UGPDA and BDAI: Robust Medical Semi-Supervised Learning Under Varied Labeled Ratios and Class-Imbalance Profiles

Semi-supervised learning (SSL) holds huge potential for medical image classification by using vast pools of unlabeled data to reduce the burden of expert annotation. Standard SSL frameworks like FixMatch experience performance degradation when deployed under real-world clinical environments because of data scarcity and class imbalance. In these environments, unfiltered pseudo-labels create confirmation bias, resulting in the model learning the majority classes. To overcome these limitations, we present a unified framework integrating two novel modules: Uncertainty Guided Pseudo-Label Acceptance (UG- PDA) and Belief Distribution Augmentation Invariance (BDAI). UGPDA acts as a targeted entropy gate that checks the validation status of weak-view predictions using Shannon entropy to filter out high-uncertainty samples. On the other hand, BDAI preserves the full soft probabilistic target distributions (belief distributions) of accepted samples. It utilizes an asymmetric Kullback-Leibler divergence constraint to align strongly augmented views without forcing minority class nuances into rigid one-hot targets. By testing on chest X-ray dataset under varied labeled ratios (5%, 10%, 15%, and 20%) shows that our approach significantly outperforms state-of-the-art baselines. At an extreme 5% labeling within a native 3 : 1 imbalanced distribution, our framework achieves an F1-Score of 0.7521, an AUC-ROC of 0.9523, and a clinical Recall sensitivity of 0.9974. This establishes robust diag- nostic stability where traditional techniques undergo catastrophic class collapse.