基于贝叶斯和cGAN的甲状腺结节生成方法

Bayesian and cGAN based thyroid nodule generation method

为克服小样本医学图像数据集对训练强大的超声图像中甲状腺结节的自动分割或分类的深度学习模型的限制,提出一种基于贝叶斯神经网络和条件生成对抗网络(cGAN)的数据生成方法。构建cGAN并以修改的结节掩膜作为条件,通过在真实样本上训练生成器生成具有不同特征的结节。使用贝叶斯神经网络识别出最有价值的生成样本,并用其构建新的数据集训练模型进行结节的分割和分类。实验结果表明,提出方法仅使用约35%的真实样本即可实现令人满意的分割和分类性能。

To overcome the limitations of small sample medical image datasets on training powerful deep learning model for automatic segmentation or classification of thyroid nodules in ultrasound images,a data generation method based on Bayesian neural network and conditional generative adversarial network(cGAN)was proposed.A cGAN was constructed and a modified nodule mask was used as the condition to generate nodule with different features by adjusting the generator on real samples.The Baye-sian neural network was used to identify the most valuable generated samples and they were used to build a new dataset to train the models for segmentation and classification of the nodules.Experimental results suggest that the proposed method can achieve satisfactory segmentation and classification performance using only about 35%of real samples.