| 杨俊杰,陈兆学.基于改进ConvNeXt模型在小样本条件下识别CT图像中的肺腺癌[J].中国医学影像技术,2025,41(12):2056~2060 |
| 基于改进ConvNeXt模型在小样本条件下识别CT图像中的肺腺癌 |
| Improved ConvNeXt model for identifying lung adenocarcinoma in CT images under small-sample conditions |
| 投稿时间：2025-07-26 修订日期：2025-11-13 |
| DOI：10.13929/j.issn.1003-3289.2025.12.025 |
| 中文关键词: 肺肿瘤 腺癌 体层摄影术,X线计算机 深度学习 |
| 英文关键词:lung neoplasms adenocarcinoma tomography, X-ray computed deep learning |
| 基金项目: |
|
| 摘要点击次数: 104 |
| 全文下载次数: 20 |
| 中文摘要: |
| 目的 观察基于改进ConvNeXt模型在小样本条件下识别CT图像中的肺腺癌的价值。方法 纳入来源于公开LungCancer4Types肺肿瘤高分辨率CT图像数据集的1 000幅CT图像,将其分为训练集(n=613)、测试集(n=315)与验证集(n=72)。在ConvNeXt模型基础上引入混合注意力Transformer(HAT)模块和SCConv卷积模块构建改进ConvNeXt模型,利用准确率、精确度、召回率、F1分数和受试者工作特征曲线的曲线下面积(AUC)评价模型识别肺腺癌效能,并与GoogLeNet、EfficientNet、DenseNet及AlexNet模型进行比较。以消融实验观察引入SCConv及HAT模块后对模型性能的影响。结果 改进ConvNeXt模型识别测试集肺腺癌的准确率、精确度、召回率、F1分数及AUC分别为95.91%、95.78%、96.93%、0.964及0.975,均高于GoogLeNet、EfficientNet、DenseNet及AlexNet模型。基础ConvNeXt模型识别肺腺癌的准确率、精确度、召回率及F1分数整体较低,引入SCConv模块后模型召回率显著提升,引入HAT模块后准确率显著提高;改进ConvNeXt模型整体性能最优。结论 改进ConvNeXt模型能在小样本条件下有效识别CT图像中的肺腺癌。 |
| 英文摘要: |
| Objective To observe the value of improved ConvNeXt model for identifying lung adenocarcinoma in CT images under small-sample conditions. Methods A total of 1 000 high-resolution CT images were obtained from publicly available LungCancer4Types dataset and divided into training set (n=613), test set (n=315) and validation set (n=72). Based on ConvNeXt model, hybrid attention Transformer (HAT) module and SCConv convolution module were embedded to construct an improved ConvNeXt model. The performance of the obtained model for identifying lung adenocarcinoma in CT images was evaluated using accuracy, precision, recall, F1 score and area under the curve (AUC) of receiver operating characteristic curve, which were all compared with those of GoogLeNet, EfficientNet, DenseNet and AlexNet models. The impact of introducing SCConv and HAT modules on model performance through ablation experiments was observed. Results The accuracy, precision, recall, F1 score and AUC of improved ConvNeXt model for identifying lung adenocarcinoma in test set was 95.91%, 95.78%, 96.93%, 0.964 and 0.975, respectively, all superior to the those of GoogLeNet, EfficientNet, DenseNet and AlexNet models. The baseline ConvNeXt model exhibited relatively lower accuracy, precision, recall and F1 score, the introduction of SCConv module significantly increasd model recall, and the introduction of HAT module significantly improved accuracy. The improved ConvNeXt model had the best overall performance. Conclusion Improved ConvNeXt model could effectively identify lung adenocarcinoma in CT images under small-sample conditions. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
