Single-energy CT energy spectrum imaging method based on deep learning
Technical Field
The invention relates to the technical field of medical engineering, in particular to a single-energy CT energy spectrum imaging method based on deep learning.
Background
At present, modern X-ray CT imaging is widely applied to medical clinical diagnosis and treatment, and has great social value and significance. In recent years, with the continuous emergence of new modalities of various CT imaging, spectral CT and its material decomposition characteristics show great clinical potential. At present, the energy spectrum CT imaging technology is widely applied to clinical diagnosis such as pathological organ contour delineation, virtual single-energy imaging, virtual non-enhanced imaging, hepatic fibrosis quantification, gout diagnosis, artery CT angiography imaging and the like.
The current energy spectrum CT imaging technology mainly comprises multiple scanning, a fast kVp switching technology, a double-layer plate detector technology, a photon counting detector technology and a double-source CT imaging technology. Multiple scans are easy to implement but increase the radiation dose and scan time. Fast kVp switching techniques can generate dual energy data in a single scan, but implementation of such schemes is challenging due to the difficulty in controlling the exposure at different tube voltages. The dual-layer detector scheme provides two overlapping detectors with different energy responses to incident X-ray photons to generate dual-energy projection data. This approach may generate consistent dual-energy projection data in a single scan, with the disadvantage that there is not enough energy reserve between the acquired high-energy and low-energy projections. Photon counting detectors can distinguish more of the incident X-ray energy spectrum, but still suffer from various limitations, including detector charge sharing, pulse packing, limited detection efficiency, and the like. Dual source CT imaging techniques mount two X-ray tubes nearly perpendicular to each other to generate dual energy data in one scan. The scheme can flexibly adjust tube voltage to obtain high-quality CT images, but is limited by cross scattering of photons generated by two bulbs, and in addition, two X-ray tube imaging schemes cannot obtain accurate and consistent dual-energy projection. The defects of the energy spectrum CT imaging technologies are that the hardware implementation difficulty is high, the reconstruction algorithm is relatively complex, and the radiation dose to a patient is increased.
Recently, deep learning has attracted a wide range of attention in the fields of machine learning and computer vision. Deep learning can efficiently learn high-level features from pixel-level data through a multi-layer hidden layer framework. Several network architectures proposed at present have been applied in the field of medical images, such as image restoration, image noise reduction, image super-high resolution improvement, image segmentation, organ classification, cell edge detection, and the like.
Inspired by the great potential of deep learning in the aspect of image processing, a single-energy CT energy spectrum imaging method based on deep learning is provided. The method can obtain energy spectrum CT data by single-energy CT by utilizing the mapping relation between CT images under different energies.
Disclosure of Invention
The invention aims to provide a single-energy CT energy spectrum imaging method based on deep learning. The method can realize the rapid, accurate and robust acquisition of the high-energy CT image by using the single-energy CT image, and meets the clinical requirements.
In order to achieve the purpose, the invention provides the following technical scheme:
a method of single energy CT spectral imaging based on deep learning, the method comprising the steps of:
taking the single-energy CT image and the label high-energy CT image as a training sample to form a training set;
constructing an image conversion model for predicting and outputting a high-energy CT image according to an input single-energy CT image based on a deep learning network, and training the image conversion model by utilizing a training set to determine model parameters of the image conversion model;
and inputting the single-energy CT image to be imaged into an image conversion model determined by the model parameters, and calculating to output a high-energy CT image.
Compared with the prior art, the invention has the beneficial effects that:
the image conversion model obtained through training can obtain a converted high-energy CT image based on the single-energy CT image, and the high-energy CT image has the image quality close to that of a real high-energy CT image and can be used for clinical diagnosis.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a flowchart of a method for single-energy CT spectral imaging based on deep learning according to an embodiment of the present invention;
fig. 2 is a related picture of high-energy CT image imaging performed by using an image transformation model constructed by a supervised learning network according to an embodiment of the present invention, where (a) is a single-energy CT image, (b) is a real high-energy CT image, and (c) is a predicted high-energy CT image;
fig. 3 is related pictures of high-energy CT image imaging performed by using an image transformation model constructed by an unsupervised learning network according to an embodiment of the present invention, where (a) is a single-energy CT image, (b) is a real high-energy CT image, and (c) is a predicted high-energy CT image.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the detailed description and specific examples, while indicating the scope of the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention.
Fig. 1 is a flowchart of a method for single-energy CT spectral imaging based on deep learning according to an embodiment of the present invention. Referring to fig. 1, the single energy CT energy spectrum imaging method includes the following steps:
s101, taking the single-energy CT image and the label high-energy CT image as a training sample to form a training set.
