Performance validation of a closed loop fully automated AI model for lung nodule stratification in screening cases

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Human-based lung cancer screening (LCS) faces challenges including high false-positive rates, inter-reader variability, and diagnostic errors.
• Existing artificial intelligence (AI) models often require manual processing, increasing time and cost.
• Bronchosolve is a novel, fully-automated software designed to enhance LCS consistency, accuracy, and throughput.

Purpose of the Study:

• To develop and validate Bronchosolve, a closed-loop, automated software for lung nodule risk classification.
• To assess the performance of Bronchosolve compared to existing methods like Lung-RADS and the Brock model.
• To improve the efficiency and accuracy of lung cancer screening workflows.

Main Methods:

• Bronchosolve integrates pre-processing, analysis, and report generation using a deep-learning convolutional neural network (CNN).
• The system processes full chest CT scans automatically, including optimal series selection, normalization, nodule detection, classification, and report generation.
• The model was trained on 2358 multi-center cases and validated on a U.S.-based cohort of 184 patients.

Main Results:

• Bronchosolve achieved an Area Under the Curve (AUC) of 0.898, significantly outperforming Lung-RADS (pAUC 0.669) and the Brock model (AUC 0.783).
• The software demonstrated high sensitivity (83.6%) and specificity (86.3%) in classifying lung nodules.
• Performance remained consistent across various scanner types and slice thicknesses, with 100% automated case completion.

Conclusions:

• Bronchosolve provides accurate, fully-automated risk classification for lung nodules.
• The software has the potential to significantly enhance non-invasive diagnostic workflows in lung cancer screening.
• Automated AI solutions like Bronchosolve can address limitations of traditional LCS methods.