<p>This study focuses on the classification of brain tumors using advanced deep learning techniques applied to 3D MRI scans. Gliomas, highly invasive and life-threatening brain tumors, present significant challenges in accurate diagnosis and treatment planning. While traditional approaches rely on either 2D models that analyze individual MRI slices or 3D models that process entire volumetric data, this research explores a novel approach by leveraging spatio-temporal deep learning models adapted as spatio-spatial models for tumor classification. By interpreting the slice dimension in MRI scans as a pseudo-temporal axis, these models efficiently capture both in-plane spatial features and inter-slice correlations. To distinguish high-grade gliomas from healthy brain tissue and categorize them into low-grade kinds, two architectures are used: ResNet (2+1)D and ResNet Mixed Convolution. These models are evaluated against the standard ResNet3D architecture using the BraTS dataset and additional healthy brain scans from the IXI dataset. Metrics, including classification accuracy, computational efficiency, and the models’ capacity to extract intricate characteristics from volumetric data, are used to assess them. The experimental results show that spatio-spatial models improve classification performance, particularly in distinguishing low-grade gliomas, while reducing computational complexity. The study also looks at the advantages of model pre-training for transfer learning. Light augmentation strategies and transfer learning further enhance model generalization and training efficiency. This spatio-spatial adaptation provides a compelling alternative to traditional volumetric analysis methods, balancing accuracy, interpretability, and computational efficiency.</p>