基于U-Net模型从低能锥形束CT图像实现高能成像的研究

目的基于U-Net网络深度学习的方法, 实现在放疗临床中低能锥形束CT(CBCT)图像转换成高能CBCT图像, 以期提供双能CBCT成像图像基础且降低辐射剂量。方法利用放疗机载CBCT设备采集CIRS电子密度模体和CIRS头部体模在80和140 kV能量下的CBCT图像数据, 数据集按10∶1分为训练集和测试集。利用U-Net网络从低能量(80 kV)CBCT图像预测高能量(140 kV)下CBCT图像。采用平均绝对误差(MAE)、结构相似度指数(SSIM)、信噪比(SNR)和峰值信号噪声比(PSNR)4种度量指标, 定量评价预测高能CBCT图像。结果预测高能图像与真实高能图像之间总体结构差异较小(SSIM:0.993 ±0.003)。预测高能图像噪声较低(SNR:15.33±4.06), 但组织间分辨力有损失。预测高能图像比真实高能图像平均CT值偏低, 在低密度组织中差异较小(<10 HU, P > 0.05), 而在高密度组织中差异大(< 21 HU, t = -7.92, P < 0.05)。结论利用深度学习方法可以从低能CBCT图像获得结构相似度高的高能...

Study on generation of high energy images from low energy CBCT images based on U-Net model

Objective To investigate the conversion of low-energy CBCT images into high-energy CBCT images in clinical radiotherapy based on the deep learning method of U-Net network, in order to provide dual-energy CBCT images and reduce radiation dose.Methods The CBCT image data of CIRS electron density phantom and CIRS head phantom at 80 and 140 kV were collected by the on-board CBCT in radiotherapy equipment. The dataset was divided into training set and test set according to 10:1. The U-Net network was used to predict CBCT images at high energy (140 kV) from low-energy (80 kV) CBCT images. Four parameters, including mean absolute error (MAE), structural similarity index (SSIM), signal-to-noise ratio (SNR) and peak signal-to-noise ratio (PSNR) were used to quantitatively evaluate predicted high-energy CBCT images.Results The overall structural difference between the predicted high-energy image and the real high-energy image was smaller (SSIM:0.993 ±0.003). The noise of predicted high-energy image was lower (SNR:15.33±4.06), but there was a loss of inter-tissue resolution. Predicted high-energy images had slightly lower average CT values than real high-energy images, with less difference in low-density tissues (<10 HU, P > 0.05) and greater differences in high-density tissues (<21 HU, t=-7.92, P < 0.05).Conclusions High-energy CBCT images with high structural similarity can be obtained from energy CBCT images by using deep learning method. The predicted high energy CBCT images have the potential to be applied to clinical dual-energy CBCT imaging technology in radiotherapy.