Fuzzy Generative Learning with FCM-Based Segmentation for Brain Tumour MRI Classification

Abstract

Brain tumour analysis from magnetic resonance imaging (MRI) remains challenging because tumour classes exhibit heterogeneous intensities, overlapping morphology, and uncertain tissue boundaries. This paper presents a fuzzy generative learning framework that integrates Fuzzy C-Means (FCM), a Generative Adversarial Network (GAN), and VGG16 for four-class MRI classification. FCM performs soft intensity-based pre-segmentation and enhances ambiguous tissue transitions; its output is then used by the GAN to model the processed image distribution and by VGG16 to learn discriminative class features. The GAN supplies quality-controlled synthetic images to increase training diversity, while transfer learning reduces the amount of task-specific data required. Experiments used 3,264 publicly available MRI images representing glioma, meningioma, pituitary tumour, and no-tumour cases. The reported internal confusion matrix contained 2,860 correct predictions from 2,870 images, corresponding to 99.65% accuracy, 99.68% macro F1-score, and 99.88% macro specificity. Most errors occurred between glioma and meningioma, indicating that their similar appearance remains the principal challenge. The framework demonstrates the value of combining fuzzy pre-segmentation, generative augmentation, and transfer learning. However, the result is an internal estimate; pixel-level segmentation metrics, ablation analysis, patient-level splitting, external multicentre validation, and prospective clinical assessment remain necessary before diagnostic deployment.