Brain-MRINet: A Parallel Multi-Scale Convolutional Framework for Brain Tumor Classification From MRI

Abstract

Brain tumor diagnosis from Magnetic Resonance Imaging (MRI) is a clinically critical and error-prone task when performed manually. This research work proposes Brain-MRINet, an automated deep learning (DL) framework for multi-level brain tumor classification, designed to reduce radiologist workload and improve diagnostic reproducibility. Brain-MRINet is built upon a VGG16 model extended with a parallel convolutional block comprising three branches of $1\times 1$ convolution for feature compression, a depthwise separable $3\times 3$ convolution for computationally efficient spatial refinement, and a standard $3\times 3$ convolution for local feature extraction. The concatenated multi-scale feature maps are further refined through additional convolutional layers with batch normalization, followed by global average pooling and fully connected layers with dropout regularization. The framework is evaluated on binary (tumor vs. no tumor), four-class (glioma, meningioma, pituitary tumor, no tumor), and seventeen-class fine-grained tumor subtype classification. Five publicly available MRI datasets (DS-I: Br35H binary dataset, DS-II: Brain Tumor Segmentation dataset, DS-III: 4-class Brain Tumor MRI dataset, DS-IV: Crystal Clean Brain Tumor MRI dataset, and DS-V: 17-class fine-grained tumor dataset; totalling over 44,000 images) are used for training and evaluation, with preprocessing including resizing, normalization, and data augmentation. Performance was benchmarked against VGG16, VGG19, DenseNet121, ResNet50, MobileNet, and EfficientNetB0. Using an 80/10/10 train-validation-test split, Brain-MRINet achieved 100% accuracy on binary classification (DS-I), maintained 99.43% and 99.45% accuracy on four-class datasets (DS-III and DS-IV), and attained 95.48% accuracy on the challenging seventeen-class fine-grained dataset (DS-V), outperforming all baseline models across all five benchmarks. EfficientNetB0 and ResNet50 exhibited poor generalization, particularly on multi-class tasks. These results demonstrate Brain-MRINet as a promising framework for scalable, automated brain tumor classification, with potential to support radiologist workflows in future clinical validation studies.