Segmentation and Classification of Magnetic Resonance Images

Abstract

Accurate diagnosis and early treating of brain tumors have a positive and significant impact on a person's quality of life. For this reason, several studies are currently being conducted to improve medical image processing and the detection and accurate localization of brain tumors. This has led to the implementation of Deep Learning (DL) in medical image analysis. Thus, this study presents a lightweight U-Net to reduce computational costs in the detection of brain tumors. This detection was performed by selecting the best slice from the Magnetic Resonance Images (MRI) of several modalities (T1C, T2W, and T2F) in the BrasTS2024 dataset. The proposed model automatically selects the best slice by extracting the slice with the major tumor
content from the mask. An adaptive learning-rate scheduler was applied during training to ensure stable convergence. Among the modalities, T2W produced the best quantitative results, achieving a Dice coefficient of 0.760, an IoU of 0.623, and a sensitivity of 0.842. In contrast, T2F yielded visually comparable yet slightly lower metrics. These findings indicate that reducing model complexity needs not compromise accuracy and highlight the T2W sequence as the most informative single slice for tumor delineation