基于特定损失函数GAN的高质量CT图像预测

High Quality CT Image Prediction by Using Generative Adversarial Network Based on Specific Loss Function

高放射剂量CT增加了患癌的风险,因此,针对治疗计划系统研究了超低剂量CT,但其含有的噪声造成诊断结果有较高的假阳性。针对上述问题,提出一种基于超低剂量CT预测高剂量CT的方法,并通过比较多种不同低剂量CT图像预测的高剂量CT图像,评估如何在保证图像质量的同时使放射剂量尽可能低。首先,基于改进的生成对抗网络模型提出一种特定损失函数,该损失函数可以有效地优化图像质量。然后,分别用10 mA·s,25 mA·s,100 mA·s和150 mA·s的低剂量CT图像数据训练不同的模型,最终在超低剂量CT中预测出令人满意的高剂量CT图像。为了验证算法的有效性,对预测结果进行主、客观评价并与存在的算法进行比较。实验结果表明,该模型可以基于10 mA·s的超低剂量预测高剂量图像的同时保证图像质量。

High-dose computed tomography(HDCT)leads to increased risk of radiation-related diseases such as cancer.Thus,extra-low-dose computed tomography(eLDCT)is researched for therapy planning system(TPS).The diagnosis always gives a relatively high false positive(FP)because of the noise.In order to address these problems,we propose a method for prediction of HDCT by using eLDCT images.By comparing the predicted high-dose CT images of different low-dose CT images,we evaluate how to ensure the image quality and make the radiation dose as low as possible.Firstly,a specific loss function is proposed based on the improved generative adversarial network model,which can effectively optimize image quality.Then,we use the low-dose CT image data of 10 mA·s,25 mA·s,100 mA·s and 150 mA·s to train different models respectively.In this way,satisfactory high-dose CT images are predicted in eLDCT images.To evaluate the effectiveness of our algorithm,the prediction results are subjectively and objectively evaluated and compared with different algorithms.The experiment results demonstrate that our proposed model can predict HDCT images based on eLDCT of 10 mA·s while ensuring image quality.