A Computationally Efficient Brain Tumor Segmentation Framework From Magnetic Resonance Imaging Using Contrast Enhancement And Region Of Interest Localization With A Standard U-Net Architecture

Abstract

Extracting brain tumor from MRI with high accuracy is a challenging task due to its low contrast, intensity inhomogeneity, and class imbalance. This study assesses the impact of standard preprocessing techniques on plain U-Net segmentation performance in a systematic manner. The pipeline applies median filtering followed by Contrast Limited Adaptive Histogram Equalization (CLAHE) and automatic region of interest (ROI) extraction. A regular 2D U-Net with weighted binary cross-entropy loss deals with class imbalance. Analysis of 504 MRI scans according to a 60/20/20 division and the utilization of stepwise ablations reveal that CLAHE contributes +4.33% Dice, ROI extraction contributes +3.62% and weighted loss contributes +1.87%. The full framework obtains a Dice of 0.8618 and IoU of 0.7591 with an inference time of 12 ms and a training time of 237 minutes, making it feasible for resource-scarce clinical setups.