Abstract Rationale Expeditious evaluation of peripheral pulmonary nodules (PPN) is central to timely lung cancer treatment. Robotic-assisted bronchoscopy (RaB) has improved peripheral biopsy reach, but current clinical and imaging tools inadequately predict diagnostic success. This study aims to incorporate radiomics, machine learning (ML), artificial intelligence, and risk calculators to develop a predictive model of the likelihood of diagnosis of PPNs, facilitate procedural planning, and streamline the pathway from nodule detection to treatment. Methods We conducted a retrospective observational cohort study of 519 patients who underwent RaB for PPN evaluation at three Chicagoland hospitals. Pre-procedural CT scans, procedural characteristics, and histopathologic outcomes were collected in a secure, de-identified database. Experienced bronchoscopists across the USA independently reviewed the de-identified scans, predicting whether the procedure would be diagnostic and whether the diagnosis would be malignant. Cases were also scored using the BIMC clinical model to predict malignancy risk. We implemented a volumetric CT classification model using DenseNet121 architecture. The model processes 3D lung CT volumes at 256 × 256×160 resolution for binary classification of diagnostic outcomes. Training employed focal loss to address class imbalance, optimizing over 100 epochs with batch size of 4. Performance of clinicians, the BIMC model, and the ML model were evaluated using AUROC. Results Twenty-two bronchoscopy experts reviewed 84 de-identified CT scans with pathologic diagnoses from a prior RaB. Each scan reviewed for diagnostic outcome received 12.39 average reads, with 55/84 (65.5%) proving diagnostic on biopsy. The 55 scans corresponding to diagnostic nodules were reviewed to predict malignancy, receiving 13.8 average reads, with 41/55 (74.5%) resulting in a malignant diagnosis. Across all reviewers, the AUROC was 0.563 for predicting a diagnostic procedure and 0.675 for predicting malignancy. The BIMC clinical model achieved an AUROC of 0.833 for malignancy prediction. The DenseNet121 ML model predicted a diagnostic procedure with an AUROC of 0.778, 69.5% accuracy, and an F1 score of 0.7748 on independent test set evaluation, demonstrating potential utility in supporting bronchoscopy planning. Conclusions Experienced bronchoscopists, without clinical context, showed limited ability to predict 1) whether a RaB procedure would be diagnostic, and 2) whether a lesion was malignant based on CT alone. The application of radiomics and ML to PPN evaluation offers a promising approach to enhance diagnostic precision, decrease nondiagnostic procedures, and facilitate earlier lung cancer detection and treatment. Future steps include the addition of clinical information and prospective internal validation within the Northwestern cohort, followed by external validation. This abstract is funded by: None
Bhargava et al. (Fri,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: