Date of Publication :5th December 2025

Abstract: Brain tumor detection remains one of the most challenging tasks in medical image analysis due to the variability in tumor shapes, sizes, and locations, as well as differences across MRI scanners. To address these challenges, we propose an Optimized Deep Learning Architecture for MRI Brain Tumor Detection Using Modified YOLOv5, incorporating a robust preprocessing pipeline to enhance model performance. The pipeline begins with N4ITK bias field correction to remove MRI intensity inhomogeneity, followed by skull stripping for eliminating non-brain tissues. Subsequently, z-score normalization ensures standardized intensity distributions, while Nyul histogram normalization harmonizes images acquired from different scanners. To further improve consistency, bilinear resampling resizes the images to the YOLOv5 input dimension (640×640). Data variability and model robustness are enhanced through extensive data augmentation, including mosaic/mixup for object detection, random affine transformations, elastic deformation for MRI-realistic distortions, and brightness/contrast adjustments for scanner variation. The proposed Modified YOLOv5, enhanced with advanced feature fusion and attention mechanisms, leverages these optimized inputs for accurate and efficient tumor localization and classification. Experimental results demonstrate improved generalization, robustness, and detection accuracy compared to conventional approaches, highlighting the potential of our method for reliable clinical application in brain tumor diagnosis.

Reference :