MSTD: A Multi-Scale Transformer-Based Method to Diagnose Benign and Malignant Lung Nodules

The identification of benign and malignant lung nodules is crucial for timely treatment to reduce the risk of the progression and metastasis of diseases. However, the varied sizes, diverse morphologies, non-fixed positions, and dynamic growth of lung nodules in computed tomography (CT) images make their accurate identification challenging. To address these issues, we propose a multi-scale transformer-based diagnosis (MSTD) method for benign and malignant lung nodules. To handle significant variations in the shapes and sizes of the lung nodules, we first design a multi-scale module based on parallel branches to extract multi-scale features. To make full use of these features, we then introduce a multi-scale transformer fusion (MSTF) module to integrate the information obtained at different scales. Unlike conventional vision transformers, our MSTF can simultaneously extract attention-based features from the spatial dimensions at different scales to enhance the accuracy of classification of lung nodules. We conducted extensive ablation experiments on multi-scale structures and transformer-based methods of fusion to explore the impact of features obtained at different scales on the accuracy of classification of lung nodules. The results of verification on the LUNA16 dataset showed that the average F1Score, Specificity, and Sensitivity of the proposed MSTD exceeded 90% (94.5%, 96.5%, and 91.1%, respectively), where this shows that it can accurately identify both benign and malignant lung nodules. Its average performance was superior to the state-of-the-art method by about 1%, 3.4%, and 3.6% in terms of the area under the curve (AUC), Accuracy, and F1Score, respectively.