基于ViT与ResNet-18的深度学习模型在肺结节良恶性识别中的性能比较

Comparing ViT and ResNet-18 Deep Learning Models for Pulmonary Nodule Benign-Malignant Classification

摘要: 目的：比较基于局部纹理特征与全局上下文信息的深度学习模型在肺结节良恶性鉴别中的性能。方法：纳入2023年6月至2025年7月在石河子大学第一附属医院经手术及病理证实的134例肺结节患者(良性51例，恶性83例)，共收集1172张含肺结节的CT切片，按7:3比例划分为训练集和内部测试集。训练集用于训练关注局部纹理的ResNet-18模型及关注全局信息的Vision Transformer (ViT)模型。通过受试者工作特征曲线下面积(AUC)和决策曲线分析(DCA)评估不同模型的诊断性能。结果：ViT模型在训练集与测试集上均表现更优，其训练集AUC为0.977，准确率为0.978；测试集AUC为0.901，准确率为0.878，均高于ResNet-18模型(训练集AUC 0.959，准确率0.930；测试集AUC 0.880，准确率0.812)。DCA进一步表明ViT模型具有更高的临床净获益。结论：在肺结节良恶性鉴别中，对全局上下文信息的建模能力优于对局部纹理的分析，是提升鉴别性能的关键。

Abstract: Objective: This paper aims to compare the performance of deep learning models focusing on local texture features versus global contextual information in differentiating benign and malignant pulmonary nodules. Methods: A total of 134 patients with pathologically confirmed pulmonary nodules, treated between June 2023 and July 2025 at the First Affiliated Hospital of Shihezi University (51 benign, 83 malignant), were included. A total of 1172 CT slices containing pulmonary nodules were collected and split into a training set and an internal test set at a 7:3 ratio. The training set was used to train a ResNet-18 model focusing on local texture features and a Vision Transformer (ViT) model emphasizing global contextual information. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA). Results: The ViT model outperformed ResNet-18 on both the training and test sets. In the training set, the ViT achieved an AUC of 0.977 and an accuracy of 0.978, while in the test set, it reached an AUC of 0.901 and an accuracy of 0.878, all higher than those of ResNet-18 (training set AUC 0.959, accuracy 0.930; test set AUC 0.880, accuracy 0.812). DCA further indicated that the ViT model provided greater clinical net benefit. Conclusion: In differentiating benign and malignant pulmonary nodules, modeling global contextual information is superior to analyzing local texture features, representing a key factor in improving diagnostic performance.