External Test of a Deep Learning Algorithm for Pulmonary Nodule Malignancy Risk Stratification Using European...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Low-dose CT screening effectively reduces lung cancer mortality.
• However, high false-positive rates in screening lead to unnecessary procedures.
• Deep learning (DL) offers potential for improved lung nodule risk stratification.

Purpose of the Study:

• To externally validate a DL algorithm for estimating lung nodule malignancy risk.
• The study utilized pooled data from three major European lung cancer screening trials.
• Performance was compared against the established Pan-Canadian Early Detection of Lung Cancer (PanCan) model.

Main Methods:

• Retrospective analysis of baseline CT scans from the Danish Lung Cancer Screening Trial, Multicentric Italian Lung Detection trial, and Dutch-Belgian Lung Cancer Screening Trial.
• A DL algorithm, trained on US data, was tested on these European cohorts.
• Performance metrics, including AUC, were evaluated across the pooled cohort and specific subsets of indeterminate and size-matched nodules.

Main Results:

• The DL algorithm demonstrated strong performance across the pooled cohort (AUCs 0.98-0.94), comparable to the PanCan model.
• In subset A (indeterminate nodules), DL significantly outperformed PanCan (AUCs 0.95-0.90 vs 0.91-0.86).
• DL achieved a 39.4% relative reduction in false-positive findings for indeterminate nodules at 100% sensitivity, and superior AUC in size-matched nodules (0.79 vs 0.60).

Conclusions:

• The DL algorithm exhibits superior performance in predicting lung nodule malignancy across diverse European screening datasets.
• It significantly reduces false-positive classifications, particularly for indeterminate nodules.
• This validates the DL algorithm's potential for enhancing lung cancer screening accuracy and efficiency.