融合注意力机制的残差双通道多尺度的胰腺囊性肿瘤分割模型

Pancreatic Cystic Tumor Segmentation with Fusion Attention Mechanism and Residual Dual-Channel Multi-Scale Approach

摘要: 本文针对胰腺肿瘤的大小形状差异大、区域边界不清晰等问题，提出了一种融合注意力机制的残差双通道多尺度的胰腺囊性肿瘤分割模型ARDM-Net (Attention Residual Dual Channel and Multi Scale-UNet)。首先，对长海医院提供的数据集进行肿瘤区域的裁剪；其次，在U-Net网络模型上，将基础的3 × 3卷积模块替换为残差双通道多尺度卷积模块，增强网络特征提取的能力；最后在跳跃连接中加入注意力模块，调整学习特征的权重。本文的方法性能表现，Dice相似系数为89.50%，豪斯多夫距离(Hausdorff Distance, HD)为2.80 mm，交并比(Intersection-over-Union, IoU)为82.46%，均优于普通的U-Net网络。该试验结果充分表现了本文方法在胰腺囊性肿瘤分割任务中的显著价值。

Abstract: This paper proposes a fusion Attention Mechanism and Residual Dual-Channel Multi-Scale model (ARDM-Net) for the segmentation of pancreatic cystic tumors, addressing the challenges of significant differences in size and shape, as well as unclear regional boundaries. Initially, tumor regions were cropped from the dataset provided by Changhai Hospital. Subsequently, the basic 3x3 convolution module in the U-Net network model was replaced with a Dual-Channel and multi-scale convolution module to enhance feature extraction capability. Finally, an attention mechanism module was incorporated into the skip connection to adjust the weight of learning features. The proposed method in this paper achieves a performance with the Dice similarity coefficient of 89.50%, Hausdorff Distance (HD) of 2.80 mm, and Intersection-over-Union (IoU) of 82.46%, all of which outperformed the standard U-Net network. These experimental results demonstrate the significant value of our approach in the segmentation of pancreatic cystic tumors.