Deep-Learning-Based Dual-Energy CT Imaging

X-ray computed tomography(CT)has the advantages of low cost,high speed and high spatial resolution,and is a routine screening tool for many diseases such as lung,abdomen,and cardiovascular diseases.However,conventional CT is associated with high radiation dose and low soft tissue resolution,which poses a significant safety risk to highly sensitive populations(e.g.,pregnant women,newborns,etc.)and imposes many constraints on its clinical application.Conventional single energy computed tomography(SECT)technology distinguishes different substances with their different absorption coefficients to the X-ray photons.Due to the presence of image noise,it is difficult to distinguish substances with similar absorption coefficients(e.g.,iodine and calcium)on SECT images.In contrast to conventional single energy CT,which uses only a single energy spectrum of X-rays,dual-energy CT makes use of two different X-ray energy spectra to scan the patient separately to obtain two tomographic images at the same location,one at low energy and the other at high energy.Using dual-energy material decomposition techniques,these two images can be further converted into atomic number and electron density at each pixel location or to the relative abundance of several substances,thus enabling differentiation between substances.Due to the limitations of existing dual-energy CT data acquisition devices in terms of acquisition modes,existing dual-energy CT imaging systems and reconstruction algorithms have to compromise between equipment cost,radiation dose,image noise,temporal resolution,and energy resolution.Therefore,we intend to develop a new dual-energy CT imaging protocol and use deep learning-based algorithms to target the difficulties in the corresponding imaging protocol,reduce the difficulty of dual-energy CT image acquisition,and achieve low-cost and high-quality dual-energy CT imaging.The main research contents of this dissertation include CT image mapping between different energy spectra,mapping between SECT images to virtual single-energy imaging,dual-energy CT imaging with polar coefficients in single-angle high-energy projection,and high-low energy slow switching dual-energy scanning mode.The main work and contributions are as follows,(1)A deep learning dual-energy CT mapping algorithm is proposed to realize the mapping of CT images between different energy spectra.The DL-DECT algorithm is proposed,which estimates the difference between the denoised high-and low-energy images by a denoising network Denoise Net and a dual-energy difference mapping network DEMap Net,and obtains the corresponding high-energy image by summing this difference with the input low-energy image.The high-energy images estimated in this way are closer to the real images in terms of noise and structure than those obtained by directly using a single mapping network.Also,it is found in the material decomposition experiments that the algorithm can obtain the base material image with lower noise level.(2)For the application of DECT on virtual monoenergetic imaging,two deep learning SECT virtual monoenergetic imaging algorithms are proposed based on DL-DECT algorithm to realize the mapping between SECT images to virtual monoenergetic images.A two-step virtual single-energy imaging algorithm based on deep learning is proposed,which further extends the DL-DECT algorithm and complements the subsequent virtual single-energy imaging algorithm on the basis of DL-DECT to achieve the mapping between single-energy CT images to virtual single-energy images by estimating dual-energy CT images from the traditional single-energy CT.Compared with the conventional virtual single-energy imaging algorithm,the method generates images with outstanding improvement in CNR.Also,the maximum CNR is shifted from 80-85 Ke V in conventional virtual monoenergetic images to 40 Ke V,and this transfer of maximum CNR is of great value at low energy levels where iodine contrast agents as well as human soft tissues have higher contrast.A one-step virtual monoenergetic imaging algorithm,DL-VMI,is also proposed to map direct monoenergetic CT images to virtual monoenergetic images using a convolutional neural network with U-Net structure,and the performance of this estimation network is further improved by adding depth supervision,2.5D input and SSIM loss function.Compared with the two-step virtual singleenergy imaging algorithm,this method is more straightforward in its thinking and no longer needs to introduce the step of estimating a dual-energy multicolor image from a single-energy multicolor image,which can more fully utilize the powerful feature learning capability of neural networks.(3)To address the problem of the lack of constraint effect of DL-DECT algorithm,the full low-energy and single high-energy DECT imaging algorithm is proposed to realize dual-energy CT imaging under low-energy CT images plus single-view extremely sparse high-energy information.The dual-energy CT imaging is achieved by directly estimating the high-energy image corresponding to the low-energy image from the low-energy image and the single-angle high-energy projection using the Full Low-Energy and Single High-energy DECT imaging algorithm(FLESH-DECT)algorithm.In FLESH-DECT,after image denoising,a set of possible "basis material" images are firstly decomposed from the low-energy images using the MD-CNN network,and the high-energy projection data are preprocessed using the PDCNN network to reduce the effect of data mismatch.Finally,the estimation of the high-energy images is achieved by least-squares matching between the "basis material" images and the corresponding single-angle high-energy projections.By incorporating redundant information and correlations into the deep learning model,this method has the potential to provide materialand energy-specific images using standard SECT equipment,which has the potential to alleviate the need for advanced DECT equipment.In addition,the use of sparse sampling at the second energy level can significantly reduce the radiation dose of DECT imaging compared to standard all-low and high-energy sampling DECT mechanisms.In contrast to the DL-DECT algorithm,the FLESH-DECT method uses the learned knowledge to adapt to the high-energy projections obtained from the actual measurements.The high-energy projections introduce a data fidelity term to constrain the projections generated from the predicted high-energy images to be consistent with the measurements,which further improves the accuracy of the predicted images.(4)A slow k Vp-switching dual-energy CT imaging algorithm is proposed to achieve high-quality dual-energy CT reconstruction in a scan protocol with slow high-and lowenergy switching.In this protocol,the trajectory of the projection source is only a complete circle,and the k Vp of the projection source alternates between low energy and high energy several times during the scanning process.A new reconstruction algorithm SKS-DECT is proposed,which combines iterative reconstruction and deep neural networks.The method first combines the interleaved high-and low-energy projections into a set of completed projection data,and reconstructs an image with relatively few artifacts from this data,then iteratively optimizes the obtained image using the high-and low-energy projection data separately,and finally fine-tunes the image using a neural network model to obtain a high-quality dual-energy image.Residual learning and attention mechanism are introduced in the network to further enhance the network learning effect.The influence of the interval of high-and low-energy switching on the final results is analyzed in the result analysis,and the index of sinogram coverage is proposed for program evaluation.The dual-energy CT imaging algorithm proposed in this paper still has significant limitations.This mainly contains two aspects.One is the limitation on the ability to generalize to different tissues,such as it is difficult for the network trained with head data to better achieve imaging on thoracic and abdominal data.Second,for different tissues that cannot be distinguished in the input single-energy images,the DL-DECT,DL-VMI,and FLESH-DECT algorithms proposed in this paper also cannot guarantee the accuracy of the output high-energy images.