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

Several limitations hinder the effectiveness of human-based lung cancer screening (LCS): high false-positive rates leading to unnecessary follow-up imaging, procedures, and surgeries; inter-reader variability; inconsistent Lung-RADS adherence; and fatigue-related diagnostic errors. Additionally, most artificial intelligence (AI) models address only one task (nodule detection or risk stratification) and require manual image processing, which is time-consuming and costly. We developed Bronchosolve, a closed-loop, fully-automated software that processes scans without manual input, aiming to improve consistency, accuracy, and throughput in LCS.

The software integrates pre-processing, analysis, and result generation, using a deep-learning convolutional neural network (CNN) for pulmonary nodule triaging. Inputs were full chest CT scans in DICOM format, without clinical or demographic data. Automated steps included: 1) optimal CT series selection, 2) normalization and preprocessing, 3) AI-based detection and classification of suspicious nodules, and 4) report generation. The model was trained on a multi-center high-prevalence set of 2358 cases (malignant and benign nodules). Validation used a U.S.-based, multi-site cohort (n = 184; 8 sites). Positive cases were biopsy-confirmed within 1 year; negatives had biopsy or ≥2-year follow-up.

All cases completed automatically (100 % success). Median age was 62.5 years (IQR 58.5-66.5); 45 % former smokers, 55 % current smokers, and 40 % female. The model achieved an AUC of 0.898 [0.851-0.940], outperforming Lung-RADS (pAUC 0.669) and the Brock model (AUC 0.783). Sensitivity was 83.6 %; specificity was 86.3 %. Performance remained consistent across scanner types and slice thicknesses.

Bronchosolve enables accurate, fully-automated risk classification of lung nodules and may enhance non-invasive diagnostic workflows.