LungNet: A Transformer-Based Deep Learning Model for Early Lung Cancer Detection from CT Images

Abstract

Early detection of lung cancer has a greater opportunity in terms of survival rate; however, the current diagnostic systems are already dealing with some difficulties in achieving high accuracy, particularly in early-stage nodules with fine-grained differences. In this paper, we present LungNet, a new deep learning model that combines Vision Transformers (ViT) with a convolutional neural network (CNN) backbone for early detection of lung cancer in computed tomography (CT) scans. Different from the conventional CNN-based networks, the global attention mechanism of transformers in LungNet helps successfully model long-range correlations and context across anatomical regions of the input images, which is crucial to accurately localize the malignant features even in complex backgrounds. We train and evaluate the proposed model on the public dataset LIDC-IDRI, with performance superior to the state-of-the-art, the accuracy, sensitivity, and F1-score are 94.6%, 96.1%, 0.942, respectively. Attention visualization indicates that LungNet pays attention to meaningful regions, thus making the model interpretable. This is a hybrid architecture that succeeds to incorporate the local detail extraction capabilities of CNNs with the global reasoning of transformers, leading to an effective and scalable intermediate-level solution for computer-aided lung cancer screening. Our findings suggest the applicability of LungNet in the radiologist’s workflow with the help of LungNet for more robust, interpretable, and early lung cancer diagnosis.