SGC-ARANet: scale-wise global contextual axile reverse attention network for automatic brain tumor segmentation

SGC-ARANet: scale-wise global contextual axile reverse attention network for automatic brain tumor segmentation

Automated brain tumour segmentation using magnetic resonance imaging (MRI) is essential for clinical decision-making and surgical planning. Numerous studies have demonstrated the feasibility of segmenting brain tumours using deep learning models such as U-shaped architectures. Unfortunately, due to the diversity of tumors and complex boundaries, it is insufficient to obtain contextual data on tumor cells and their surroundings from a single stage. To overcome this limitation, we proposed a Scale-wise Global Contextual Axile Reverse Attention Network (SGC-ARANet) consisting of four modules that improve segmentation performance. We begin by creating three global multi-level guidance (GMLG) modules to provide various levels of global contextual data. Additionally, we develop a scale-wise multi-level blend module (SWMB) that dynamically blends multi-scale contextual data with high-level features. Following that, we demonstrated how a partial decoder (PD) connected in parallel to the encoder is utilized to aggregate high-level and SWMB feature maps to create a global map. The axile reverse attention (ARA) module is then presented to simulate multi-modality tumor regions and boundaries using global and GMLG feature maps. We evaluate our model using the publicly available BraTS 2019 and 2020 brain tumor segmentation datasets. The results indicate that our SGC-ARANet is competitive or outperforms numerous State-of-the-art (SOTA) algorithms for several segmentation measures.