Intelligent Diagnosis of Peripheral Pulmonary Lesions Based on Endobronchial Ultrasound Images Using Multi-modal Features.

Abstract

Lung cancer is a leading cause of cancer-related morbidity and mortality worldwide. For peripheral pulmonary lesions (PPLs) that cannot be definitively diagnosed by computed tomography (CT) or positron emission tomography-computed tomography (PET-CT) imaging, endobronchial ultrasound-guided transbronchial lung biopsy (EBUS-TBLB) for pathological examinations still carries a significant possibility of indefinite diagnosis. Endobronchial ultrasound (EBUS) serves as not only a valuable and widely used guidance technique but also a superior diagnostic technique for PPLs compared with CT or PET-CT, by enabling direct visualization of the structures within the lesion and the echo distributions outside the lesion. Aimed at classifying PPLs as benign or malignant, this study enrolled 313 patients for the final analysis. Multi-modal features including clinical, radiomic, and deep features were integrated and fed into an artificial neural network for feature selection, fusion, and prediction based on the long-axis diameter of PPLs measured on CT images with a threshold of 3 cm. To reduce manual effort and maximize the diagnostic performance of multi-modal features, we employed the particle swarm optimization algorithm to automatically optimize the combination of selected features and model structure. Our method achieves an average area under the receiver operating characteristic curve of 0.981 (95% CI: 0.953-1.000), accuracy of 0.964, sensitivity of 0.963, specificity of 0.969, and F1-score of 0.976 on the independent test dataset after four-fold cross-validation. The outstanding diagnostic performance for PPLs based on EBUS images, with all the evaluation metrics exceeding 0.95, can provide a solid foundation for clinical decision-making, biopsy site selection, and follow-up observations.