Multiside graph neural network-based attention for local co-occurrence features fusion in lung nodule classification

Abstract

Early diagnosis of lung cancer is critical as it can save people’s lives. Long-range dependencies within volumetric medical images are essential attributes for accurate lung nodule classification. Many deep learning-based methods are used for lung nodule classification; however, the construction of the lung nodule is not axial and can be any shape. Thus, the nodules are interrelated through their adjacent slices axially and diagonally, locally and globally, making their capture through local convolutional operations challenging. In this article, we benefited from local co-occurrences of texture features and graph neural networks (GNNs) to effectively capture important patterns considering the long-range dependencies among adjusted slices. The proposed framework comprises a multi-side graph construction layer (MSGCL) that computes informative texture features and captures spatial relationships from cross-sectional and longitude orientations of the nodule, creating two sets of nodes. Further, the graph-based fusion of long-range dependency layers (GFLL) is used to deeply fuse and generate attentive edges among fused nodes. The LIDC-IDRI dataset, used for training and testing the proposed Multi-side Graph Neural Network-based Attention for Local Co-occurrence Features Fusion (MS-GNN-ALCFF), achieves state-of-the-art. The LUNGx dataset used as an unseen dataset shows that our model is generalisable compared to baselines. The results of the proposed method when the LIDC-IDRI dataset was used for train and testing were 87.17 ± 0.84 %, 91.01 ± 1.14 %, 88.6 ± 1.42 %, 89.7 ± 0.87 %,95 ± 1.22 % and 76.7 ± 1.4 % in terms of Accuracy, Precision, Recall, F1-score, AUC and MCC respectively. When the LUNGx dataset is used for testing as an unseen dataset, the results were 69.86 ± 2.4 %, 75 ± 2.7 %, 71.4 ± 3.1 %, 73.17 ± 3.7 %, 70.2 ± 1.36 % and 71 ± 3.8 % for Accuracy, Precision, Recall, F1-score, AUC and MCC respectively. These findings represent the significance of the ability of GNNs to construct a multi-set of nodes through the proposed MSGCL layer and fuse deeply through the proposed GFLL layer.