Early and reliable identification of intracranial tumors from MRI can shorten time to treatment and reduce invasive diagnostic procedures. This study investigates supervised deep learning for four-way classification of brain MRI into glioma, meningioma, pituitary tumor, and no-tumor using a curated dataset of 7,023 T1-weighted images split into disjoint training and test sets. Five modern convolutional backbones MobileNetV2, EfficientNetB0, ResNet50, DenseNet121, and ConvNeXt-Tiny were fine-tuned with identical preprocessing stratified sampling and class balanced mini batches to ensure fair comparison. To improve generalization the pipeline applied contrast limited adaptive histogram equalization intensity standardization center crop and resize to 224×224 and light spatial photometric augmentation training used label smoothing cosine learning rate decay and early stopping. Among the evaluated models, ConvNeXt-Tiny produced the most consistent performance on the held-out test set, achieving an overall accuracy of 0.96, an F1-score of 0.96. While EfficientNetB0 and DenseNet121 offered competitive accuracy with fewer parameters, ConvNeXt-Tiny demonstrated superior robustness in distinguishing tumor subtypes. MobileNetV2 provided the best compute-latency trade-off for edge deployment. The results indicate that careful normalization and balanced fine-tuning narrow the gap between compact and heavier architectures, and that modern ConvNeXt designs can yield strong accuracy without excessive model size.