106P - AI negative predictive performance exceeds that of radiologists in volumetric-based risk stratification of lung nodules detected at baseline in a lung cancer screening population

Background

Lung cancer is the leading cause of cancer-related mortality due to the late stage at which it is diagnosed. When detected at stage 4, 5-year survival is <5%. Whereas lung cancer detected at stage 1, has a 5-year survival >75%. Early detection can be achieved through LDCT lung cancer screening in a high-risk population. Despite being proven effective, widespread implementation remains a challenge for multiple reasons. One of which, the ever-increasing workloads which radiologists face. AI could be crucial in overcoming this challenge.

Methods

Performance of different versions of an AI-prototype (AVIEW LCS; v1.0.34, v1.1.39.14, and v1.1.42.56) was analysed in 283 ultra-LDCT baseline scans. Volumetric nodule measurements from independent reads of the AI-prototypes were compared that of five experienced radiologists, and an independent consensus reference read was performed. Discrepancies between individual reads and consensus were classified as follows; positive misclassifications (nodules classified by the reader/AI as ≥100 mm³, which at the reference consensus read were <100 mm³) and negative misclassifications (nodules classified as a <100 mm³ by the reader/AI, which at consensus read were ≥100 mm³).

Results

Using AVIEW LCS v1.0.34, 1149 nodules were detected of which 878 were classified as solid non-calcified. For v1.1.39.14 and v1.1.42.55, 1502 nodules were detected of which 1019 were solid non-calcified. Overall, v1.0.34 had 61 discrepancies (53 positive and 8 negative misclassifications), and v1.1.39.14 and v1.1.42.56 both had 32 discrepancies (28 positive and 4 negative). The improved AI versions (v1.1.39.14 and v1.1.42.56) had a better negative predictive value than 4 radiologists and equivalent negative predictive value to the 5th radiologist.

Conclusions

AI can surpass the negative predictive performance of experienced radiologists, meaning it could provide a solution to reducing radiologists’ workload associated with lung cancer screening if used as a first read filter.

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

D. Yu: Other, Personal, Full or part-time Employment: Coreline Soft. J. Yi: Other, Personal, Full or part-time Employment: Coreline Soft. All other authors have declared no conflicts of interest.