基于多尺度注意力的特征自适应聚合脑肿瘤图像分割

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，本模型还在Kaggle公开的脑肿瘤数据集上完成对比实验，证明本模型具有良好的可泛化性。

Brain tumor image segmentation using feature adaptive aggregation based on multi-scale attention

A multi-scale feature aggregation method based on multi-scale attention is proposed to address the problem of low recognition accuracy in automatic segmentation of stroke lesions in brain magnetic resonance images due to complex segmentation target edges and diverse scale changes. This method utilizes attention mechanisms to adjust the weights of different channels of intermediate features and adaptively selects features of different scales for fusion. A series of experiments conducted on the public dataset ATLAS for ischemic stroke and selected Dice coefficient, Hausdorff distance, overlap, accuracy and recall as evaluation indicators. The result demonstrate that the proposed model achieves leading segmentation performance in stroke lesion segmentation. Additionally, the model also conducted comparative experiments on the brain tumor dataset publicly available on Kaggle, proving its good generalization ability.