In the embodiment, high-energy CT images obtained from Siemens SOMATOM Force X-ray computed tomography equipment are used as a data set, wherein a CT image with one energy is used as a single-energy CT image, and CT images with the rest energy are used as label high-energy CT images. And dividing the data set into a training set and a verification set, wherein the training set is used for optimizing model parameters of the image conversion model, and the verification set is used for verifying the prediction effectiveness degree of the image conversion model.
S102, an image conversion model used for predicting and outputting a high-energy CT image according to an input single-energy CT image is built on the basis of a deep learning network, and the image conversion model is trained by utilizing a training set to determine model parameters of the image conversion model.
In this embodiment, the image transformation model may be obtained by optimizing network parameters using a supervised learning network. Specifically, the image conversion model takes a deep convolutional neural network as a network basis, a training set is used for training network parameters of the deep convolutional neural network, a joint residual learning framework is adopted to learn a residual compensation image between an input single-energy CT image and a labeled high-energy CT image in the training process, and the mapping function of the residual learning framework is as follows:
fspectral=M(fsingle)+fsingle
wherein f isspectralIs a high-energy CT image, fsingleIs a monoenergetic CT image, M (-) is an end-to-end mapping function used to train residual compensation images, M (f)single) Is a residual compensated image.
During training, the minimization loss function is:
wherein N is the total number of training samples in the training set, f
spectralIs a high-energy CT image, f
singleIs a single-energy CT image and can be used,
is the optimized mapping function result;
and solving the minimized loss function to obtain an optimized mapping function result.
In this embodiment, the image conversion model may also be obtained by optimizing network parameters using an unsupervised learning network. Specifically, the image conversion model is based on a cycle generation countermeasure network as a network basis, and the cycle generation countermeasure network comprises a generator and two discriminators which form two cycle networks and compete with each other until convergence to an equilibrium point; the discriminator is used for distinguishing a real high-energy CT image from a predicted high-energy CT image, the generator is used for generating the predicted high-energy CT image with the distribution which is finally matched with the real high-energy CT image, and the discriminator is deceived to classify the predicted high-energy CT image into real data to realize an unsupervised learning mode.
Wherein, the generator GAFor generating high-energy CT images from monoenergetic CT images, generator GBUsed for generating a single-energy CT image according to the high-energy CT image; discriminator DAFor distinguishing true high-energy CT images and for generating high-energy CT images, discriminator DBForDistinguishing a real single-energy CT image and generating a single-energy CT image;
in this embodiment, the generator specifically adopts a residual error network structure, the discriminator specifically adopts a PatchGAN network structure,
training the loop to generate a confrontation network by using a training set, wherein a generator loss function adopted during training is as follows:
the discriminator loss function is used as:
wherein,
and
is a generator G
AAnd generator G
BLoss function of F
cycle-spectralAnd F
cycle-singleIs a loss function of two cycles, λ is a weight coefficient,
and
as a discriminator D
AAnd discriminator D
BIs measured.
In this embodiment, an adaptive moment estimation algorithm is used as a solver to optimize network parameters of a supervised learning network and an unsupervised learning network.
S103, inputting the single-energy CT image to be imaged into an image conversion model determined by the model parameters, and calculating to output a high-energy CT image.
In the single-energy CT energy spectrum imaging method based on deep learning, after a high-energy CT image is obtained by using an image conversion model, the effect of energy spectrum CT conversion is evaluated by taking root mean square error and structural similarity as quantitative evaluation indexes of the image.
Fig. 2 is a related picture of high-energy CT image imaging performed by using an image transformation model constructed by a supervised learning network according to an embodiment of the present invention, where (a) is a single-energy CT image, (b) is a real high-energy CT image, and (c) is a predicted high-energy CT image;
fig. 3 is related pictures of high-energy CT image imaging performed by using an image transformation model constructed by an unsupervised learning network according to an embodiment of the present invention, where (a) is a single-energy CT image, (b) is a real high-energy CT image, and (c) is a predicted high-energy CT image.
By analyzing the images in fig. 2 and 3, the accuracy of predicting the CT number of the high-energy CT image is 10HU compared with the real high-energy CT image, and the texture features and the boundary features of the original high-energy CT image are well reproduced.
The above-mentioned embodiments are intended to illustrate the technical solutions and advantages of the present invention, and it should be understood that the above-mentioned embodiments are only the most preferred embodiments of the present invention, and are not intended to limit the present invention, and any modifications, additions, equivalents, etc. made within the scope of the principles of the present invention should be included in the scope of the present invention.