Classification of Pulmonary Nodules Using Multimodal Feature-Driven Graph Convolutional Networks with Specificity Proficiency

Abstract

Graph neural networks could compare the difference among all samples (nodes in graph) and transmit the interrelationship among them to obtain a global landscape. Compared with radiomics and clinical feature-based machine learning methods, whether a graph convolutional neural network (GCNN) based on radiomics and clinical features improve the performance in distinguishing benign and malignant pulmonary nodules is not well studied. We propose an approach based on multimodal GCNNs that integrates patients’ lung computed tomography images with clinical information to differentiate between benign and malignant pulmonary nodules. Leveraging large-scale and multisource data from multiple hospitals (i.e., 6033/290/524 patients for three hospitals respectively) enhances the diversity of features. Accuracy, sensitivity, specificity, precision, and area under the receiver operating characteristic curve (AUROC) are used to evaluate the performance. We achieved the average accuracy/sensitivity/specificity/AUROC of 0.8612/0.9425/0.6786/0.9025 for the main dataset via the novel GCNN proposed, respectively, maintaining the robustness of the deep learning procedures. Especially for the external testing dataset (hospital 2/hospital 3), the specificity is much higher than comparison methods (0.6250–0.6731 vs. 0.2569–0.2788). The graph neural network-based deep learning method holds the potential to assist clinicians, aiding in treatment planning, patient management, follow-up strategies, resource optimization, and overall healthcare decision-making.