基于D-Unet神经网络的鼻腔鼻窦肿瘤分割算法

Nasal Cavity and Paranasal Sinuses Tumor Segmentation Algorithm Based on D-Unet Neural Network

鼻腔鼻窦肿瘤为多发性疾病,其CT影像具有形态不规则、分界不均匀等特征,而现有的U-Net分割算法对图片细节不敏感且割裂了图像局部与整体特征的一致性,难以实现精准诊断。提出一种基于D-Unet深度神经网络的改进算法,根据鼻腔鼻窦肿瘤空间形变特点,将可变形卷积融入U-Net网络,并利用可变形卷积能依据目标形态拥有自适应感受野的特点,充分学习图像细节,从而提升算法的特征提取能力。在此基础上,使用损失函数Tversky解决数据集样本失衡问题,从而获得更高的灵敏度和泛化能力。为方便进一步研究,建立鼻腔鼻窦肿瘤分割数据集。实验结果表明,所提算法能有效提高鼻腔鼻窦肿瘤分割精度,相比U-Net、Res-Unet和Attention U-Net算法,分割精度分别提高了5.01%、2.56%和0.48%。

Nasal cavity and paranasal sinus tumor are a type of multiple disease,and their CT images are characterized by irregular shape as well as uneven boundary.When dealing with these images,the existing U-Net segmentation algorithms are not sensitive to image details,and split the consistency of local and overall features of the picture,so the accuracy of diagnosis is reduced.To address the problem,this paper proposes an improved D-Unet based algorithm.Considering the spatial deformation of nasal cavity and paranasal sinus tumors,deformable convolutions are introduced into U-Net.The convolutions can adapt the receptive field to the target shape,fully learn image details,and enhance the feature extraction ability.On this basis,Tversky loss function is used to address imbalanced sample data sets,so higher sensitivity and generalization ability are obtained.In addition,a data set of nasal cavity and paranasal sinus tumor segmentation is established to facilitate further research.Experimental results show that the proposed algorithm can effectively improve the segmentation accuracy of nasal cavity and paranasal sinus tumors.The accuracy is 5.01%higher than U-Net,2.56%higher than Res-Unet and 0.48%higher than Attention U-Net.