Deep Learning Architectures for Brain Tumor MRI Classification: A Comparative Benchmarking Study

Abstract

Brain tumor classification from Magnetic Resonance Imaging (MRI) is an essential challenge in clinical neuro-oncology. This thesis presents a systematic comparative study of eight deep learning architectures ResNet50, DenseNet121, VGG19, EfficientNet-B3, EfficientNetV2- S, ConvNeXt-Tiny, ConvNeXt-Small, and Swin Transformer Tiny evaluated on a dataset of MRIs collected from Nanfang Hospital, Guangzhou, China, and General Hospital, Tian jin Medical University, China, from 2005 to 2010 containing 3,064 T1-weighted contrast enhanced images from 233 patients across three tumor classes: meningioma, glioma, and pi tuitary adenoma. A significant methodological addition of this work is the adoption of a strict patient-level data splitting technique, which avoids data leaking, a critical weakness present in the majority of earlier work on this dataset. All models are trained via transfer learning from ImageNet, a large-scale benchmark of 1.28 million natural images across 1,000 cate gories, and adapted to the MRI classification task using a two-phase protocol: frozen back bone training followed by complete fine-tuning, including inverse-frequency class weighting to solve class imbalance. Results demonstrate that the Swin Transformer Tiny obtains the best accuracy of 93.13 % and the highest meningioma F1-score of 0.9118, demonstrating the advantage of hierarchical attention mechanisms for this task. Modern convolutional archi tectures (ConvNeXt family) closely follow, while older CNNs (ResNet50, EfficientNet-B3) show significantly lower generalization. Meningioma categorization is found to be the hard est class across all models. These findings establish a clinically meaningful framework for future studies on this dataset.