Optimizing the confidence threshold in pseudo-labeling is a critical technical challenge in Semi-Supervised Learning (SSL) for multi-class medical image classification, since an overly strict threshold limits the utilization of unlabeled data, while an overly loose threshold introduces low-quality pseudo-labels into the training process. This study applies the FixMatch method with a DenseNet-169 architecture as the backbone to classify three lung disease classes COVID-19, Pneumonia, and Normal under conditions of highly limited labeled data. The dataset used is the Covid19, Pneumonia, and Normal Chest X-Ray Images dataset from Mendeley Data, comprising 5,218 images, split into 70% training, 10% validation, and 20% testing. The experiment was systematically designed using three proportions of labeled data (5%, 10%, 15%) and three confidence threshold values (τ = 0.90; 0.95; 0.99), resulting in nine experimental scenarios. The results show that τ = 0.95 with 15% labeled data achieved the best performance (accuracy 97.41%, F1-Score 97.49%, AUC 0.9963) because it balances pseudo-label selectivity with a sufficient effective data volume: at a low label ratio (5%), the limited volume of labeled data means the lower mask rate at τ = 0.95 is not adequately compensated, so τ = 0.99 performs slightly better; whereas at a high label ratio (15%), the selectivity of τ = 0.95 produces high-quality pseudo-labels with adequate volume, driving improved generalization. This study contributes an empirical analysis of confidence threshold sensitivity in FixMatch for multi-class CXR classification with limited labeled data. These findings reveal that the effectiveness of the confidence threshold is contextual to label availability, and determining the optimal threshold cannot be separated from the available labeled data ratio.