Classification of peripheral pulmonary lesions in Endobronchial ultrasonography image using a multi-branch framework and voting ensemble

Lung cancer stands as a significant contributor to cancer-related fatalities worldwide. Endobronchial ultrasonography plays a crucial role in the early diagnosis of lung cancer. In this study, our objective is to formulate a deep learning-based Computer-aided Diagnosis (CAD) system for lung cancer, aiming to assist medical professionals in achieving more precise and efficient diagnoses. In this research, acknowledging the pronounced issue of extreme data imbalance, we propose a multi-branch framework. Additionally, to enhance the performance of the CAD system further, we employ a majority voting mechanism, integrating multiple branches to generate the final output. The design of this multi-branch framework aims to better adapt to the distribution differences among various categories, thereby augmenting the model's capability to recognize minority classes. Furthermore, we explored a coordinate system transformation approach, wherein the original Endobronchial Ultrasonography (EBUS) images are converted from polar coordinates to Cartesian coordinates. Such a transformation may contribute to reducing the complexity of image processing, providing deep learning models with clearer and more consistent inputs, thereby augmenting the model's ability to extract features related to lung cancer. The proposed multi-branch CAD diagnostic system, utilizing EBUS images transformed through coordinate system conversion, has demonstrated good performance. This method achieved a level of 0.80 in terms of Area Under the Curve (AUC), with an accuracy of 0.78, F1 score of 0.80, positive predictive value of 0.77, negative predictive value of 0.83, sensitivity of 0.85, and specificity of 0.72. The utilization of a multi-branch framework and ensemble learning proves to be more effective in addressing data imbalance issues. Furthermore, image transformation based on coordinate systems contributes to optimizing the model's understanding of image structures, which can further enhance performance.