Diagnosing Tumor through an Efficient Lightweight CNN with Leakage-Free Evaluation on MRI Dataset

Deep learning models for brain tumor classification from MRI routinely report accuracies above 98% on public benchmarks.Many of these benchmarks, however, contain duplicate and near-duplicate images shared between training and test sets, which inflates reported performance.This study proposes a lightweight convolutional neural network (CNN) for four-class brain tumor classification (glioma, meningioma, pituitary tumor, no tumor) and evaluates it under a leakagefree protocol.The widely used Kaggle Brain Tumor MRI dataset (7,200 images) was first deduplicated with perceptual hashing and pixel-level verification, removing 1,994 duplicate images (27.7%); 39.5% of the original test images had a near-duplicate in the original training set.The remaining images were re-split into stratified training, validation and test sets.The proposed network combines depth wise-separable inverted residual blocks with efficient channel attention (ECA), and has only 0.60 M parameters and 0.17 G multiply-accumulate operations (MACs).On the deduplicated test set it achieves 97.3% accuracy and 97.0% macro F1-score, compared with 98.2% accuracy for the best of five standard CNN baselines (EfficientNet-B0), which have 3.7 to 224 times as many parameters.Five-fold cross-validation gives 97.2 ± 0.6% accuracy.Training the same models on the original split raised accuracy by 1.3-2.0percentage points, quantifying the effect of leakage.To know the effect of each component, an ablation experimentation is performed and for explainability, Grad-CAM visualizations has been performed to show that predictions focus on tumor regions.The proposed model is much situated to clinical application with low resources as it can classify an MRI image in 10.1 ms on an average.