Integrating Confidence Regularized Self-Training with XAI for Brain Tumor Classification

Brain tumor classification is a critical aspect of medical diagnosis, where accuracy and speed can significantly impact patient outcomes. The core problem in medical image analysis is the scarcity of labeled data, especially in complex tasks like brain tumor classification, where gathering labeled samples can be costly and time-consuming. Standard self-training approaches often fail in such settings due to overconfidence in incorrect predictions, leading to error propagation. To address these challenges, Confidence Regularized Self-Training for Brain Tumor Classification (CRST-BTC), an approach that enhances classification accuracy in limited data is proposed. To enhance interpretability, the research uses Explainable AI (XAI) using the SHAP framework, offering transparency in feature importance. Comparing the suggested approach to state-of-the-art (SOTA) models demonstrates its competitive advantage. The research thoroughly examines the effects of using different proportions of unlabeled data and evaluates various backbone networks, with VGG16 emerging as the top performer. The model shows optimal performance when 60% of the data is unlabeled, paired with a confidence threshold of 0.9. Notably, the VGG16 backbone achieves an accuracy of 97.01%. This research advances the field and offers important insights for professionals working in medical image analysis.