A Continuous Feature Learning-Optimized Semi-Supervised Framework for Medical Image Classification

Medical image classification tasks often suffer from limited labeled data, while unlabeled data is plentiful. As a result, semi-supervised learning has become a crucial approach to enhance model performance. This paper presents a feature learning-optimized semi-supervised learning framework, which significantly improves classification accuracy through a two-stage training strategy. In the warm-up phase, we combine weak supervision and contrastive learning to optimize feature learning using a prototype-based parameterized classifier. In the retraining phase, we assess the confidence of unlabeled data using a Gaussian Mixture Model (GMM) to generate pseudo-labels. These labels are incorporated into joint training with labeled data, while the rest of the unlabeled data undergoes cross-prediction to refine the model's decision boundaries. Experimental results demonstrate that the proposed framework excels on the NCT-CRC-HE and ISIC2018 medical datasets, significantly improving classification performance and outperforming existing semi-supervised learning methods.