摘要：

COVID-19是由新型冠状病毒毒株引发的疾病。现有COVID-19影像诊断模型存在优质样本匮乏、未充分挖掘样本间关系等问题。提出面向COVID-19影像诊断的注意力蒸馏对比互学习(ADCML)模型。构建递进式数据增强策略，包括自动数据增强与样本过滤，通过扩充图像数量并确保其质量主动应对优质样本匮乏问题；设计注意力蒸馏对比互学习框架，利用注意力蒸馏驱动异构网络互相学习各自注意力关注的病理知识，进而提取样本间对比关系，改善特征判别性；采用自适应模型融合模块，充分挖掘异构网络间互补性，完成COVID-19影像诊断。在3个公开数据集(包括CT图像和X-Ray图像)上的实验结果表明，该模型准确率分别达到89.69%、98.16%和98.91%，F1值分别达到88.62%、97.58%和98.47%，AUC分别达到88.95%、97.77%和98.90%。ADCML模型优于主流基线，具有较强鲁棒性，且递进式数据增强、注意力蒸馏及对比互学习形成合力，共同推动模型性能提升。

关键词: COVID-19影像诊断, 注意力蒸馏, 对比互学习, 样本过滤, 自适应模型融合

Abstract:

COVID-19 is an illness caused by a strain of the novel coronavirus. Existing COVID-19 imaging diagnostic models face challenges such as the lack of high-quality samples and insufficient exploration of inter-sample relationships. This paper proposes a novel model called Attention Distillation Contrastive Mutual Learning (ADCML) for COVID-19 diagnosis, to address these two issues. First, a progressive data augmentation strategy is constructed, which includes AutoAugment and sample filtering, and the lack of quality samples is proactively addressed by expanding the number of images and ensuring their quality. Second, the ADCML framework is built, which employs attention distillation to motivate two heterogeneous networks to learn from each other the pathological knowledge concerned with their attention. The implicit contrastive relationships among the diverse samples are then fully mined to improve the discriminative ability of the extracted features. Finally, a new adaptive model-fusion module is designed to fully mine the complementarity between the heterogeneous networks and complete the COVID-19 image diagnosis. The proposed model is validated on three publicly available datasets-including Computed Tomography (CT) and X-ray images-with accuracies of 89.69%, 98.16%, and 98.91%; F1 values of 88.62%, 97.58%, and 98.47%; and Area Under the Curve (AUC) values of 88.95%, 97.77%, and 98.90%, respectively. These results show that the ADCML model outperforms the mainstream baselines and has strong robustness, and that progressive data augmentation, attention distillation, and contrastive mutual learning form a type of joint force that promotes the final classification performance.

Key words: COVID-19 image diagnosis, attention distillation, contrastive mutual learning, Sample Refinement (SR), adaptive model fusion