基于改进的深度残差网络(ResNet),提出更加适合肺部组织的计算断层扫描(CT)图像模式分类模型。为克服医学图像分析中可用数据集稀少的困难,采用迁移学习方法来减小神经网络模型对数据量大的需求,以减小过拟合。迁移学习的策略是将肺内大量可用的无标签区域作为预训练的数据,使用深度互信息最大化和先验分布匹配的方法进行无监督表征学习。通过对比实验发现,改进的深度ResNet可以得到更高的分类精度,迁移学习算法可以有效地利用肺内无标签区域的数据,从而提升网络模型的分类表现。

We propose a deep model for pattern classification of computed tomography (CT) images of lung tissues based on the improved deep resiual netwk (ResNet). To address the problem of lack of availability training data, we adopt a transfer learning method to reduce the requiement of a neural network model for large data, thereby decreasing overfitting. The transfer learning strategy uses massively available unlabeled lung CT data as the pre-training data. We perform unsupervised representation learning by maximizing the deep mutual information and matching the prior distribution. The results of contrast experiments show that the improved ResNet achieves improved classification accuracy, the effectiveness of utilizing the unlabeled lung CT data for transfer learning and the classification performance of the network model is improved.