HAB‐Net: Hierarchical asymmetric convolution and boundary enhancement network for brain tumor segmentation

Abstract

AbstractBrain tumour segmentation (BTS) is crucial for diagnosis and treatment planning by delineating tumour boundaries and subregions in multi‐modality bio‐imaging data. Several BTS models have been proposed to address specific technical challenges encountered in this field. However, accurately capturing intricate tumour structures and boundaries remains a difficult task. To overcome this challenge, HAB‐Net, a model that combines the strengths of convolutional neural networks and transformer architectures, is presented. HAB‐Net incorporates a custom‐designed hierarchical and pseudo‐convolutional module called hierarchical asymmetric convolutions (HAC). In the encoder, a coordinate attention is included to extract feature maps. Additionally, swin transformer, which has a self‐attention mechanism, is integrated to effectively capture long‐range relationships. Moreover, the decoder is enhanced with a boundary attention module (BAM) to improve boundary information and overall segmentation performance. Extensive evaluations conducted on the BraTS2018 and BraTS2021 datasets demonstrate significant improvements in segmentation accuracy for tumour regions.