Abstract:

Proper brain tumor segmentation is crucial for medical image analysis and treatment planning. Manual segmentation is time-consuming, is subjective in nature, and can greatly differ among radiologists. This article introduces a deep learning approach employing an Attention U-Net architecture trained on the BraTS 2020 dataset to perform automated brain tumor segmentation. The suggested 2D Attention U-Net utilizes self-attention to emphasize region-of-interest with respect to tumors in MRI slices. It aids in the detection of tumor sub-regions such as the enhancing tumor, necrotic core, and edema. The model was utilized in the PyTorch MONAI framework and optimized using a Dice and Cross- Entropy combined loss to counter class overlap and class imbalance. In the experiments, the model yielded a Dice Similarity Coefficient (DSC) of 0.8461, reflecting robust segmentation. The results prove that the implementation of attention gates in the encoder-decoder structure improves attention to trouble spots and minimizes false positives. This yields a firm foundation for pre-operative planning and clinical decision support.