Abstract Background Radial endobronchial ultrasound (R-EBUS) guides transbronchial sampling of peripheral pulmonary lesions (PPLs), but B-mode image interpretation is operator-dependent and contributes to nondiagnostic sampling and variable diagnostic yield. Evidence for deep-learning models that analyze R-EBUS images, particularly multimodal approaches integrating routine clinical information, remains limited. Objective To develop and validate image-only and multimodal image-text deep-learning models for malignancy discrimination in PPLs using R-EBUS images, and to test whether integrating routine clinical descriptors improves performance. Methods We performed a single-center retrospective diagnostic study of patients contributing 869 R-EBUS images, with patient-level randomization (7:2:1) into training, validation, and test sets. Four ImageNet-pretrained convolutional backbones (EfficientNet-B0, ConvNeXt-Tiny, DenseNet-201, ResNet-50) were fine-tuned for image-only classification. A multimodal image-text model used Chinese Contrastive Language-Image Pre-training (Chinese-CLIP) bi-encoders to represent paired R-EBUS images and structured Chinese-language routine clinical descriptors; visual and textual embeddings were concatenated and passed to a shallow classifier. Model development used standard augmentation and early stopping. Test performance was summarized across five independent runs. The primary outcome was area under the receiver operating characteristic curve (AUC); sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score were reported with 95% confidence intervals (CIs). Results Across all backbones, clinical fusion improved test-set discrimination. The best multimodal model (EfficientNet-B0 plus clinical) achieved AUC 0.847 (95% CI 0.807-0.887). At the prespecified threshold, recall was 0.870 (95% CI 0.839-0.901), precision 0.851 (95% CI 0.829-0.873), and F1-score 0.860 (95% CI 0.852-0.868). The image-only EfficientNet-B0 achieved AUC 0.805 (95% CI 0.757-0.852). Gains of similar direction and magnitude were seen with ConvNeXt-Tiny, DenseNet-201, and ResNet-50, with stable estimates across five repeated runs. Conclusion Multimodal fusion of routine clinical descriptors with R-EBUS images improves discrimination between malignant and benign PPLs and can support more precise biopsy targeting, reducing nondiagnostic sampling and repeat procedures. Prospective multicenter validation and formal clinical-utility studies are warranted before implementation. This abstract is funded by: None
Qi et al. (Fri,) studied this question.