CT图像层间插值方法的深入分析验证

Research on Interslice Interpolation of CT Images

目的:通过CT图像层间插值的方法提高图像质量,从而保证放疗中患者摆位验证的精度。方法:采用基于深度卷积神经网络(DCNN)算法,利用图像层间的关联信息重建中间层图像。应用卷积层、膨胀卷积层、池化层和上采样层交织的端到端的神经网络对CT图像进行处理。由于头颈部和胸腹盆部的结构差异,分别采用70例头颈部和75例胸腹盆部患者的图像数据进行训练,建立不同的头颈部和胸腹盆部模型。采用留一交叉验证的方法验证模型,将DCNN与线性插值和生成对抗网络(GAN)的性能进行比较。结果:头颈部患者中,DCNN的平均绝对误差(MAE)为36 HU,优于线性插值和GAN的62 HU和50 HU。除此之外,骨骼的Dice相似系数(DSC)为0.95,超过线性插值和GAN的0.88和0.91;胸腹盆部位患者中,DCNN的MAE为29 HU,优于线性插值和GAN的47 HU和34 HU。除此之外,骨骼的DSC为0.94,同样超过线性插值和GAN的0.87和0.91。结论:与线性插值和GAN相比,DCNN算法在重建薄层CT图像方面具有更高的精度;其显著减少了插值伪影、图像失真和锯齿效应,从而提高了放疗患者摆位验证的精度。

Purpose:To improve the quality of CT images through interslice interpolation,thereby ensuring accurate patient setup verification in radiotherapy.Methods:Using an algorithm based on a deep convolutional neural network(DCNN)to reconstruct intermediate CT slice images by utilizing the correlation between image layers.An end-to-end neural network comprising convolutional,dilated convolutional,pooling,and upsampling layers was utilized to process CT images.Distinct models were developed for head-neck and chest-abdomen-pelvis regions due to their structural differences,and 70 head-neck and 75 chest-abdomen-pelvis datas were used for training,respectively.The model was tested using the method of leave-one-out cross validation.The performance of the DCNN was compared against linear interpolation and generative adversarial networks(GAN).Results:In headneck patients,the mean absolute error(MAE)of DCNN was 36 HU,outperforming linear interpolation(62 HU)and GAN(50HU).Additionally,the Dice similarity coefficient(DSC)of bone was 0.95,surpassing both linear interpolation(0.88)and GAN(0.91).For chest-abdomen-pelvis patients,the DCNN's MAE was 29 HU,outperforming linear interpolation(47 HU)and GAN(34 HU).DSC of bone was 0.94,also exceeding linear interpolation(0.87)and GAN(0.91).Conclusion:The DCNN algorithm demonstrates superior accuracy in reconstructing thin-slice CT images compared to linear interpolation and GAN methods.It significantly reduces interpolation artifacts,image distortion,and jagged effects,thus enhancing the precision of radiotherapy patient setup verification.