DSA: Deep Self-Attention Medical Transformer Neuro-Technology for Brain Tumor Segmentation

Abstract

Transformer-based methods have shown remarkable outcomes in medical image segmentation tasks. Specifically, the Swin Transformer has proven to be an impressive approach for segmentation jobs, demonstrating its potential to further the discipline. Extensive research on integrating Swin Transformer architecture with U-Net models has shown significant progress toward improving segmentation accuracy. Currently, researchers are looking for innovative methods to improve the challenging segmentation accuracy of enhanced tumor regions due to their heterogeneous and indistinct boundaries. To improve its accuracy, we have proposed a modified version of Swin UNETR, DSA, which is deeper and more focused on extracting global features by an enhanced self-attention mechanism in the later stages of the encoder. It outperformed the enhancing tumor class with comparative performance for the other two classes. By fine-tuning some hyperparameters, we achieved SOTA performance for brain tumor segmentation. The proposed deep self-architecture obtained a mean dice score value of 0.889 and a mean Jaccard score of 0.806, respectively. A comparison was conducted with some recent state-of-the-art techniques, which showed improved accuracy and outperformed the recent best-performing UNet and transformer architectures.