<![CDATA[Brain and central nervous system tumors remain a serious global health issue. In 2022, there were 308,102 new cases and 251,329 deaths worldwide, highlighting the urgency of this issue from a clinical perspective. Magnetic Resonance Imaging (MRI) is widely used to examine brain tumors, but the interpretation process remains highly dependent on radiologists and can be quite time-consuming. This study proposes an enhanced ResNet-50 architecture with optimized bottleneck layers to classify three types of brain tumors: glioma, meningioma, and pituitary tumors. The proposed model modifies the conventional ResNet-50 by embedding structured bottleneck units within the residual blocks. Each bottleneck unit applies a convolution for channel reduction, a convolution for spatial feature extraction, and a subsequent convolution for feature projection. The channel configuration is progressively arranged from 32 to 64, 128, and 256 channels across the deeper bottleneck units, enabling more discriminative hierarchical feature learning while maintaining architectural efficiency. The research process includes MRI image preprocessing, data augmentation, model training, confusion matrix-based evaluation, single-prediction testing, and estimated McNemar’s test. The proposed method achieves an accuracy of 98.96%, precision of 97.0%, recall of 98.2%, specificity of 99.5%, and an F1 score of 97.6%. The estimated McNemar's test yields values of 89.01, 15.06, 54.02, and 4.17 when compared to VGGNet-16, DenseNet-201, AlexNet, and the standard ResNet-50, respectively. All values exceed the critical value of 3.84 at the 0.05 significance level, indicating that the proposed method produces significantly fewer classification errors than the compared models. These results indicate that optimizing the bottleneck layer can improve the classification performance of ResNet-50 and provide a reliable framework.]]>
