Self-Supervised Graph Vision Transformer Network For Early Pulmonary Nodule Detection And Malignancy Classification From Ct Imaging

Abstract

Early detection of pulmonary nodules is very important in enhancing survival rates among lung cancer patients. Nevertheless, the small nodules are difficult to identify and classify in computed tomography (CT) images because of their insignificant appearance, change in size and shape, and the limited amount of large annotated medical data. To mitigate these shortcomings, this paper suggests a Self-Supervised Graph Vision Transformer Network (SS-GVTNet) in detecting and classifying pulmonary nodules and malignancies through CT scanning. The suggested framework combines self-supervised representation learning with the graph-based spatial modeling and transformer-based global feature extraction. First, self-supervised pretraining module trains powerful feature representations using massive amounts of unlabeled CT scans, which removes reliance on manual labels. A Self-Supervised Graph Vision Transformer (SGVT) backbone is then used to capture cross-slice contextual dependencies of CTs, and successful modeling of complex anatomy patterns is possible. To further improve the spatial cognition, a Enhanced Granular Neural Network (EGNN) module is built to form connectivity relations among possible nodule areas to enable the neural net to learn structural associations and contextual interactions in lung tissues. The hybrid architecture enhances the local feature removal and the global contextual reasoning. Empirical testing of publicly available lung CT data proves that the suggested model has higher detection accuracy, sensitivity, and specificity than the existing convolutional neural networks and mixed deep learning methods. Moreover, the framework offers credible classification of malignancy, which help clinicians in early diagnosis and treatment planning. The proposed solution is a scalable and efficient computer-aided lung cancer screening system that involves the integration of self-supervised learning, the graph-based modeling of features, and transformer architecture.