<scp>MR</scp> Brain Tumor Image Segmentation Based on <scp>EA</scp> ‐ <scp>UNet</scp>

ABSTRACT Brain tumors pose a significant threat to human health. Accurate segmentation is crucial for clinical diagnosis and treatment. Although the U‐Net model has been extensively utilized for medical image segmentation, it faces challenges in accurately segmenting brain tumors, particularly in capturing tumor feature information and effectively extracting global features. We propose a novel brain tumor segmentation model based on Efficient feature extraction and Adaptive enhancement to mitigate these challenges, called EA‐UNet. Our model incorporates an efficient GhostDenseVgg (GhostDV) Block in the encoder part to extend the convolution receptive field and improve the quality of features by stacking small convolution kernels and feature fusion. Additionally, the decoder part includes a Parallel Multi‐dimensional Dynamic Convolution (PMDC) block, combined with the residual mechanism, which enhances the adaptability of the input features and ensures the output feature map retains rich contextual information. On the BraTS 2019 dataset, our model achieves a Dice similarity coefficient of 81.42% for the whole tumor area, 72.46% for the tumor core area, and 71.99% for the enhanced tumor area, surpassing similar methods. Furthermore, the model's stability and robustness are effectively validated using the BraTS 2020 test set as an independent external dataset, providing strong evidence of its generalization ability for clinical brain tumor segmentation. The code is publicly available on https://github.com/chtxcl/EA‐UNet .