Lung Nodule Malignancy Classification On 3D CT Images Using a Cosine Similarity-Enhanced Graph Attention Network

Lung cancer remains a major contributor to global cancer-related deaths, largely due to diagnoses occurring at advanced stages. Accurately distinguishing between benign and malignant pulmonary nodules plays a pivotal role in facilitating early detection and timely intervention. To address the limitations of traditional methods that consider only individual CT slices and suffer from high false-positive rates, this work proposes an end-to-end patient-level classification framework for pulmonary nodule malignancy based on a Cosine Similarity-Enhanced Graph Attention Network (CSEGAT). The proposed method utilizes a pre-trained Vision Transformer (ViT) to derive features from individual slices, which are subsequently employed as node attributes within the constructed graph. Edges are constructed based on both spatial adjacency and cosine similarity, enabling the graph neural network to capture spatial and contextual relationships between slices. Each patient is represented as a graph, and the CSEGAT model is employed for patient-level benign/malignant classification. By introducing a similarity-enhanced attention mechanism, CSEGAT dynamically adjusts edge weights, thereby improving information propagation and classification performance. This method enables end-to-end learning from slice-level feature extraction to patient-level diagnosis, eliminating the need for complex feature engineering while enhancing diagnostic accuracy and reducing the false-positive rate. Experimental findings indicate that the proposed method attains a peak testing accuracy of 90.85%, accompanied by an average sensitivity of 88.84% and an average specificity of 90.65%. Experimental results on the LIDC-IDRI dataset demonstrate that CSEGAT outperforms state-of-the-art methods in terms of accuracy, sensitivity, specificity, and AUC, providing a robust solution for pulmonary nodule diagnosis.