融合U-Net 3+和ACM的医学图像分割模型

Medical Images Segmentation Model Combining U-Net 3+ and Active Contour Model

活动轮廓模型方法不需要训练集且能较好利用边缘信息,但对初始轮廓较敏感,在处理复杂背景图像时分割不够精确。U-Net 3+网络可以分割更复杂的医学图像,但需要大量的人工标记,且模型的特征提取机制导致其在非典型边界特征的决策时通常是不准确的。因此,针对训练集较小的医学图像,提出了一种融合卷积神经网络和活动轮廓模型的医学图像自动分割模型。模型通过U-Net 3+网络获得目标先验信息,使用先验信息构造拟合能量项,并融合到活动轮廓模型中约束曲线演化。在皮肤镜病变和胸部X光片图像上测试,该模型的分割精度高于单独使用U-Net 3+网络和活动轮廓模型的分割结果。

The active contour model can use edge information without a training set. However, it is sensitive to the initial contour, and the segmentation is not accurate enough when dealing with complex background images. The U-Net 3+network can segment complex medical images, but required numerous manual labeling, and the unique feature extraction mechanism of the model leads to ambiguity in the decision-making of atypical boundary features. Therefore, for small medical images in the training set, a medical image automatic segmentation model merging with convolutional neural network and active contour model are proposed in this paper. This model obtains the prior information to construct the fitting energy term through the U-Net 3+network, and then integrates it into the constraint curve evolution. Tested on the dermoscopy lesions and chest X-ray images, the segmentation accuracy of the proposed model is higher than that of using the U-net 3+network and active contour model alone.