用于腹部CT肝脏肿瘤分割的注意力引导模型

Attention-guided model for abdominal computed tomography liver tumor segmentation

针对肝脏肿瘤在腹部CT影像中占比低,人为分割与传统分割效果、性能差的问题,本文提出一种高效的两阶段注意力引导的肝脏肿瘤分割模型,该模型由肝脏器官分割模块和肝脏肿瘤分割模块构成。肝脏器官分割模块中凭借卷积神经网络得到肝脏器官分割图像作为输出,和原始CT影像叠加得到新的输入影像,最后将新输入图像导入肝脏肿瘤分割模块得到精确的肝脏肿瘤分割影像,给出了每个模块的损失函数以监督训练。利用肝脏肿瘤分割挑战提供的数据集进行了实验研究和定性比较分析。结果在Dice系数等指标上优于其他模型,提高了肝脏肿瘤分割准确率和肝脏肿瘤区域定位准确率,从而使分割出的肝脏肿瘤更接近真实边界。

Owing to the low frequency of liver tumors in abdominal computed tomography(CT)images and the limitations of artificial and traditional segmentation methods,this study presents a high-efficiency two-stage attention-guided liver tumor segmentation model.This model consists of a liver organ segmentation module and a liver tumor segmentation module.In the first module,a convolutional neural network generates a segmented image of the liver organ,which is superimposed with the original CT image to obtain a new input image.This image is subsequently imported into the liver tumor segmentation module to obtain accurate liver tumor segmentation images.Each module incorporates a specific loss function for supervised training.This study uses the data set from a liver tumor segmentation challenge,conducting experimental research and qualitative comparative analysis.Our model outperforms existing models across several metrics,including the Dice coefficient.The improvements in liver tumor segmentation accuracy and the precision of tumor localization result in segmented images that more closely align with the real boundaries of liver tumors.