3D Brain Tumor Segmentation in Multimodal MRI Images

In brain tumor research, the use of machine learning based methods using MRI data to locate brain tumors accurately has significantly increased and new techniques are proposed to facilitate the medical experts. There is still room for improvement as tumors have complex anatomy which requires efforts to explore further and propose new deep learning based solutions. In this research, we propose 3D UNet based multimodal brain tumor segmentation for Non Enhancing Tumor (NET), Peritumoral Edema (PE), and Enhancing Tumor (ET). Data preprocessing and augmentation techniques are employed to optimise the model's performance. The dataset used in this study to train, validate, and test the model performance is the BraTS 2020 dataset. The model achieved improved accuracy and IoU after implementing the mentioned techniques. Results before and after data augmentation demonstrate a significant enhancement in model performance. The model achieved an improved mIoU of 0.77 and accuracies of 98.21% and 99.36% for actual data and augmented data for three brain tumor classes, i.e., NET, PE, and ET. Our model performed multiclass brain tumor segmentation with high accuracies, and we aim to enhance this study to propose a valuable medical assistance model.