Preprocessing to Visualization of Brain Tumor Segmentation with 3D U-Net

Abstract

Segmentation of brain tumors and its classification from MRI scans is a stimulating task in medical imaging. In this paper, we propose a hybrid deep learning framework that combines 3D U-Net with Transformer modules to improve brain tumor segmentation and classification accuracy. The 3D U-Net architecture captures local spatial features while the Transformer modules capture long-range dependencies in the image data. The proposed framework includes preprocessing steps such as bias field correction, Gibbs artifact removal and automatic region-of-interest extraction, followed by a hybrid encoder-decoder architecture with skip connections. The model also uses a multi-head attention mechanism to improve feature extraction. We use a combination of Dice loss and BCE loss function to handle class imbalance. Experimental results on the BraTS 20 21 dataset show that our hybrid model outperforms existing state-of-the-art methods with a Dice similarity coefficient (DSC) score of 93.38 %. We also demonstrate that our model achieves faster training time by 15% compared to a standard 3D U-Net. Additionally, ablation studies show the effectiveness of Transformer modules and preprocessing steps in our proposed framework. This model can be extended to other imaging modalities and can be deployed in real-world scenarios. Future work includes exploring real-time deployment, multimodal medical image segmentation, and model efficiency.