Multi-Modal MRI Brain Tumor Segmentation Using Attention Modules

Gliomas are one of the most aggressive brain tumors which can be a potential threat to life in their advanced stages. Manual detection of brain tumor is computationally expensive. Therefore, timely and precise localization of such tumors is crucial for diagnosis and further treatments. This study demonstrates the deployment of Attention Modules (AMs) in the baseline nnUNet (AM-nnUNet) for multi-sequence MRI brain tumor segmentation to achieve state-of-the-art performance. The AM-nnUNet utilizes BraTS 2020 dataset for training and validation using 5 -fold cross-validation (CV), exploring several variants incorporating Focal Tversky loss function (FTL), aggressive data augmentation (DA) techniques, and group normalization (GN) instead of instance normalization. Comprehensive experiments were carried out on BraTS dataset for different variant of proposed AM-nnUNet: AM-nnUNet + FTL, AM-nnUNet + FTL + DA, AM-nnUNet + FTL +GN and AMnnUNet + FTL +GN+DA. Experimental findings suggest AMnnUNet + FTL + GN + DA outperformed other AM-nnUNet varaints. AM-nnUNet + FTL + GN + DA, achieved dice scores of $0.9025,0.9387$, and 0.9568 for enhancing tumor (ET), necrotic tumor core (NC), and peritumoral edematous/invaded tissue (ED) labels, respectively. Comparative analyses were conducted with state-of-the art CNN architectures and nn-UNet based frameworks demonstrated that proposed framework yielded better segmentation performance.