Research on Image Segmentation Model of Pancreatic Lesions

胰腺病变分割是医学图像分析中的一项重要任务，但由于胰腺的大小、形状不规则以及位置不固定，导致该任务充满挑战。胰腺病变的早期检测对于患者的治疗和预后至关重要，因此受到了广泛关注。本文提出了一种改进的U-Net网络模型，通过引入MedSAM编码器和高效通道注意力(ECA)模块，显著提升了胰腺病变分割的精度和鲁棒性。MedSAM编码器对输入图像进行预处理，提取多层次、多尺度的特征。MedSAM编码器能够有效捕捉胰腺的复杂结构和病变区域的细节信息，为后续分割提供高质量的特征表达。其次，在上采样过程中引入高效通道注意力(ECA)模块，增强通道间的依赖关系，防止降维导致的通道信息丢失。在MSD_Pancreas数据集上的实验结果表明，本文方法在Dice相似系数(DSC)和豪斯多夫距离(HD95)两个评价指标上均取得了显著提升。具体而言，本文方法的DSC分数达到82.7%，HD95值为10.2 mm，相比基线U-Net分别提高了4.4%和2.3 mm。实验结果表明，本文方法在胰腺病变分割任务中显著优于基线模型，尤其是在处理复杂病变和边界模糊区域时表现突出。Pancreatic lesion segmentation is an important task in medical image analysis, but it is challenging due to the irregular size and shape of the pancreas as well as its location. Early detection of pancreatic lesions is crucial for patient treatment and prognosis, and thus has received much attention. In this paper, we propose an improved U-Net network model that significantly improves the accuracy and robustness of pancreatic lesion segmentation by introducing the MedSAM encoder and the Efficient Channel Attention (ECA) module. The MedSAM encoder preprocesses the input image and extracts multi-level and multi-scale features. The MedSAM encoder is able to efficiently capture the complex structure of the pancreas and the lesion region’s detail information, providing high-quality feature representation for subsequent segmentation. Second, the Efficient Channel Attention (ECA) module is introduced in the up-sampling process to enhance the inter-channel dependencies and prevent the loss of channel information due to dimensionality reduction. The experimental results on the MSD_Pancreas dataset show that this paper’s method achieves significant improvement in both the evaluation metrics of Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD95). Specifically, the DSC score of this paper’s method reaches 82.7% and the HD95 value is 10.2 mm, which are improved by 4.4% and 2.3 mm, respectively, compared with the baseline U-Net. The experimental results show that this paper’s method significantly outperforms the baseline model in the task of pancreatic lesion segmentation, especially when dealing with complex lesions and regions with fuzzy boundaries.