Integrating ResNet50V2 and MobileNetV2 Models for Meningioma Detection
Femila Irene G, R Surendran · 2025
Detecting brain tumors early, especially meningioma, is important for treatment and depends on MRI scans. While deep learning methods like convolutional neural networks (CNNs) are useful in identifying tumors, most studies rely only on clean images and single models, which may not perform well when MRI quality is poor. This paper presents a novel ensemble-based approach that combines ResNet50V2 and MobileNetV2 and uniquely trains on both clean and noisy images. The key contribution is the inclusion of Gaussian noise during training, helping the model learn better generalization and noise resistance for real clinical scenarios where image quality often varies. The two networks extract different features: ResNet50V2 handles complex patterns with deep layers, and MobileNetV2 keeps the model lightweight for fast use in clinical environments. Their features are combined before classification. The model was trained using transfer learning on preprocessed and augmented MRI scans. On clean images, it achieved 100% accuracy, and precision, recall, and F1-scores of 1.00 or near 1.00 for both normal and abnormal classes. On noisy data, the model remained robust, with accuracy above 0.98 up to σ = 0.1 and 0.9168 at σ = 0.2. Although Grad-CAM visualizations couldn’t be applied due to the model’s branched structure, the experimental results show strong performance and robustness. These findings make the proposed model a promising and practical choice for real-world brain tumor detection, especially in hospitals with varied image quality. Future work can explore new ways of model interpretability and validation on larger datasets.