Deep Learning-Based 3D Graph Attention Networks for Volumetric Lung Cancer Detection in CT Scans

One of the principal causes of mortality in men and women around the world is lung cancer, which causes five million deaths per year. Cancer is a deadly disease usually caused by a range of pathological changes and genetic disorders. Early diagnosis and accuracy remain challenges for practitioners, even though automated approaches are exploited. To tackle these challenges, a novel 3D Multi-Scale Residual Graph-Attention Network is introduced for accurate detection of lung cancer through volumetric computed tomography scans. In this model, three volumetric computed tomography datasets are utilized, and they are preprocessed by the steps of resampling, intensity normalization, volume cropping and padding, and data augmentation to ensure consistency assurance and model generalization improvement. For hierarchical volumetric feature extraction from the preprocessed computed tomography images, the 3D Multi-Scale Residual Fusion Transformer is employed, which captures long-range contextual dependencies and local textures. The 3D Graph-Attention Residual Detection Network employs graph attention to model the relationships among contextual and spatial information across voxels, enabling precise detection and localization of malignant nodules. Experimental results demonstrate the effectiveness of the proposed 3D Multi-Scale Residual Graph-Attention Network, achieving a 98.83% accuracy, a 98.71% recall, and a 98.63% F1-score, which characterizes the detection of lung cancer in volumetric computed tomography scans.