基于混合密集注意力机制和NRAdam算法的图像分割

Dense Attention Network with NRAdam Algorithm for Image Segmentation

针对人工分割脑肿瘤耗时耗力的问题,提出一种基于U-Net的自动分割模型。设计了以U-Net模型为主体,Dense模块和注意力机制相结合的分割网络结构,使分割后的精度有了明显的提升。同时在RAdam优化算法的基础上加入Nesterov动量,即NRAdam优化算法,不仅加快收敛速度,还提高了模型的鲁棒性。将所提出的方法应用于MRI低级神经胶质瘤图像分割任务,得到Dice值和参数大小分别为0.9168和1.91M。实验结果表明,与原有的U-Net方法和其他的分割方法相比,改进后的方法在Dice值上得到了更好的结果,并通过减少参数量大小来实现模型轻量化。

In order to solve the time-consuming and labor-consuming of artificial segmentation,an automatic segmentation model based on U-Net is proposed.This method designs a segmentation network structure that takes the U-Net model as the main body,and combines the dense block and the attention mechanism,which significantly improves the accuracy after segmentation.Meanwhile,adding Nesterov momentum to the RAdam optimization algorithm is not only speed up the convergence speed,also make the model more robust.Applied the proposed method to LGG on MRI segmentation task,the Dice value and number of parameters were 0.9168 and 1.91 M respectively.The experimental results show that,compared with the U-Net and other segmentation methods,improved method obtains better results on the Dice value,and achieves our network more lightweight by reducing the number of parameters.