EnGAN:医学图像分割中的增强生成对抗网络

EnGAN: enhancement generative adversarial network in medical image segmentation

原始采集的医学图像普遍存在对比度不足、细节模糊以及噪声干扰等质量问题,使得现有医学图像分割技术的精度很难达到新的突破。针对医学图像数据增强技术进行研究,在不明显改变图像外观的前提下,通过添加特定的像素补偿和进行细微的图像调整来改善原始图像质量问题,从而提高图像分割准确率。首先,设计引入了一个新的优化器模块,以产生一个连续分布的空间作为迁移的目标域,该优化器模块接受数据集的标签作为输入,并将离散的标签数据映射到连续分布的医学图像中;其次,提出了一个基于对抗生成网络的EnGAN模型,并将优化器模块产生的迁移目标域用来指导对抗网络的目标生成,从而将改善的医学图像质量知识植入模型中实现图像增强。基于COVID-19数据集,实验中使用U-Net、U-Net+ResNet34、U-Net+Attn Res U-Net等卷积神经网络作为骨干网络,Dice系数和交并比分别达到了73.5%和69.3%、75.1%和70.5%,以及75.2%和70.3%。实验的结果表明,提出的医学图像质量增强技术在最大限度保留原始特征的条件下,有效地提高了分割的准确率,为后续的医学图像处理研究提供了一个更为稳健和高效的解决方案。

The quality issues commonly found in original medical images,such as insufficient contrast,blurred details,and noise interference,make it difficult for existing medical image segmentation techniques to achieve new breakthroughs.This study focused on the enhancement of medical image data.Without significantly altering the appearance of the image,it improved the quality problems of the original image by adding specific pixel compensation and making subtle image adjustments,thereby enhancing the accuracy of image segmentation.Firstly,it introduced a new optimizer module,which generated a continuous distribution space as the target domain for transfer.This optimizer module took the labels of the dataset as input and mapped the discrete label data to the continuous distribution of medical images.Secondly,it proposed an EnGAN model based on generative adversarial networks(GAN),and used the transfer target domain generated by the optimizer module to guide the target generation of the adversarial network,thereby implanting the knowledge of improving medical image quality into the model to achieve image enhancement.Based on the COVID-19 dataset,convolutional neural networks,including U-Net,U-Net+ResNet34,U-Net+Attn Res U-Net,were utilized as the backbone network in the experiment,and the Dice coefficient and intersection over union reached 73.5%and 69.3%,75.1%and 70.5%,and 75.2%and 70.3%respectively.The empirical results demonstrate that the proposed medical image quality enhancement technology effectively improves the accuracy of segmentation while retaining the original features to the greatest extent,providing a more robust and efficient solution for subsequent medical image processing research.