SSTL-FM: Self-supervised transfer learning-based fusion model for the classification of benign-malignant lung nodules

Abstract

Studies on the classification of lung nodules have been extensively conducted because of their significant role in diagnosing lung cancer. However, it is challenging to obtain several labeled nodules in the real world owing to the costly labeling. The limitations posed by a scarcity of labeled nodules in supervised learning hinder the development of more effective network models to improve classification performance. To address this challenge, we propose a novel self-supervised transfer learning strategy (SSTL) and develop a fusion model (FM) to extract the image features and clinical attributes of lung nodules by studying nodule-related data. The reconstruction task and transfer learning in the SSTL can learn robust knowledge by using non-nodules, thereby enhancing the subsequent classification task. FM effectively combines image features and clinical attributes to provide a more comprehensive representation of the nodule information, while the feature refinement module (FRM) further enhancing the fusion of these features. Furthermore, we explored the impact of varying amounts of non-nodule data on the classification performance and the ability of clinical attributes to distinguish between benign and malignant lung nodules. Extensive ablation and comparative experiments have demonstrated that the proposed method significantly improves performance by utilizing nodule-related data.