Abstract
Lung cancer is one of the leading fatal diseases that causes millions of deaths each year, but early detection and treatment can improve survival. Electronic nose (e-nose) is a recently developed gas sensor that can help us obtain information from the exhaled breath of patients. Its advantages include low cost and no residual radiation risk. Our classifier uses Convolutional Neural Network(CNN) architecture for diagnosing lung cancer based on analyzing e-nose sensory signals. Due to the differences between e-nose devices and background environment, the exhaled breath of same patient can result in different responses when read by different sensors. Therefore, we used transfer learning to enhance the recognition performance of our model for data from new devices, by using a part of pretrained parameters from a previously trained model. To train and evaluate performance, we collected e-nose data from multiple devices and clinical environments. In experiments, our method outperformed other machine learning methods, achieving an accuracy rate of up to 93%.