Semisupervised Classification of Brain Tumor Images Using Gradient Adversarial Consistency on Mixup

The scarcity of labeled brain MRI datasets presents a significant challenge, as manual labeling is resource‐intensive. This lack of data can impede the training and generalization of convolutional neural network (CNN) models, increasing the risk of overfitting and diminishing performance on real‐world clinical data. By incorporating labeled and unlabeled data, semisupervised learning techniques aim to build precise models. Effectively using unlabeled data to boost the accuracy of models trained on limited labeled data remains a crucial challenge. To address this, we propose a gradient adversarial consistency on Mixup (GAM), a semisupervised framework. GAM uses a Mixup of labeled images to shift the classifier’s decision boundary toward the interclass region. It also applies the GAM of unlabeled data to smooth the boundary. This allows GAM to leverage unlabeled patterns and learn robust features, improving brain tumor classification. The research evaluates a semisupervised CNN‐13 classifier on a combined brain dataset from three imaging studies, exploring the impact of generating novel augmentation as gradient noise on the Mixup of unlabeled images. GAM achieved statistically significant accuracy improvements ( p < 0.05), reaching 0.8154 and 0.8352 with 50 and 100 labeled images per class, respectively, averaged over five runs, compared to baseline accuracies of 0.7833 and 0.8225. Experimental results demonstrate that GAM outperforms previous techniques, particularly in enhancing performance near decision boundaries. The findings indicate that GAM effectively leverages both labeled and unlabeled data, resulting in improved classification accuracy for brain tumor detection, especially in situations where labeled data are scarce in medical imaging applications.