基于轴向反向注意力机制的BS-Net脑瘤分割算法  

BS-Net Brain Tumor Segmentation Algorithm Based on Axial Reverse Attention Mechanism

作　　者：唐明[1] 张文静 白俊卿 吴晨俣 徐东 TANG Ming;ZHANG Wenjing;BAI Junqing;WU Chenyu;XU Dong(Xi'an Peihua University,Xi'an 710125;Xi'an Shiyou University,Xi'an 710065)

机构地区：[1]西安培华学院,西安710125 [2]西安石油大学,西安710065

出　　处：《计算机与数字工程》2025年第3期648-651,724,共5页Computer & Digital Engineering

摘　　要：针对核磁共振图像中脑瘤与周围组织对比不明显,图像中病灶占比较低导致的脑瘤误检漏检等问题,论文提出一种基于轴向反向注意力机制的BS-Net脑瘤分割算法。首先采用具有多尺度残差单元的Res2Net提取图像的全局特征,加强对脑瘤大小多样的关注。同时通过特征金字塔融合不同感受野的特征图,获得小目标脑瘤丰富的语义信息。其次使用轴向反向注意力模块,获取含有较多病灶空间位置与语义信息的特征信息,细化脑瘤病灶区域的边界。最后BS-Net网络在BraTS 2018数据集进行训练得到脑瘤图像分割模型。论文提出的模型从客观评价指标和视觉分割效果进行对比,实验表明该网络对小目标脑瘤分割效果较优,对脑瘤病灶分割的形状边缘更接近生物学,在脑瘤的临床应用中有重要意义。Aiming at the problems of brain tumor and surrounding tissue in the MRI image is not obvious,and the proportion of lesions in the image is relatively low,resulting in missed detection of cerebellar tumors,this paper proposes a BS-Net brain tumor segmentation algorithm based on axial reverse attention mechanism.Firstly,Res2Net with multi-scale residual units is used to extract the global characteristics of the image,which strengthens the attention to the diversity of brain tumor sizes.At the same time,the feature maps of different receptor fields are fused through the feature pyramid to obtain rich semantic information of small target brain tumors.Secondly,the axial reverse attention module is used to obtain the characteristic information containing more spatial location and semantic information of the lesion,and refine the boundary of the brain tumor lesion area.Finally,the BS-Net network is trained on the BraTS 2018 dataset to obtain a brain tumor image segmentation model.The model proposed in this paper compares the objective evaluation index and the visual segmentation effect.Experiments show that the network has a better effect on the detection of small targets,and the shape edges of the brain tumor lesion segmentation are closer to biology,which is of great significance in the clinical application of brain tumors.

关 键 词：轴向反向注意力机制 Res2Net 特征金字塔 脑瘤图像分割 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]