Brain tumor segmentation in 3D MRI images using W-Net architecture

Segmenting the 3D MRI images by the computer to identify the brain tumors is a very critical and challenging task till the invention of the deep learning algorithms. The previous works used some traditional methods like Mathematical Morphological Reconstruction (MMR), superpixel-level features extracted from 3D volumetric MR images, ensemble approaches, CNN, U-Net, etc., to achieve enhanced accuracy in segmenting different tumor regions. This study presents an innovative 3D brain tumor segmentation method using an extended W-Net architecture, a derivative of UNet, leveraging deep learning. Python programming on Google Colab facilitated the study, employing MRI scans from the BraTS dataset. The training dataset achieved a remarkable Dice Similarity Coefficient (DSC) and accuracy score of 0.98, showcasing the model's precision in tumor localization. The Matthews Correlation Coefficient (MCC) achieved 0.75, confirming the model's comprehensive segmentation quality. Generalization testing mirrored training outcomes, maintaining DSC and accuracy at 0.98, highlighting the model's robustness. The MCC, at 0.76, strengthened the model's ability to generalize to new data. This approach offers dependable and consistent segmentation outputs for 3-D brain MRI scans with tumor labels.