A hybrid approach based on deep learning to improve segmentation, classification and localization of different types of brain tumors in MRI images

Abstract

Brain tumors pose significant health risks, and early detection is critical for improving patient outcomes. Although Magnetic Resonance Imaging (MRI) is widely used for diagnosis, accurate interpretation remains challenging, particularly in resource-limited clinical settings. To address this issue, this study investigates both dedicated single-task models and a unified multi-task deep learning framework for automated brain tumor analysis. The single-task models were developed to perform classification, segmentation, and localization independently, while the proposed multi-task framework employs a shared DenseNet121 backbone with task-specific branches to jointly perform all three tasks within a single end-to-end architecture. This unified framework enables simultaneous tumor classification, segmentation, and localization from a single MRI scan, providing comprehensive diagnostic information with improved efficiency and interpretability. Experimental results demonstrate that the proposed multi-task framework achieved a classification accuracy of 96.25%, a Dice coefficient of 81.71%, and a localization mean squared error of 0.004, highlighting its potential for clinical decision support, particularly in low-resource healthcare environments.