Diffusion-Enhanced Brain Tumor Classification Using Fast-DDPM and EfficientNet-B0 on MRI: A Controlled Comparative Evaluation of Baseline and Denoising-Enhanced Accuracy

Noise in clinical Magnetic Resonance Imaging (MRI) is a chronic and consequential impediment to automated brain tumor classification, degrading feature clarity at tumor boundaries and systematically limiting the discriminative power of deep learning models. This study presents a controlled comparative framework that directly quantifies the diagnostic benefit of Fast-DDPM generative denoising within a dual-branch evaluation pipeline: raw noisy MRI images are classified first by EfficientNet-B0 to establish a baseline, then the same images are restored by the Fast-DDPM denoiser and reclassified by the identical model checkpoint, and the two accuracy scores are compared. The classifier in both branches is EfficientNet-B0 with ImageNet-pretrained weights (IMAGENET1K_V1), a modified Linear(1280→4) classification head, Adam optimiser at lr = 1×10⁻⁴, CrossEntropyLoss, batch size 32, and five training epochs with Resize(224×224) preprocessing. The Brain Tumor MRI dataset (Figshare/Kaggle, four classes: Glioma, Meningioma, Pituitary, No Tumor) is partitioned into original and Fast-DDPM-enhanced splits. The baseline branch achieves 65.03% classification accuracy on raw noisy images. After Fast-DDPM preprocessing, the enhanced branch achieves 72.80%—an absolute gain of 7.77 percentage points, representing an 11.94% relative improvement, delivered without any modification to the classifier architecture, training procedure, or hyperparameters. These results establish generative diffusion preprocessing as a deployable, lightweight clinical strategy for improving automated brain tumor diagnostic accuracy.