深度学习模型训练时需要大量的注释样本,但在医学领域注释数据难以获取。针对此问题,提出了一种结合部分注释数据的自监督学习算法,以提高3D肺结节的分类性能。在传统自监督训练的网络结构基础上,设计了一种多任务学习的网络结构,以同时利用医学图像处理任务中大量未注释数据和少量注释数据。通过先训练未注释数据然后加入注释数据继续训练的方式,实现了注释数据与未注释数据间部分网络结构和参数的共享。相较于传统自监督学习方法,所提算法在保证模型泛化能力的同时能够学习到更多与肺结节相关的鉴别特征,因此将模型迁移学习用于肺结节分类时也能表现出更佳的性能。所提算法在公开数据集LIDC-IDRI上的分类准确率达0.886,曲线下面积(AUC)值达0.929,实验结果表明,所提算法能够有效提升肺结节的分类性能。

The process of training a deep learning model requires many annotation samples, even though annotation data is difficult to obtain in the medical field. A self-supervised learning algorithm combined with partial annotation data is proposed as a solution to this problem, in order to improve classification performance of 3D pulmonary nodules. Based on the traditional self-supervised training network structure, a multitask learning network structure is designed to address a large amount of unannotated data and a small amount of annotated data obtained from medical image processing tasks. First, the proposed algorithm trains the unannotated data, and then explores the annotation data to continuously train the model. Thus, this algorithm manages to share partial network structures and parameters between the annotated and unannotated data. Compared to traditional self-supervised learning methods, the proposed algorithm can learn to recognize the discriminant features of pulmonary nodules to ensure the model's capacity to generalize, therefore, model transfer learning can also perform better when applied to the classification of pulmonary nodules. The classification accuracy of the proposed algorithm on LIDC-IDRI dataset is 0.886, and the area under the curve (AUC) is 0.929. The results of the investigation indicate that the proposed algorithm can effectively improve classification performance of pulmonary nodules.