Multi-Class Brain Tumor MRI Classification Using MS-GHOST-RAAE on a Combined Figshare-Br35H-SARTAJ Dataset

This study proposes a hybrid deep learning framework for multi-class brain tumor classification using a combined MRI dataset constructed from Figshare, Br35H, and SARTAJ sources. The dataset includes 7023 MRI images belonging to four clinically important classes: glioma, meningioma, no-tumor, and pituitary tumor. In the proposed approach, a dedicated preprocessing pipeline is first applied to enhance tumor-related image regions and reduce irrelevant background information. Then, the Multi-scale GHOST Residual Attention Autoencoder (MS-GHOST-RAAE) is used for deep feature extraction. This architecture integrates multi-scale GHOST modules, residual connections, and attention mechanisms to obtain compact, stable, and discriminative feature representations. The extracted features are subsequently classified using conventional machine learning classifiers. Experimental results show that the proposed hybrid framework achieves 98.75% accuracy on the combined dataset, with a total processing time of 543 + 269.4571 s. These findings indicate that MS-GHOST-RAAE provides strong classification performance on heterogeneous MRI images and offers an effective computer-aided decision-support approach for brain tumor classification.