A Cascaded Dynamic Attention-Based Deep Learning Model Segments Brain Tumors

Brain tumor diagnosis and treatment need segmentation, Precision automated segmentation of brain tumors is difficult due to their size, shape, unexpected placements, and fuzzy boundaries. U-Net is a popular medical picture segmentation model due to its simple architecture and high performance. Underutilized contextual information, limited local receptive fields, and lost spatial information are issues. For brain tumor segmentation, CDAU-Net, a novel model with Dynamic Convolution and Non-Local Attention Mechanism, addresses these issues. Two-stage cascaded three-dimensional U-Nets rebuild brain tumor spatial information at better resolution. The cascaded network's lateral connections receive maximal attention, improving its ability to collect long-distance interconnections and use tumor context information. Finally, dynamic convolutions with local adaptive capabilities replace normal convolutions in the cascaded network to increase local features capture. Extensive experiments were conducted on publicly available BraTS 2019 and 2020 datasets, comparing the proposed method with other representative approaches. Proposed method segments brain tumors. Brain tumor segmentation dice values of 0.897/0.903 for whole tumor, 0.826/0.828 for tumor core, and 0.781/0.786 for enhanced tumor segmentation worked in the BraTS 2019/2020 validation set.