Hybrid Segmentation Algorithm for Medical Image Segmentation Based on Generating Adversarial Networks, Mutual Information and Multi-Scale Information
Hybrid Segmentation Algorithm for Medical Image Segmentation Based on Generating Adversarial Networks, Mutual Information and Multi-Scale Information
复制标题
DOI：
10.1109/access.2020.3005384
复制
发表时间：
2020-01-01
期刊：
影响因子：
3.9
通讯作者：
Zhai, Zhaoyu
中科院分区：
文献类型：
--
作者：
Sun, Yi;Yuan, Peisen;Zhai, Zhaoyu
提出了3D-MedGAN、MLU-Net和Info-Max-Net模型，以克服医学图像分割中标注数据的不足，提取图像的多层次特征。3D-MedGAN旨在解决医学图像中缺乏标签数据的问题。它使用产生式对抗网络来模拟数据，然后从模型学习的分布中提取新生成的样本。将生成的样本与真实样本混合训练分割模型，可以有效地提高分割模型的效果。MLU-Net使用多层不同卷积角度从医学图像的多个角度提取特征信息。MLU-Net采用注意力机制对多层次特征信息进行融合，提高了特征表达和分割效果。Info-Max-Net是针对医学图像中的噪声问题而设计的。当图像中的信息比较复杂时，很难进行特征提取。利用互信息来度量图像与提取的特征之间的依赖关系，可以有效地降低图像中的噪声，提高分割效果。同时，为了解决图像的高维使得互信息难以度量的问题，本文使用了一个下界的BL-估计器来度量优化后的图像与提取出的特征之间的互信息。因此，该模型在逼近互信息真值的同时，仍能保持较快的收敛速度。考虑到3D-MedGAN生成的图像质量不如原始图像，我们将3D-MedGAN、MLU-Net和Info-Max-Net结合起来，以提高混合模型的敏感度和特征提取能力。通过在LIVER100数据集上对3D-MedGAN、MLU-Net、Info-Max-Net和混合模型的实验，验证了该模型的有效性。
This paper proposes 3D-MedGAN, MLU-Net and Info-Max-Net models for overcoming the lack of labeled data and extracting the multi-level feature of images in medical image segmentation. 3D-MedGAN is aimed at dealing with the lack of labeled data in medical images. It uses a generative adversarial network to simulate data and then draws newly generated samples from the distribution learned by the model. Training the segmentation model by mixing generated samples with real samples can effectively improve the effect of the segmentation model. MLU-Net uses multiple layers of different levels of convolutional angles to extract feature information from multiple angles in medical images. By adopting the attention mechanism to fuse the multi-level feature information, MLU-Net is able to improve the feature expressions and segmentation effect. Info-Max-Net is aimed at handling the noise problem in medical images. When the information in the images is complex, it is difficult to extract features. Using mutual information to measure the dependency between the image and the extracted features can effectively reduce noise in the image and improve the effect of segmentation. At the same time, for solving the problem that the high dimension of the image makes it difficult to measure mutual information, this paper uses a lower bound BL-estimator to measure the mutual information between the optimized image and the extracted features. Therefore, the model can maintain a high convergence speed as it approaches the true value of mutual information. Considering that the quality of images generated by 3D-MedGAN are not as good as the original images, we combine the 3D-MedGAN, MLU-Net, and Info-Max-Net to improve the sensitivity and the power of feature extraction of the hybird model. The effectiveness of our model is verified through experiments of the 3D-MedGAN, MLU-Net, Info-Max-Net, and the hybrid model over the LIVER100 dataset.