<![CDATA[Brain tumors are among the most serious neurological diseases, requiring accurate diagnosis to support effective treatment. Magnetic Resonance Imaging (MRI) is widely used for brain tumor detection because it provides detailed visualization of brain structures. However, manual MRI interpretation is time-consuming and highly dependent on radiologists' expertise, potentially leading to inconsistent diagnoses. This study evaluates the performance of several deep learning architectures, including VGG16, DenseNet121, Custom CNN, and SqueezeNet, for multiclass brain tumor classification using MRI images. To address the computational limitations of deep neural networks in resource-constrained medical environments, a structured pruning approach based on polynomial decay sparsity scheduling was applied to the SqueezeNet model to reduce model complexity while preserving predictive performance. Experiments were conducted on a multiclass MRI dataset consisting of glioma, meningioma, pituitary tumor, and non-tumor images. The proposed pruned SqueezeNet achieved a classification accuracy of 98.70% and an AUC of 0.998, while reducing the model size to 2,970 KB and achieving an inference time of 105.40 ms. These results demonstrate that structured pruning effectively improves computational efficiency without significantly compromising classification accuracy, making the model suitable for deployment in resource-limited medical settings. Nevertheless, further validation using larger and multi-institutional MRI datasets is required before clinical implementation]]>
