The Potential of Transformer-Based Models for Automated Lung Cancer Detection from CT Scans

Abstract

Lung cancer is the most common type of cancer worldwide and the leading cause of cancer-related deaths. Early diagnosis and treatment can significantly increase the survival rate of this disease. Radiological methods used in the diagnosis of lung cancer, especially Computed Tomography (CT) imaging, allow tumors to be detected more precisely. However, manual analysis of these images is time consuming and error prone due to human factors. In this study, we compared the potential of three different transformer-based state-of-the-art models (ViT, DeiT and Swin Transformer) for automatic lung cancer detection. We collected 690 CT images including small cell lung cancer (SCLC), non-small cell lung cancer (NSCLC) and normal findings from a local hospital. Each image was carefully reviewed and labeled by our expert radiologist, and these labeled images were used to train the models. The ViT, DeiT and Swin Transformer models achieved accuracy rates of 91.3%, 84.1% and 80.4% respectively on the test samples. This study shows that the use of transformer-based models for lung cancer classification is promising in overcoming the difficulties in manual analysis.