TRANSFORMER NETWORK FOR BRAIN GLIOMA SEGMENTATION IN MRI IMAGES

Glioma is a form of tumor that grows in the brain or spinal cord, forming a mass that can press on surrounding tissue and cause symptoms. To diagnose glioma and to assess the tumor volume, manual segmentation of gliomas in MRI images is normally performed. However, manual segmentation takes time and is exposed to human errors due to diagnostic variability among experts. This study proposes a deep learning approach using a Transformer Network to enhance segmentation accuracy and improve diagnostic efficiency.The research utilizes the BraTS 2021 dataset, consisting of 374 MRI scans with ground truth labels, to train and evaluate the Transformer Network model. The model incorporates an EfficientNet B1 backbone for computational efficiency and is trained with optimal parameters: a learning rate of 0.0001, batch size of 5, and 200 epochs. Results indicate that the Transformer Network achieved a Dice coefficient of 0.921, significantly outperforming the baseline deep learning segmentation method, which is U Net model (0.827), demonstrating superior segmentation accuracy.In conclusion, the Transformer Network proves more effective and accurate than traditional methods for brain glioma segmentation. Future research should focus on expanding datasets and computational resources to further enhance model performance. This study is expected to contribute to an improved glioma diagnosis and treatment planning.