Peripheral Pulmonary Lesions Classification Using Endobronchial Ultrasonography Images Based on Bagging Ensemble Learning and Down-Sampling Technique

Abstract

Lung cancer is the second most common cancer in the world, with an average five-year survival rate of 15 percent. Approximately 238,340 people were diagnosed in the US in 2023 based on the estimation of the American Cancer Society, and 127,070 people died from it. Cancer has always been a big problem for scientists. There has never been a good solution. So, the early detection of cancer is particularly important. In recent years, endobronchial ultrasonography (EBUS) images have been used more and more in the diagnosis of lung cancer because of their advantages of good real-time performance, no radiation, and superior performance. This research aims to develop a computer-aided diagnosis (CAD) system to differentiate benign and malignant peripheral pulmonary lesions (PPLs). The efficacy of this framework was evaluated on a dataset comprising 69 cases of lung carcinoma, encompassing 59 malignant instances and 10 benign cases. The final experimental results of accuracy, F1-Score, AUC, PPV, NPV, sensitivity, and specificity were 0.7, 0.63, 0.75, 0.84, 0.68, 0.56, and 0.85, respectively. From the experiment results, the developed CAD system has the potential ability to diagnose PPLs by using the EBUS images based on Deep Learning.

Keywords

• lung cancer
• endobronchial ultrasonography (EBUS)
• peripheral pulmonary lesions (PPLs)
• ensemble learning
• down-sampling
• convolutional neural network (CNN)