针对遥感图像建筑物的轮廓分割不完整、边界分割模糊和阴影干扰等导致的错误分割问题,提出一种基于VGG16的卷积块注意力深度可分离卷积U-Net网络(VGG16 Convolutional Block Attention and Deep Separable Convolution U-Net,VCDG-UNet...针对遥感图像建筑物的轮廓分割不完整、边界分割模糊和阴影干扰等导致的错误分割问题,提出一种基于VGG16的卷积块注意力深度可分离卷积U-Net网络(VGG16 Convolutional Block Attention and Deep Separable Convolution U-Net,VCDG-UNet)。为对建筑物特征进行提取,编码器部分模型以具有强大特征提取能力的VGG16作为骨干网络;解码器部分用深度可分离卷积代替普通卷积来减少参数量并融合不同尺度的特征;引入卷积块注意力模块(Convolutional Block Attention Module,CBAM)加入跳跃连接中,使其更有效地从不同尺度的图像中提取上下文信息并提高其对重要区域的关注度;为解决网络训练过程中的梯度消失问题,使用了高斯误差线性单元(Gaussian Error Linear Unit,GELU)。实验结果显示,改进后的网络在WHU和INRIA数据集上的平均交并比(mean Intersection over Union,mIoU)和F1-score分别达到了94.20%、96.83%和89.69%、94.51%,相较于基础模型高出了1.59%、0.76%和2.8%、1.59%。展开更多
Cone-beam computed tomography(CBCT) is mostly used for position verification during the treatment process. However,severe image artifacts in CBCT hinder its direct use in dose calculation and adaptive radiation therap...Cone-beam computed tomography(CBCT) is mostly used for position verification during the treatment process. However,severe image artifacts in CBCT hinder its direct use in dose calculation and adaptive radiation therapy re-planning for proton therapy. In this study, an improved U-Net neural network named CBAM-U-Net was proposed for CBCT noise reduction in proton therapy, which is a CBCT denoised U-Net network with convolutional block attention modules. The datasets contained 20 groups of head and neck images. The CT images were registered to CBCT images as ground truth. The original CBCT denoised U-Net network, sCTU-Net, was trained for model performance comparison. The synthetic CT(SCT) images generated by CBAM-U-Net and the original sCTU-Net are called CBAM-SCT and U-Net-SCT images, respectively. The HU accuracies of the CT, CBCT, and SCT images were compared using four metrics: mean absolute error(MAE), root mean square error(RMSE), peak signal-to-noise ratio(PSNR), and structure similarity index measure(SSIM). The mean values of the MAE, RMSE, PSNR, and SSIM of CBAM-SCT images were 23.80 HU, 64.63 HU, 52.27 dB, and 0.9919, respectively,which were superior to those of the U-Net-SCT images. To evaluate dosimetric accuracy, the range accuracy was compared for a single-energy proton beam. The γ-index pass rates of a 4 cm × 4 cm scanned field and simple plan were calculated to compare the effects of the noise reduction capabilities of the original U-Net and CBAM-U-Net on the dose calculation results. CBAM-U-Net reduced noise more effectively than sCTU-Net, particularly in high-density tissues. We proposed a CBAM-U-Net model for CBCT noise reduction in proton therapy. Owing to the excellent noise reduction capabilities of CBAM-U-Net, the proposed model provided relatively explicit information regarding patient tissues. Moreover, it maybe be used in dose calculation and adaptive treatment planning in the future.展开更多
文摘针对遥感图像建筑物的轮廓分割不完整、边界分割模糊和阴影干扰等导致的错误分割问题,提出一种基于VGG16的卷积块注意力深度可分离卷积U-Net网络(VGG16 Convolutional Block Attention and Deep Separable Convolution U-Net,VCDG-UNet)。为对建筑物特征进行提取,编码器部分模型以具有强大特征提取能力的VGG16作为骨干网络;解码器部分用深度可分离卷积代替普通卷积来减少参数量并融合不同尺度的特征;引入卷积块注意力模块(Convolutional Block Attention Module,CBAM)加入跳跃连接中,使其更有效地从不同尺度的图像中提取上下文信息并提高其对重要区域的关注度;为解决网络训练过程中的梯度消失问题,使用了高斯误差线性单元(Gaussian Error Linear Unit,GELU)。实验结果显示,改进后的网络在WHU和INRIA数据集上的平均交并比(mean Intersection over Union,mIoU)和F1-score分别达到了94.20%、96.83%和89.69%、94.51%,相较于基础模型高出了1.59%、0.76%和2.8%、1.59%。
基金Digital Medical Equipment Research and Development Project,Ministry of Science and Technology,China:The development of Synchrotron-based proton therapy system(2016YFC0105400).
文摘Cone-beam computed tomography(CBCT) is mostly used for position verification during the treatment process. However,severe image artifacts in CBCT hinder its direct use in dose calculation and adaptive radiation therapy re-planning for proton therapy. In this study, an improved U-Net neural network named CBAM-U-Net was proposed for CBCT noise reduction in proton therapy, which is a CBCT denoised U-Net network with convolutional block attention modules. The datasets contained 20 groups of head and neck images. The CT images were registered to CBCT images as ground truth. The original CBCT denoised U-Net network, sCTU-Net, was trained for model performance comparison. The synthetic CT(SCT) images generated by CBAM-U-Net and the original sCTU-Net are called CBAM-SCT and U-Net-SCT images, respectively. The HU accuracies of the CT, CBCT, and SCT images were compared using four metrics: mean absolute error(MAE), root mean square error(RMSE), peak signal-to-noise ratio(PSNR), and structure similarity index measure(SSIM). The mean values of the MAE, RMSE, PSNR, and SSIM of CBAM-SCT images were 23.80 HU, 64.63 HU, 52.27 dB, and 0.9919, respectively,which were superior to those of the U-Net-SCT images. To evaluate dosimetric accuracy, the range accuracy was compared for a single-energy proton beam. The γ-index pass rates of a 4 cm × 4 cm scanned field and simple plan were calculated to compare the effects of the noise reduction capabilities of the original U-Net and CBAM-U-Net on the dose calculation results. CBAM-U-Net reduced noise more effectively than sCTU-Net, particularly in high-density tissues. We proposed a CBAM-U-Net model for CBCT noise reduction in proton therapy. Owing to the excellent noise reduction capabilities of CBAM-U-Net, the proposed model provided relatively explicit information regarding patient tissues. Moreover, it maybe be used in dose calculation and adaptive treatment planning in the future.
