Salient feature extraction using Attention for Brain Tumor segmentation

Abstract

The brain tumor is recognized as one of the most frequent tumors, with a significant mortality rate associated with its development. Segmentation of brain tumors involves distinguishing normal brain tissue from malignant tissue. When evaluating brain tumors, it is possible to determine the existence of tumor tissue quickly. However, accurate and reproducible segmentation and characterization of anomalies are not readily achievable. Consequently, several researchers have proposed various biomedical image segmentation methods to distinguish between tumor and normal brain tissue reliably. However, state-of-the-art segmentation has not been achieved by the existing brain tumor segmentation models, and they often come with high model complexity. Att-Sharp-U-net, a model influenced by the actual U-net model utilized in various medical image segmentation research, is presented as a contribution by this study. Two critical alterations to the underlying U-net model have been incorporated into the model: a grid-based attention block and a sharp block. By doing this, we were able to address the constraints of the U-net model while simultaneously enhancing segmentation performance with increasing negligible computational complexity. Experiments on the Brats2020 dataset, a recent publicly available benchmark dataset in brain tumor segmentation, showed that the proposed model improved segmentation performance with a dice score of 0.9275 and Jaccard score of 0.8684 when compared to the baselines.