LSKA-Unet：结合大型可分离核注意力机制与Unet的细胞分割模型

LSKA-Unet: Large Separable Kernel Attention Mechanism and Unet Modeling Combined for Cell Segmentation

摘要: 医学图像中的细胞分割是生物医学研究和临床诊断的重要步骤。为了解决当前细胞分割方法存在细胞边界分割不连续、精度低等问题，提出了一种基于Unet的改进模型LSKA-Unet。首先，针对医学图像受噪声、伪影和其他干扰的影响导致图像质量低的问题，引入了大型可分离核注意力模块(Large Separable Kernel Attention, LSKA)来减少背景噪声对分割结果的干扰，通过自适应地学习空间注意力权重，强调重要特征，增强特征的表达能力；其次，使用Leaky ReLU激活函数替换ReLU激活函数，通过允许小的负输出，Leaky ReLU能保持一定的特征信息，有助于更好地捕捉医学图像中的细微变化；最后，使用BCE Loss (Binary Cross Entropy Loss)替换Dice Loss，针对细胞分割二分类这类问题，能够逐像素地评估细胞和背景的分类，从而提高模型在边界区域的准确性。实验结果表明，LSKA-Unet网络在细胞数据集Drosophila EM上，Precision、Dice系数和MIoU三大指标达到了94.82%、84.08%和84.92%，满足医学图像分割需求。

Abstract: Cell segmentation in medical images is an important step in biomedical research and clinical diagnosis. To solve the problems of discontinuous cell boundary segmentation and the low accuracy of current cell segmentation methods, an improved model LSKA-Unet based on Unet is proposed. Firstly, to address the problem that medical images are affected by noise, artifacts, and other interferences that lead to low image quality, a large detachable kernel attention module (LSKA) is introduced to reduce the interference of background noise on the segmentation results, and to emphasize the important features and enhance the expression of the features by adaptively learning the spatial attention weights; second, the Leaky ReLU activation function is used to replace the ReLU activation function, and by allowing a small negative output, the Leaky ReLU can maintain certain feature information, which helps to better capture subtle changes in medical images; finally, BCE Loss (Binary Cross Entropy Loss) is used to replace Dice Loss for such problems as cell segmentation binary classification, which can evaluate the classification of cells and background pixel by pixel, thus improving the accuracy of the model in the boundary region. The experimental results show that the LSKA-Unet network achieves 94.82%, 84.08%, and 84.92% in Precision, Dice coefficient, and MIoU on the cellular dataset Drosophila EM, which meets the requirements of medical image segmentation.