Multi-Modal Features for Intelligent Differential Diagnosis of Solitary Pulmonary Tumors by Using Endobronchial Ultrasonography Images

Lung cancer is the deadliest cancer worldwide. Endobronchial ultrasonography (EBUS) images contain effective features that are helpful for the diagnosis of lung cancer. This study develops a multiparametric ultrasound method by extracting multi-modal features from EBUS images based on ultrasound tissue characterization methods and convolutional neural network (CNN) for differentiating diagnosis of benign and malignant solitary pulmonary tumors. The proposed intelligent diagnosis method with three modalities of features (that is, clinical features, Radiomic features and deep features) and feature selection achieve good performance with AUC of 85.14%, accuracy of 80.55%, sensitivity of 80.14%, specificity of 81.69%, and the F1-score of 80.88 %. Ablation experiments suggest that the multi-modal features are better and more robust than single- or dual- modal features. In conclusion, the proposed method could be a competitive tool to achieve intelligent diagnosis of benign and malignant solitary pulmonary tumors.