Main Article Content
Abstract
Brain tumor classification and segmentation are essential tasks for medical diagnosis and treatment planning. This study presents a comprehensive approach for simultaneous brain tumor classification and segmentation using a standard U-Net architecture applied to 2D brain MRI scans. The primary contribution of this work is the development of a novel balanced multi-class dataset comprising 6,380 MRI images created by merging and balancing two publicly available datasets, ensuring equal representation across four classes: no tumor, glioma, meningioma, and pituitary tumors (1,595 images per class). Our unified framework performs both pixel-level segmentation and image-level classification in a single forward pass, where classification is derived from segmentation outputs through spatial probability analysis. The standard U-Net model achieved robust performance with test accuracy of 99.62\%, Dice coefficient of 0.8423, and IoU of 0.9913. Image-level classification demonstrated precision and recall values ranging from 0.89 to 0.97 across all tumor classes. The perfectly balanced dataset eliminates class imbalance issues commonly encountered in medical imaging, enabling fair model evaluation and robust performance across all tumor types. This work provides a strong baseline for brain tumor analysis and demonstrates the effectiveness of proper dataset curation combined with classical deep learning architectures for medical image analysis applications.