CN-120689524-B - Three-dimensional CT image reconstruction method, program product and device based on multi-view projection

CN120689524BCN 120689524 BCN120689524 BCN 120689524BCN-120689524-B

Abstract

The invention discloses a three-dimensional CT image reconstruction method, a program product and equipment based on multi-view projection, wherein the method comprises the steps of obtaining a plurality of projection images collected under different view angles of a three-dimensional object to be reconstructed and initial volume data corresponding to the three-dimensional object to be reconstructed, expanding channel dimensions of the projection images by utilizing a dimension expansion network in a pre-trained three-dimensional reconstruction model to obtain expanded projection images corresponding to the projection images, respectively taking each expanded projection image as input of a feature extraction network in the three-dimensional reconstruction model to extract multi-scale feature data corresponding to the projection images, obtaining multi-scale implicit neural representation data corresponding to the three-dimensional object to be reconstructed by utilizing an implicit neural learning network in the three-dimensional reconstruction model based on the initial volume data, reconstructing the three-dimensional image by utilizing a feature reconstruction network in the three-dimensional reconstruction model to output the three-dimensional reconstruction object.

Inventors

SUN XUEQIN
WU YANGXU
ZHAO XIAOJIE
KONG HUIHUA
CHEN PING
HU HAI
LI YIHONG
LI YU
WU YANFANG
WANG SUKAI
GUO LINA
WEI JIAOTONG

Assignees

中北大学

Dates

Publication Date: 20260508
Application Date: 20250710

Claims (8)

