Abstract
Accurate classification of brain tumors on magnetic resonance
imaging (MRI) is essential for diagnosis, treatment planning, and
prognostic assessment. However, manual interpretation is time-
consuming and subject to inter-observer variability. Automated deep
learning-based approaches may improve diagnostic consistency and
efficiency. To develop and comparatively evaluate a deep learning
framework using pretrained convolutional neural networks (CNNs) for
multi-class brain tumor classification on MRI. In this retrospective
experimental study, brain MRI images were categorized into four
classes: glioma, meningioma, pituitary tumor, and no tumor. Eight
pretrained CNN architectures—VGG16, VGG19, InceptionV3,
Xception, ResNet18, DenseNet121, MobileNetV2, and
EfficientNetB0—were evaluated under identical training and testing
conditions to ensure fair comparison. Transfer learning was applied to
all models. Performance was assessed using accuracy, precision, recall,
F1-score, and confusion matrix analysis. External validation was
performed on the BRISC 2025 public benchmark dataset to evaluate
generalizability. Lightweight architectures demonstrated performance
comparable to or exceeding deeper networks. ResNet18 achieved the
highest validation accuracy of 99.10%, with consistently balanced
precision, recall, and F1-scores across all tumor classes. Confusion
matrix analysis showed reduced misclassification between clinically
overlapping entities. Validation on the BRISC 2025 benchmark
confirmed the robustness and generalization capability of the proposed
framework. Pretrained lightweight CNN models, particularly
ResNet18, provide highly accurate and computationally efficient multi-
class brain tumor classification on MRI. The findings support their
potential integration into computer-aided diagnostic systems to
enhance radiological workflow and decision-making.
Authors
Shailesh Amarsinh Chaudhari
Veer Narmad South Gujarat University, India
Keywords
Brain Tumor, Magnetic Resonance Imaging, Deep Learning, Transfer Learning, Convolutional Neural Network, Computer-Aided Diagnosis