Clinical Impact and Generalizability of a Computer-Assisted Diagnostic Tool to Risk-Stratify Lung Nodules With CT

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Lung nodules require accurate classification and follow-up to manage potential malignancy.
• Current risk assessment tools, like Lung-RADS, have limitations in sensitivity and specificity.

Purpose of the Study:

• To evaluate an imaging classifier for improving lung nodule classification and follow-up in radiology practice.
• To assess the performance of a machine learning-based malignancy risk score (mSI) compared to existing clinical models.

Main Methods:

• Developed and trained a machine learning classifier (mSI) using National Lung Screening Trial (NLST) data.
• Validated mSI on external screening and non-screening CT datasets.
• Compared mSI combined with Lung-RADS against Lung-RADS alone and other risk calculators.

Main Results:

• mSI demonstrated comparable accuracy (AUC=0.89) to existing risk models (AUC=0.86-0.88).
• mSI significantly increased sensitivity (25%-117%) and specificity (17%-33%) when combined with Lung-RADS.
• Use of mSI could lead to earlier diagnoses and reduced follow-up, potentially identifying 42% more malignant nodules.

Conclusions:

• Computer-assisted diagnosis software (mSI) improves risk classification for lung nodules detected on CT scans.
• mSI provides independent predictive value beyond existing radiological and clinical variables.
• The tool shows generalizability and potential for straightforward clinical implementation to enhance lung nodule management.