基于多尺度注意力的特征自适应聚合脑肿瘤图像分割

Brain Tumor Image Segmentation Using Feature Adaptive Aggregation Based on Multi-Scale Attention

针对脑部磁共振图像中脑卒中病灶的自动分割因分割目标边缘复杂、尺度变化多样而造成的识别精度不高的问题,提出一种基于多尺度注意力的多尺度特征聚合方法,该方法利用注意力机制调节中间特征不同通道的权重,并自适应地选择不同尺度的特征进行融合,在缺血性脑卒中的公开数据集ATLAS上进行的一系列实验,选取Dice系数、豪斯多夫距离、重叠度、准确率和召回率作为评价指标,结果表明所提出的模型在脑卒中病变的分割问题上取得了较好的分割效果;另外,该模型还在Kaggle公开的脑肿瘤数据集上完成对比实验,证明该模型具有良好的可泛化性。

Aimed at the low recognition accuracy for complex segmentation target edges and diverse scale changes in the automatic segmentation of stroke lesions in brain magnetic resonance images,a multi-scale feature aggregation method based on multi-scale at-tention is proposed.This method utilizes attention mechanism to adjust the weights of different channels for intermediate features and adaptively selects the features of different scales for fusion.A series of experiments are conducted on the public dataset ATLAS for is-chemic stroke,and the Dice coefficient,Hausdorff distance,overlap,accuracy and recall are selected as evaluation indicators.The experimental results show that the proposed model achieves a better segmentation performance in stroke lesion segmentation.Addi-tionally,the comparative experiments of the model are also conducted on the publicly available brain tumor dataset Kaggle,proving that it has a good generalization ability.