OdCFU-Net for Brain Tumor Segmentation Based on MRI

In the field of medical image analysis, the accurate segmentation of brain tumors plays a crucial role in the diagnosis and treatment of brain-related diseases. This study aims to address key challenges in this domain, with a particular focus on improving feature extraction, information integration, and model generalization capabilities. By concurrently introducing multi-dimensional attention mechanisms in convolutional operations, a method has been developed to dynamically generate convolution kernels in spatial, input channel, and output channel dimensions. This approach optimizes the limitations of traditional static convolutional kernels, maintaining high-performance efficiency while reducing the model’s parameter count. To overcome the deficiencies of traditional convolutional networks in capturing global information, channel and spatial attention mechanisms have been integrated between convolutional layers. This enhances adaptability and global information retrieval capabilities, allowing the model to dynamically adjust the weight of feature maps based on different task requirements. Leveraging the advantages of Vision Transformer in handling global information and addressing its limitations in capturing details and training on small datasets, the algorithm aligns and learns feature representations effectively by incorporating channel attention and normalizing attention weights within its multi-head attention. When applied to the BraTs2021 multi-modal brain tumor image database, the proposed algorithm accurately delineates brain tumor regions, demonstrating excellent performance.