ResNet50 Versus VGG16 for MRI Brain Tumor Detection: A Transfer-Learning Study with Visual Explanation

Abstract

Accurate detection of brain tumors from magnetic resonance imaging (MRI) is essential for early diagnosis and treatment planning. However, manual interpretation of MRI scans is time-consuming, requires experienced radiologists, and is subject to inter-observer variability. To address these challenges, this study proposes a transfer-learning-based framework for binary brain tumor classification using ResNet50 and VGG16 pretrained on ImageNet. The data come from the public Kaggle brain-MRI collection of 253 scans. On top of it we fine-tuned two ImageNet-pretrained networks through a single shared pipeline: ResNet50, whose residual connections keep gradients from vanishing in deep stacks, and VGG16 as a plain-CNN reference. ResNet50 outperformed the VGG16 baseline. On the held-out test set it reached 91.1% accuracy, 91.7% precision, 88.0% recall on the tumor class, and a ROC-AUC of 0.9559, beating the VGG16 baseline. Because interpretability is important in clinical applications, Grad-CAM was incorporated to provide visual explanations of the model's predictions by highlighting image regions that contributed most to the classification decisions.