文献类型：期刊文献

英文题名：VSNet: classification of pulmonary nodules in 3D using vision transformer and sequence spatial attention mechanism

作者：Tang, Dongfang[1]; Xiao, Ting[2]; Yang, Fan[2]; Zhang, Conghao[2]; Wang, Zhe[2]; Gao, Wen[1]

机构：[1] Department of Thoracic Surgery, Huadong Hospital Affiliated to Fudan University, Shanghai, 200040, China; [2] Department of Computer Science and Engineering, East China University of Science and Technology, Shanghai, 200237, China

年份：2025

卷号：84

期号：14

起止页码：13885

外文期刊名：Multimedia Tools and Applications

收录：EI(收录号：20242416253484)

语种：英文

外文关键词：Classification (of information) - Computer aided diagnosis - Computerized tomography - Positron emission tomography

摘要：Accurate classification of benign and malignant nodules in Computed Tomography (CT) scans is crucial for the early detection of lung cancer and Computer-Aided Diagnosis (CAD) systems. Despite significant advancements, challenges such as the interpretability of the reasoning process and the lack of fine-grained representations persist. To address these challenges, we propose a 3D VSNet. Our approach incorporates a Sequence Spatial Attention Module (SSAM) that automatically locates the sequence of the encoder’s output and the receptive field region required, achieving the acquisition of key features of pulmonary nodules. We leverage the characteristics of shallow and deep features based on 3D Vision Transformer (ViT) and Convolutional Neural Networks (CNNs) to obtain fine-grained representations and improve nodule classification. Additionally, we implement a new training strategy using Supervised Contrastive (SC) loss and Proxy-Anchor (PA) loss to optimize the embedding feature of similar samples and the direct cosine distance, and enhance the convergence speed of tuple sampling. Our experiments on the LIDC-IDRI dataset demonstrate the effectiveness of our proposed technique, achieving an accuracy of 90.28% and a precision of 91.63%. Furthermore, we conduct ablation experiments to analyze the contribution and influence of each component of our method. ? The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.