Robust Semi-supervised Medical Image Classification: Leveraging Reliable Pseudo-labels

The field of semi-supervised learning (SSL) has fostered new techniques to increase the performance of machine learning models in interpreting medical images. This paper introduces a groundbreaking approach for medical image classification, which combines pseudo-loss approximation and adversarial distortion. Our model improves the learning process by offering a more accurate evaluation of pseudo-labels attached to unlabeled data. It uses pseudo-labeled data that are both trustworthy and meaningful, resulting in higher categorization accuracy. Moreover, adversarial distortion is added to unlabeled data through a cross pseudo-loss approximation strategy. This unique technique allows us to unlock the hidden value in previously ignored data, thereby further boosting our model’s performance. We have conducted extensive experiments on two medical datasets, including the NCT-CRC-HE, to illustrate our model’s efficacy and adaptability under various test scenarios. Comparative results, showing a consistent performance improvement over other SSL methods, underline the potential of our approach in redefining boundaries in semi-supervised medical image classification tasks, highlighting its promise to significantly contribute to the medical image analysis field.