Comprehensive Summary
This retrospective study assessed whether pathomics-based machine learning (ML) could improve the diagnostic yield of LungPro navigational bronchoscopy for peripheral pulmonary lesions (PPLs), which often suffer from high false-negative rates. A total of 144 patients were included, with 94 diagnosed as malignant and 50 as benign. Whole-slide biopsy images were processed with convolutional neural networks and multiple instance learning to extract pathomic features. Logistic regression identified age, lesion boundary, and mean CT attenuation as independent malignancy risk factors. Among tested models, ExtraTrees achieved AUCs of 0.792 (95% CI 0.680–0.903) in training and 0.777 (95% CI 0.531–1.000) in testing. When clinical variables were combined with pathomics, diagnostic accuracy improved substantially, reaching AUC 0.909 (95% CI 0.812–1.000) in training and 0.848 (95% CI 0.695–1.000) in testing. In the LungPro-negative cohort (n=50), the integrated model correctly identified 71.4% (20/28) of malignant cases and 68.2% (15/22) of benign cases. Grad-CAM heatmaps localized nuclear atypia and conspicuous nucleoli as key malignant features, enhancing interpretability.
Outcomes and Implications
This work suggests that combining clinical and pathomic data with ML could reduce the limitations of LungPro bronchoscopy by providing more accurate malignancy risk stratification in non-diagnostic cases. At the bedside, this may help clinicians decide when additional biopsy, surgery, or close follow-up is warranted, potentially reducing delays in lung cancer diagnosis. While external validation is still needed, the study highlights how AI-driven pathology can complement interventional pulmonology to improve early lung cancer detection and patient management.