School of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea.
Med Phys. 2021 May;48(5):2185-2198. doi: 10.1002/mp.14701. Epub 2021 Mar 22.
Medical image analysis using deep neural networks has been actively studied. For accurate training of deep neural networks, the learning data should be sufficient and have good quality and generalized characteristics. However, in medical images, it is difficult to acquire sufficient patient data because of the difficulty of patient recruitment, the burden of annotation of lesions by experts, and the invasion of patients' privacy. In comparison, the medical images of healthy volunteers can be easily acquired. To resolve this data bias problem, the proposed method synthesizes brain tumor images from normal brain images.
Our method can synthesize a huge number of brain tumor multicontrast MR images from numerous healthy brain multicontrast MR images and various concentric circles. Because tumors have complex characteristics, the proposed method simplifies them into concentric circles that are easily controllable. Then, it converts the concentric circles into various realistic tumor masks through deep neural networks. The tumor masks are used to synthesize realistic brain tumor images from normal brain images.
We performed a qualitative and quantitative analysis to assess the usefulness of augmented data from the proposed method. Data augmentation by the proposed method provided significant improvements to tumor segmentation compared with other GAN-based methods. Intuitive experimental results are available online at https://github.com/KSH0660/BrainTumor.
The proposed method can control the grade tumor masks by the concentric circles, and synthesize realistic brain tumor multicontrast MR images. In terms of data augmentation, the proposed method can successfully synthesize brain tumor images that can be used to train tumor segmentation networks or other deep neural networks.
使用深度神经网络进行医学图像分析已得到广泛研究。为了对深度神经网络进行准确的训练，学习数据应具有足够的数量、良好的质量和广泛的特征。然而，在医学图像中，由于招募患者困难、专家对病变进行注释的负担以及侵犯患者隐私等原因，很难获取足够的患者数据。相比之下，健康志愿者的医学图像可以很容易地获取。为了解决这个数据偏差问题，提出了一种从正常脑图像合成脑肿瘤图像的方法。
我们的方法可以从大量的正常脑多对比度磁共振图像和各种同心圆中合成大量脑肿瘤多对比度磁共振图像。由于肿瘤具有复杂的特征，该方法将其简化为易于控制的同心圆。然后，通过深度神经网络将同心圆转换为各种现实的肿瘤掩模。利用肿瘤掩模从正常脑图像中合成逼真的脑肿瘤图像。
我们进行了定性和定量分析，以评估该方法生成的增强数据的有用性。与其他基于 GAN 的方法相比，该方法的数据增强显著提高了肿瘤分割的性能。直观的实验结果可在 https://github.com/KSH0660/BrainTumor 上查看。
该方法可以通过同心圆控制肿瘤掩模的等级，并合成逼真的脑肿瘤多对比度磁共振图像。在数据增强方面，该方法可以成功地合成可用于训练肿瘤分割网络或其他深度神经网络的脑肿瘤图像。
