基于多尺度特征信息的脑肿瘤MRI图像分割网络

Brain tumor MRI image segmentation network based on multi-scale feature information

针对脑肿瘤核磁共振成像因大脑组织边界重叠和图像噪声干扰导致分割精度低的问题,提出一种基于多尺度特征信息的脑肿瘤分割模型.该模型将注意力机制等最新技术引入2D U-Net网络,通过独特的信息融合及由Transformer和卷积神经网络并行结构组成的双分支模块,提取全局和局部区域的多尺度信息特征,以突出肿瘤区域的病变信息.并用标准的Figshare脑肿瘤数据集评估此模型.实验结果表明,该模型在Dice分数、平均Jaccard系数、Precision和Recall上分别提高了3.01%、2.6%、3.08%和4.73%,HD95降低了0.1187,评估指标性能高于现有先进方法.同时,消融实验表明,信息融合模块和双分支模块有助于提高现有脑肿瘤磁共振成像的分割精度.

To address the issue of low segmentation accuracy in brain tumor magnetic resonance imaging(MRI)due to overlapping brain tissue boundaries and image noise interference,a brain tumor segmentation model based on multi-scale feature information is proposed.This model incorporates the latest technology such as the attention mechanism into a 2D U-Net network.It utilizes a unique information fusion approach and a dual-branch module composed of Transformer and convolutional neural network parallel structures to extract multi-scale information features from both global and local regions,thereby highlighting the pathological information in tumor areas.The model was evaluated based on the standard Figshare brain tumor dataset.The experimental results show improvements in Dice score,average Jaccard index,Precision,and Recall by 3.01%,2.6%,3.08%and 4.73%,respectively,while the HD95 metric decreased by 0.1187.These evaluation metrics outperform existing state-of-the-art methods.Additionally,ablation experiments demonstrate that both the information fusion module and the dual-branch module contribute to enhancing the accuracy of existing brain tumor MRI segmentation.