MobileNetV2 with transfer learning for brain tumor classification

Abstract

This document presents a deep learning-based approach for the automatic classification of brain tumors from magnetic resonance imaging (MRI) images, using the lightweight MobileNetV2 model combined with transfer learning and fine-tuning techniques. The study aims to address the constraints of resource-limited medical environments, where model speed and lightweight design are as important as accuracy. The dataset used is a fusion of three public databases (Figshare, SARTAJ, and BR35H), comprising 4,480 images for training and 1,600 for testing, divided into four classes: glioma, meningioma, pituitary tumor, and healthy brain. The methodology includes several steps: resizing the images to 224×224 pixels, normalizing pixel values between 0 and 1, augmenting the data through random rotations, shifts, and zooms to avoid overfitting, and then extracting features using MobileNetV2 pre-trained on ImageNet. The strategy adopted comprises two phases: first, transfer learning where only the layers added at the top of the model are trained for 10 epochs, then fine-tuning consisting of unfreezing the last 20 layers of the base model and retraining them with a reduced learning rate for 5 epochs. The results obtained show an overall accuracy of 86.56% after fine-tuning, with a macro-mean area under the ROC curve (AUC) of 0.9645, indicating excellent discriminatory power. The confusion matrix reveals that the "no tumor" class achieves perfect performance (400 out of 400), while the "meningioma" class remains the most difficult to classify, often confused with gliomas and pituitary tumors. Compared to more resource-intensive models like VGG16, ResNet50, or EfficientNet, MobileNetV2 offers an optimal balance between performance and lightweight design, with a significantly lower number of parameters, making it particularly well-suited to resource-constrained environments. The authors conclude that this approach provides reliable diagnostic support for radiologists, accelerating tumor detection without replacing medical expertise. Future directions include clinical validation on multi-center data, integration of an automated segmentation step, and exploration of newer architectures such as attention mechanisms.