritised reporting of and speed up the cancer pathway.

What this study adds

This study describes a method that sites deploying AI can use to determine the relative prevalence of AI finding on CXR in patients with lung cancer and in the referral population. The results indicate that prioritisation approaches focusing on single abnormalities such as lung nodule detection will miss most lung cancers on CXR.

How this study might affect research, practice or policy

Prioritisation strategies need to optimise the detection of lung cancer and other significant pathology. Using AI to prioritise abnormal CXR requires ethical consideration as moving some patients to the front of the reporting queue can also have a detrimental effect on patients that are not prioritised. This study indicates a range of AI findings need to be prioritised to maximise the detection rate of lung cancer.