<p> MRI Brain Tumor Classification using Patch-Level Graph Attention Networks</p>

Magnetic Resonance Imaging (MRI) plays a crucial role in the diagnosis and characterization of brain tumors; however, automated classification remains challenging due to tumor heterogeneity, subtle boundary variations, and the complex spatial relationships among tumor regions. Conventional convolutional neural networks (CNNs) primarily focus on local feature extraction and often lack explicit mechanisms to model long-range inter-patch dependencies within medical images. To address these limitations, this work proposes a Patch-Level Graph Attention Network (PL-GAT) framework for MRI brain tumor classification. In the proposed approach, each MRI slice is partitioned into non-overlapping patches, which are independently encoded using a lightweight CNN to extract discriminative local features. These patch embeddings are then represented as nodes in a graph, where spatial adjacency relationships define the graph structure. A graph attention network is employed to selectively weight inter-patch interactions, enabling adaptive aggregation of region-level information. The model is trained end-to-end using multi-class cross-entropy loss and evaluated on a multi-class MRI brain tumor dataset. Experimental results demonstrate that the proposed CNN-GAT hybrid model effectively captures both local texture patterns and global structural dependencies, leading to improved classification performance. Attention heatmaps and Receiver Operating-Area Under the Curve analysis further provide interpretability by highlighting diagnostically relevant image regions, making the approach suitable for clinically explainable brain tumor analysis.