HViT4Lung: Hybrid Vision Transformers Augmented by Transfer Learning to Enhance Lung Cancer Diagnosis
Reza Roofegari Nejad, Sahar Hooshmand
Abstract
Lung cancer is the leading cause of mortality among other forms of cancer worldwide. Early and accurate recognition of lung nodules is crucial for the better life quality of patients. Although the chest Computed Tomography (CT) scan is the principal imaging procedure to evaluate and recognize lung cancer, the radiologists evaluation based on CT images is subjective and afflicted from a low accuracy compared to the post-surgery pathological tests. Computer Aided Diagnosis (CAD) has been proven to be beneficial in this context by increasing accuracy and minimizing expert involvement. Nevertheless, due to various factors including size and location inconsistency of nodules, the errorless detection of cancerous cases is still a challenge for CAD systems. Motivated by this fact, this paper presents a novel and effective method, called HViT4Lung (Hybrid Vision Transformers for Lung cancer detection), to enhance lung cancer diagnosis. The proposed deep learning-based hybrid framework combines Transformers and Convolution Neural Networks, augmented by transfer learning that extracts features from the chest CT images to detect lung nodules and predict their malignancy. The proposed pipeline of HViT4Lung is implemented with various blocks of deep learning and tested on a sample lung CT images dataset. The results of the proposed model are very promising compared to the other approaches in this field, achieving 99.20% training accuracy, 99.09% validation accuracy, and 99.09% testing accuracy for the classification of lung CT scans of 1190 images into three different classes of normal, benign, and malignant.