Enhanced brain tumor detection and classification using a deep image recognition generative adversarial network (DIR-GAN): a comparative study on MRI, X-ray, and FigShare datasets

Given its significant threat to human health, early and precise diagnosis of brain cancer is crucial for improving patient prognosis and treatment efficacy. Although magnetic resonance imaging (MRI) is essential for detecting brain tumors, current deep learning techniques face challenges in managing noise interference and accurately defining tumor boundaries. This study introduces a novel hybrid framework called deep image recognition generative adversarial network (DIR-GAN), which aims to enhance the accuracy and robustness of brain tumor detection and classification in MRI scans. The DIR-GAN approach was designed to address the limitations of the existing methods in processing and analyzing brain imaging data. The proposed methodology integrates cutting-edge techniques in multiple stages. Noise reduction is achieved through adaptive median-bilateral filtering (AMBF), which effectively removes noise while preserving fine structural detail. To achieve precise tumor segmentation, the deep superpixel-based attention segmentation (DSAS) method combines superpixel generation with a hierarchical attention mechanism to achieve highly accurate tumor region delineation. Feature extraction is optimized using a hybrid approach based on the gray-level co-occurrence matrix (GLCM) and chaotic dragonfly algorithm (CDA), leveraging chaotic maps to enhance convergence speed and exploratory efficiency. These optimized features are fed into the DIR-GAN, which employs attention-enhanced multi-scale feature extraction and adversarial training to synthesize high-quality MRI images and strengthen the classification performance. Hybrid residual connections in the generator and discriminator further improved feature learning and classification reliability. The proposed framework was implemented in Python and evaluated on three datasets: Share, MRI, and X-ray. The DIR-GAN demonstrated state-of-the-art performance, achieving an accuracy of 98.90% on the FigShare dataset, 98.60% on the MRI dataset, and 97.93% on the X-ray dataset. This hybrid framework offers a robust and interpretable solution for the early diagnosis and classification of brain tumors, setting a new benchmark for clinical applications and significantly improving patient outcomes. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2025.