基于MD-CGAN的脑部肿瘤图像生成方法研究

深度学习已广泛用于脑部磁共振（MR）图像分析中，但脑部肿瘤MR图像样本不足会严重影响深度学习模型的性能.提出基于多鉴别器循环一致性生成对抗网络（MD-CGAN）的样本生成方法 .利用所提出的MD-CGAN生成脑部肿瘤病理区域图像，将生成的脑部肿瘤病理区域图像覆盖脑部正常图像子区域，合成得到脑部肿瘤MR图像. MD-CG...

Research on Brain Tumor Image Generation Method Based on MD-CGAN

The samples are insufficient due to the difficulty of obtaining real brain tumor MR images, which se? riously affects the performance of deep learning models. Therefore, a sample generation method based on the Multiple Discriminator Cycle-consistent Generative Adversarial Network (MD-CGAN) is proposed in this paper. Firstly, the MD-CGAN model is used to generate brain tumor pathological region images, and then these pathological region im? ages are overlaid with the normal sub-regions of brain images to synthesize brain tumor MR images. Among them, the double adversarial loss introduced by MD-CGAN avoids the problem of model collapse, and the cycle consistency loss function introduced can ensure that the normal brain sub-region images generate the pathological region images of brain tumors, so that the images generated by MD-CGAN have high quality and diversity. Taking the Fréchet In? ception Distance(FID) as the evaluation index, the MD-CGAN proposed in this paper and the more classic generative networks in recent years are used to generate the images of brain tumor pathological regions and calculate the FID value. The experimental results show that the FID of our MD-CGAN is 26.43%, 21.91%, and 12.78% lower than those of SAGAN, StyleGAN, and StyleGAN2, respectively. To further demonstrate the effectiveness of our proposed method, we use the generated brain tumor images to expand the training set and then train the segmentation models on this expanded dataset. The experimental results show that the performance of segmentation networks trained on the expanded dataset is better. Based on the above experimental results, it can be concluded that the brain tumor MR images generated by our proposed method have high quality and rich diversity. These samples can be used to expand the training set and effectively solve the problem that brain tumor MR images are insufficient.