MSA-Net: multiple self-attention mechanism for 3D lung nodule classification in CT images.

Abstract

Purpose: Lung cancer is a life-threatening disease that poses a significant risk to human health. Accurate differentiation between benign and malignant lung nodules, based on computed tomography (CT), is crucial to assess lung health. Developing an automated computer-aided diagnostic method for this differentiation is essential. We introduced a streamlined 3D model structure to solve the problems of 2D models cannot extract spatial information effectively and 3D models have high complexity and large occupation of computing resources.

Methods: We proposed an MSA (multiple self-attention-based) model to address the limitations of 2D models in extracting spatial information effectively and the high complexity associated with 3D models. Our approach introduced the 3D RTConvBlock, which employed multiple self-attention mechanisms for the extraction of spatial features. This enabled the extraction of specific spatial feature information by combining local features, global information, and dependencies between features.

Results: The MSA model demonstrates exceptional performance with an accuracy of 0.953, a sensitivity of 0.963, and an AUC (area under curve) of 0.993 in the LUNA16 dataset, which is higher than state-of-the-art methods. Compared with existing 2D models, we extract spatial information features better, resulting in higher accuracy.

Conclusion: These results have significant implications for enhancing the accuracy and reliability of lung nodule classification, providing robust auxiliary support for physicians diagnosing lung diseases.