Model-Data Coupling Driven Spectral CT Material Decomposition Method

Compared with traditional CT,Spectral CT can achieve the quantitatively analysis of materials through the basis material decomposition using the X-ray spectral characteristics and the material attenuation-energy correlation characteristics of the scanning target.Nevertheless,Spectral CT has some limitations,such as high ionizing radiation in Dual Energy CT increase risk of cancer.At present,low milliampere second scanning technology can directly reduce dose radiation.However,the decrease of dose will result in a reduction in the number of photons measured by the detector and intensification of the quantum noise interference,further affecting the accuracy of Spectral CT and basis material decomposition.How to obtain high-quality Spectral CT basis material images at low radiation dose is a hotspot issue in the field.At present,the basis material decomposition methods are divided into two types:model-driven and data-driven methods.The model-driven methods construct the linear attenuation coefficient of the scanned object as a linear combination of the mass attenuation coefficient and its density,which can be decomposed quickly.However,owing to the material decomposition matrix is ill-posed,the quantitative images are sensitive to noise;The iterative decomposition model embedded with prior knowledge can effectively improve the accuracy of decomposition,but the task of prior knowledge design is challenging.The data-driven methods can improve the image quality of the basis materials by constructing an end-to-end deep network model to learn the potential feature mapping between the attenuation coefficient of the materials and the quantitative labels;However,these methods are insufficient in mining the intrinsic imaging mechanism of basis material decomposition and depends on plenties of paired training samples.Regarding the issue above,we propose two intelligent quantitative estimation algorithms for low-dose energy spectrum CT materials:(1)Aiming at the limitations of the data-driven algorithm’s inherent imaging mechanism mining,we propose a model-data-based adaptive material decomposition network(MDAMD-Net)driven by data-model coupling.Through coupling iterative optimization algorithm and deep learning technology,the prior knowledge of basis material images and decomposition coefficients optimization of adaptive learning of deep network are incorporated into the iterative optimization framework of gradient descent,and the decomposition coefficients and basis material images are optimized and updated synchronously with two sub-modules.It can solve the problem of decomposition coefficients accuracy and the design of prior knowledge in the iterative decomposition process and achieve robust energy spectrum CT accurate basis material imaging.Experiments on a large number of clinical patient simulation data have proved the effectiveness of the model.The quantitative results of basis materials have excellent performance in qualitative and quantitative evaluation.(2)Aiming at the over-dependence of the real data-driven method on the paired label data,we propose a Weakly supervised learning model of Spectral CT material decomposition(WSLMD-Net)via prior perception.The method includes a supervised sub-module and a self-supervised sub-module.In the supervised sub-module,the mapping relationship between the low SNR basis images and the high SNR basis images is constructed by a small amount of label data through the mean square error loss function.In the self-supervised sub-module,the image restoration loss function is constructed based on the consistency information of a large number of unlabeled image data and the total variation prior characteristics of different basis materials.It is weighted and integrated into the deep network training.It can effectively suppress the strong noise artifacts generated in the decomposition process.Experiments on clinical patient data indicate that the method can improve the signal-to-noise ratio of the basis material images effectively and maintain the important feature structure.Moreover,the unlabeled data characteristics can be fully used to improve the generalization performance of the model,effectively avoiding the problem of high demand for matching samples for training.