结合空洞特征增强与改进HR-Net的胰腺分割方法
Pancreas Segmentation Method Combining Atrous Feature Enhancement and Improved HR-Net

在胰腺癌辅助诊断领域，精确的胰腺分割是实现高效疾病诊断的基石，同时也是医学图像处理领域中一项具有挑战性的任务。现有的主流胰腺分割方法大多依赖于U型的编码–解码路径，然而该架构在上下采样的过程中，容易导致关键空间信息的丢失，影响分割准确率。针对上述问题，本文提出了一种结合空洞特征增强与改进HR-Net的胰腺分割方法。首先，对原有的二维HR-Net进行三维化改造，以便更全面地挖掘三维医学图像内蕴的空间信息。其次，设计多尺度空洞特征增强模块以重构网络深层结构，并行捕获广泛的多尺度深度特征，并采用通道加权对重要特征赋予更高的关注度。接着，在模型多分辨率融合模块之后引入空间注意力机制，使模型更加专注于融合特征图中的重要空间区域。最后，本文提出分层渐进特征融合头，以缓解传统分割头在大幅度上采样时遭受的信息丢失。实验结果表明，在NIH- Pancreas数据集上测试的召回率、准确率和Dice相似系数分为84.5%、86.3%、85.6%，在分割性能上优于现有的主流方法。本文方法能有效保留胰腺细节特征，并展现出较强的特征提取能力，对腹部CT图像中的胰腺器官有较好的分割结果。
In the field of pancreatic cancer auxiliary diagnosis, accurate pancreatic segmentation is the cornerstone for efficient disease diagnosis and also a challenging task in the field of medical image processing. The existing mainstream pancreatic segmentation methods mostly rely on a U-shaped encoding-decoding path; however, this architecture tends to lose critical spatial information during the upsampling and downsampling processes, affecting segmentation accuracy. To address these issues, this paper proposes a pancreatic segmentation method that combines atrous feature enhancement with an improved HR-Net. Firstly, the original two-dimensional HR-Net is transformed into three dimensions to more comprehensively mine the spatial information inherent in three-dimensional medical images. Secondly, a multi-scale atrous feature enhancement module is designed to reconstruct the deep structure of the network, capturing a wide range of multi-scale deep features in parallel and assigning greater attention to important features through channel weighting. Subsequently, a spatial attention mechanism is introduced after the model's multi-resolution fusion module, making the model more focused on important spatial regions in the merged feature maps. Lastly, this paper proposes a hierarchical progressive feature fusion head to alleviate the information loss suffered by traditional segmentation heads during significant upsampling. Experimental results show that on the NIH- Pancreas dataset, the recall rate, accuracy, and Dice similarity coefficient are 84.5%, 86.3%, and 85.6%, respectively, surpassing existing mainstream methods in segmentation performance. The method proposed in this paper can effectively preserve pancreatic detail features and demonstrates strong feature extraction capability, yielding better segmentation results for pancreatic organs in abdominal CT images.