基于交替训练融合模型的COVID-19的CT影像辅助诊断

CT image auxiliary diagnosis of COVID-19 based on alternating training ensemble model

目的:在单图形处理器(GPU)上利用并探究人工智能的方法对COVID-19的CT图像进行图像识别。方法:基于深度学习提出一种基于融合模型(ensemble model)的交替训练算法。算法中融合模型由两个DenseNet以一种特定的结构组合而成融合模型,并引用自监督学习技术MoCo的思想,探究融合模型在多任务学习(multi-task learning, MTL)中的可行性。同时,在训练过程中基于多任务学习提出一种交替训练模式(alternate training mode, ALTM)来解决小数据集的过拟合问题,增加模型的泛化性能。结果:在新冠肺炎-CT影像(COVID-CT)测试集上,精确率(precision):0.86;召回率(recall):0.91;F1分数(F1-score):0.89;准确率(accuracy):0.88;曲线下面积(AUC):0.92。结论:通过在COVID-CT数据集上进行多组实验,证明了融合模型和ALTM的可行性,融合模型经过ALTM算法训练后融合模型性能有所提升。与其他基于COVID-CT数据集研究的方法相比,本方法训练的模型整体性能更好泛化性更强。

Objective:Image recognition on CT images of COVID-19 using artificial intelligence on a single GPU. Methods: Based on deep learning, we propose an alternate training which is based on ensemble model. In this paper, the Ensemble Model, is composed of two DenseNet with a specific structure, and explores the feasibility of our ensemble model in multi-task learning(MTL) by referring to the idea of MoCo. At the same time, an alternate training mode(ALTM) is proposed based on multi-task learning to solve the over-fitting problem of small dataset and increase the generalization performance of the model. Results: On the COVID-19 CT test set, our ensemble model achieve up to 0.86 precision, 0.91 recall, 0.89 F1-score, 0.88 accuracy and 0.92 AUC. Conclusion:The result on the COVID-CT dataset verify the feasibility of our ensemble model and ALTM. After ALTM training, the performance of the ensemble mode is improved. Compared with other methods based on the COVID-CT dataset, the ensemble model trained by ALTM has better overall performance and stronger generalization.