<![CDATA[Optimizing the confidence threshold in pseudo-labeling is a critical technical challenge in Semi-Supervised Learning (SSL) for multi-class medical image classification. A threshold that is too strict limits the utilization of unlabeled data, whereas a threshold that is too lenient introduces low-quality pseudo-labels into the training process. This study applies the FixMatch method with the DenseNet-169 architecture as the backbone network to classify three lung disease categories COVID-19, Pneumonia, and Normal under conditions of extremely limited labeled data. The dataset used is the COVID-19, Pneumonia, and Normal Chest X-Ray Images dataset from Mendeley Data, consisting of 5,218 chest X-ray images, divided into 70% training, 10% validation, and 20% testing sets. The experiments were systematically designed using three labeled-data proportions (5%, 10%, and 15%) and three confidence threshold values (τ = 0.90, 0.95, and 0.99), resulting in nine experimental scenarios. The results demonstrate that τ = 0.95 with 15% labeled data achieved the best performance, obtaining 97.41% accuracy, a 97.49% F1-score, and an AUC of 0.9963. This performance was achieved by balancing pseudo-label selectivity with a sufficient volume of effective training data. At a low labeled-data ratio (5%), the limited amount of labeled data meant that the lower mask rate at τ = 0.95 could not be adequately compensated, allowing τ = 0.99 to perform slightly better. In contrast, at a higher labeled-data ratio (15%), the selectivity of τ = 0.95 produced high-quality pseudo-labels while maintaining sufficient data volume, leading to improved generalization performance. This study contributes an empirical analysis of confidence threshold sensitivity in FixMatch for multi-class chest X-ray classification under limited labeled-data conditions. These findings reveal that the effectiveness of the confidence threshold is highly dependent on the availability of labeled data, and that determining an optimal threshold cannot be separated from the proportion of labeled data available.]]>
