Background and objective: Self-supervised learning is rapidly advancing computer-aided diagnosis in the medical field. Masked image modeling (MIM) is one of the self-supervised learning methods that masks a subset of input pixels and attempts to predict the masked pixels. Traditional MIM methods often employ a random masking strategy. In comparison to ordinary images, medical images often have a small region of interest for disease detection. Consequently, we focus on fixing the problem in this work, which is evaluated by automatic COVID-19 identification. Methods: In this study, we propose a novel region-guided masked image modeling method (RGMIM) for COVID-19 detection in this paper. In our method, we devise a new masking strategy that employed lung mask information to identify valid regions to learn more useful information for COVID-19 detection. The proposed method was contrasted with five self-supervised learning techniques (MAE, SKD, Cross, BYOL, and, SimSiam). We present a quantitative evaluation of open COVID-19 CXR datasets as well as masking ratio hyperparameter studies. Results: When using the entire training set, RGMIM outperformed other comparable methods, achieving 0.962 detection accuracy. Specifically, RGMIM significantly improved COVID-19 detection in small data volumes, such as 5% and 10% of the training set (846 and 1,693 images) compared to other methods, and achieved 0.957 detection accuracy even when only 50% of the training set was used. Conclusions: RGMIM can mask more valid lung-related regions, facilitating the learning of discriminative representations and the subsequent high-accuracy COVID-19 detection. RGMIM outperforms other state-of-the-art self-supervised learning methods in experiments, particularly when limited training data is used.

翻译：背景与目的：自监督学习正快速推动医学领域的计算机辅助诊断发展。掩码图像建模（MIM）是一种自监督学习方法，通过对输入像素的子集进行掩码并尝试预测被掩码的像素来实现学习。传统MIM方法通常采用随机掩码策略。与普通图像相比，医学图像中与疾病检测相关的感兴趣区域往往较小。为此，本文聚焦于解决该问题，并通过自动COVID-19识别任务进行评估。方法：本研究提出一种新颖的基于区域引导的掩码图像建模方法（RGMIM），用于COVID-19检测。该方法设计了一种新的掩码策略，利用肺部掩码信息识别有效区域，从而学习对COVID-19检测更有用的特征。我们将所提方法与五种自监督学习技术（MAE、SKD、Cross、BYOL、SimSiam）进行对比，并基于公开COVID-19胸部X光数据集进行定量评估，同时开展掩码率超参数研究。结果：使用完整训练集时，RGMIM优于其他对比方法，检测准确率达0.962。特别地，在小数据量场景下（如使用训练集的5%和10%，对应846张和1,693张图像），RGMIM显著提升了COVID-19检测性能；即便仅使用50%训练集，其检测准确率仍可达0.957。结论：RGMIM能够更有效地掩码与肺部相关的有效区域，促进判别性表征的学习及后续高精度COVID-19检测。实验表明，尤其在训练数据有限的情况下，RGMIM性能优于其他当前最优的自监督学习方法。