Comprehensive evaluation of preprocessing pipeline depth in deep learning-based brain tumor classification using CNN and vision transformer architectures

Abstract Brain tumor classification via MRI remains a critical challenge. This study presents a systematic evaluation of how preprocessing pipeline depth influences the performance of deep learning models, including both convolutional neural networks (CNNs) and Vision Transformer (ViT) architectures. A five-stage progressive preprocessing framework (denoising, contrast enhancement, edge sharpening, gamma correction, normalization) was designed and evaluated on a balanced MRI dataset of 8,000 images (glioma, meningioma, pituitary, normal). Comprehensive analysis revealed significant accuracy improvements (up to + 4.5%) with deeper preprocessing, especially for DenseNet121 and ViT Large. Stability analysis identified ViT Base R50 and VGG19 as the most robust architectures across varying preprocessing intensities. A composite clinical balance score, integrating performance, efficiency, and parameter load, ranked ViT Base R50 as the most suitable model for clinical deployment. This study emphasizes the pivotal role of preprocessing and proposes evidence-based guidelines for its design in clinical AI.