Artificial neural networks improve LDCT lung cancer screening: a comparative validation study

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Lung cancer risk assessment is crucial for early detection and intervention.
• Standardized structured reports from low-dose computed tomography (LDCT) provide valuable data for risk prediction.
• Existing risk assessment tools may have limitations in certain populations.

Purpose of the Study:

• To develop and validate an artificial neural network (ANN) prediction model for lung cancer risk.
• To compare the performance of the ANN model against the Lung-RADS assessment criteria.
• To evaluate the utility of a data-driven approach for lung cancer risk stratification.

Main Methods:

• A prospective cohort of 836 asymptomatic patients undergoing LDCT was analyzed.
• An ANN prediction model was constructed using a derivation cohort of 602 participants.
• The ANN model's performance was comparatively validated against Lung-RADS using a separate prospective cohort of 234 participants.
• Receiver operating characteristic (ROC) curve analysis was employed to compare model performance, using the area under the curve (AUC) as a metric.

Main Results:

• The ANN model achieved an AUC of 0.873, significantly outperforming Lung-RADS (AUC = 0.764, P=0.01).
• At its optimal cut-off, the ANN demonstrated higher sensitivity (75.0%) and negative predictive value (99.0%) compared to Lung-RADS (sensitivity 12.5%, NPV 96.9%).
• Partially solid nodules and ground-glass nodules sizes were identified as key predictors by the ANN.

Conclusions:

• The ANN model offers superior sensitivity for lung cancer detection in an Asian population compared to Lung-RADS.
• The ANN provides more refined discriminative ability for lung cancer risk stratification, incorporating population-specific demographic characteristics.
• Data-driven ANNs utilizing standardized LDCT reports show promise for enhancing lung cancer prediction beyond conventional rule-based criteria.