MRI-Based Brain Tumor Diagnosis Using Preprocessing Pipeline and Transfer Learning Models

Brain tumors represent a significant global health challenge, and their accurate and timely diagnosis is critical for effective treatment planning and improved patient outcomes. While transfer learning has shown promise in this domain, its performance is highly dependent on the quality of input data. This research introduces a novel, comprehensive 8-step preprocessing pipeline designed to significantly enhance the quality and feature visibility of Magnetic Resonance Imaging (MRI) scans for automated brain tumor classification. The pipeline includes resizing, grayscale conversion, Gaussian blurring, Otsu thresholding, contour isolation, Region of Interest (ROI) cropping, normalization, and a Power-law transform. To validate the efficacy of our proposed pipeline, we utilized a merged dataset of 10,287 MRI images from the Masoud and SARTAJ collections, encompassing four classes: glioma, meningioma, pituitary, and normal. This enhanced dataset was used to train and evaluate seven state-of-the-art transfer learning models: Xception, DenseNet-121, GoogLeNet, MobileNet, MobileNet-v2, VGG-19, and ResNet-50. Our rigorous preprocessing resulted in exceptional classification performance, with the Xception model achieving a peak accuracy of 98.68%. This study demonstrates that a meticulous and well-designed preprocessing pipeline is a critical and often overlooked component in developing highly accurate and reliable Computer-Aided Diagnosis (CAD) systems for clinical applications.