Enhancing Pulmonary Tumor Diagnosis Based on Endobronchial Ultrasound Images Through Cross-Domain Features Via Supervised Contrastive Learning

Lung cancer stands as the deadliest malignancy globally. Compared to computed tomography (CT), endobronchial ultrasound (EBUS) exhibits superior capabilities in detecting pulmonary tumors. Currently, EBUS image analysis involves qualitative analysis, radiomic analysis, and neural network analysis. Qualitative analysis relies on subjective physician judgment, while radiomic analysis lacks in revealing intricate nonlinear relationships. Previous studies based on neural network construct encoders using transfer learning and pre-trained models from natural image datasets, overlooking the distinctive characteristics specific to EBUS images. In this study, considering the acquisition process and structural characteristics of EBUS images, we propose a supervised contrastive learning strategy to extract multi-domain features from the Cartesian coordinate domain, polar coordinate domain and frequency domain. Bilinear pooling is adopted to construct cross-domain features for more differential diagnosis of pulmonary tumors. Ablation experiments showcase individual contributions of each module in our method. Compared to other methods, our method achieves state-of-the-art performance on our EBUS image dataset.