1. A three-dimensional CT image reconstruction method based on multi-view projection, the method comprising: acquiring a plurality of projection images acquired under different view angles of a three-dimensional target to be reconstructed and initial volume data corresponding to the three-dimensional target to be reconstructed; Performing channel dimension expansion on the projection images by using a dimension expansion network in a pre-trained three-dimensional reconstruction model to obtain expansion projection images corresponding to the projection images; Respectively taking each extended projection image as the input of a feature extraction network in the three-dimensional reconstruction model, and extracting multi-scale feature data corresponding to each projection image; Based on the initial volume data, utilizing an implicit neural learning network in a three-dimensional reconstruction model to obtain multi-scale implicit neural representation data corresponding to the three-dimensional target to be reconstructed; Based on the multi-scale feature data corresponding to each projection image and the multi-scale implicit neural representation data, carrying out three-dimensional image reconstruction by utilizing a feature reconstruction network in the three-dimensional reconstruction model, and outputting a three-dimensional reconstruction target, wherein the feature extraction network comprises an encoder and a decoder, the decoder is connected with the feature reconstruction network, the implicit neural learning network is connected with the feature reconstruction network, the encoder comprises a plurality of encoding layers, the decoder comprises a plurality of decoding layers, the encoding layers positioned in the same scale layer are in jump connection with the corresponding decoding layers, the feature reconstruction network comprises a plurality of reconstruction layers with the same layer number as the decoder and a reconstruction output layer connected with the reconstruction layers, the input of the decoding layers positioned in the same scale layer is used as the input of the corresponding reconstruction layer, the feature reconstruction network comprises a multi-scale fusion network, the implicit neural learning network is connected with the multi-scale fusion network, the multi-scale fusion network is respectively connected with each layer of the feature reconstruction network, the multi-scale fusion network is used for outputting each decoding layer, the current three-dimensional neural image reconstruction image is obtained by utilizing the current three-dimensional neural image reconstruction model, the current three-dimensional image reconstruction target is obtained by utilizing the current three-dimensional neural representation data, the three-dimensional reconstruction method comprises the steps of obtaining current scale feature data output by a current reconstruction layer, obtaining maximum scale feature data output by a last decoding layer corresponding to each projection image, obtaining current scale implicit neural representation data corresponding to the current layer number output by a multi-scale fusion network based on initial volume data, obtaining current scale feature data output by the current reconstruction layer based on current scale feature input data corresponding to each projection image, output of the last reconstruction layer and the current scale implicit neural representation data, sequentially executing until final reconstruction feature data of the last reconstruction layer are obtained, obtaining maximum scale feature data corresponding to each projection image and output by the last decoding layer corresponding to each projection image, obtaining maximum scale feature data corresponding to each projection image, and outputting the predicted three-dimensional reconstruction target through the reconstruction output layer based on the maximum scale feature data corresponding to each projection image, the implicit neural representation data corresponding to the reconstruction output layer and the final reconstruction feature data.
2. The multi-view projection-based three-dimensional CT image reconstruction method according to claim 1, wherein the dimension expansion network includes a gradient calculation layer, a gradient stitching layer and a dimension expansion layer, the channel dimension expansion is performed on the projection images by using the dimension expansion network in the pre-trained three-dimensional reconstruction model, so as to obtain expanded projection images corresponding to the projection images, and the method includes: calculating gradients in the horizontal direction and the vertical direction by utilizing a gradient calculation layer based on the projection image to obtain a gradient image corresponding to the projection image; summing the projection image and the gradient image corresponding to the projection image by utilizing a gradient stitching layer to obtain a composite image corresponding to the projection image, and performing channel stitching on the projection image, the gradient image corresponding to the projection image and the composite image corresponding to the projection image to obtain a three-channel initial image corresponding to the projection image; and gradually extracting feature images in the three-channel initial image based on a plurality of cascaded residual blocks in the dimension expansion layer to obtain an expansion projection image corresponding to the projection image.
3. The multi-view projection-based three-dimensional CT image reconstruction method according to claim 1 or 2, further comprising: acquiring a training data set, wherein each training sample in the training data set comprises a plurality of sample projection images acquired for a three-dimensional sample target at multiple angles of view, initial sample volume data corresponding to the three-dimensional sample target and a three-dimensional reconstruction volume label corresponding to the three-dimensional sample target; Based on the training data set, acquiring an input training sample of the current iteration, iteratively training an initial three-dimensional reconstruction model to obtain a three-dimensional reconstruction sample of the current iteration, and calculating a current total loss value based on a three-dimensional reconstruction volume label corresponding to the input training sample and the three-dimensional reconstruction sample of the current iteration; based on the current total loss value, judging whether the current iteration meets an iteration termination condition, if the current iteration does not meet the iteration termination condition, continuing to acquire an input training sample for iterative training until the current iteration meets the iteration termination condition, and taking the three-dimensional reconstruction model meeting the iteration termination condition as a pre-trained three-dimensional reconstruction model.
4. The multi-view projection-based three-dimensional CT image reconstruction method of claim 3, wherein calculating a current total loss value based on the three-dimensional reconstructed volume label corresponding to the input training sample and the three-dimensional reconstructed sample of the current iteration comprises: Calculating a current reconstruction loss value by using a first loss function based on a three-dimensional reconstruction volume label corresponding to the input training sample and a three-dimensional reconstruction sample of a current iteration, wherein the current reconstruction loss value is used for indicating the structural difference between the predicted three-dimensional reconstruction sample and the three-dimensional reconstruction volume label; Calculating a current structure similarity loss value by using a second loss function based on the three-dimensional reconstruction volume label corresponding to the input training sample and the three-dimensional reconstruction sample of the current iteration, wherein the current structure similarity loss value is used for indicating the similarity difference between the predicted three-dimensional reconstruction sample and the three-dimensional reconstruction volume label; calculating a current gradient loss by using a third loss function based on the three-dimensional reconstruction volume label corresponding to the input training sample and the three-dimensional reconstruction sample of the current iteration, wherein the current gradient loss is used for indicating the difference of image detail information between the predicted three-dimensional reconstruction sample and the three-dimensional reconstruction volume label; calculating current projection loss by using a fourth loss function based on the three-dimensional reconstruction volume label corresponding to the input training sample and the three-dimensional reconstruction sample of the current iteration, wherein the current projection loss is used for indicating physical process loss of CT imaging; And weighting the current reconstruction loss value, the current structure similarity loss value, the current gradient loss and the current projection loss to obtain the current total loss value.
5. The multi-view projection based three-dimensional CT image reconstruction method according to claim 4, wherein calculating the current gradient loss using a third loss function based on the three-dimensional reconstruction volume label corresponding to the input training sample and the three-dimensional reconstruction sample of the current iteration comprises: Calculating horizontal gradient loss in the horizontal direction between the three-dimensional reconstruction volume label corresponding to the input training sample and the three-dimensional reconstruction sample of the current iteration by using a mean square error function, and calculating vertical gradient loss in the vertical direction between the three-dimensional reconstruction volume label corresponding to the input training sample and the three-dimensional reconstruction sample of the current iteration; the current gradient penalty is calculated based on the horizontal gradient penalty and the vertical gradient penalty.
6. The multi-view projection-based three-dimensional CT image reconstruction method of claim 4, wherein calculating the current projection loss using a fourth loss function based on the three-dimensional reconstruction volume label corresponding to the input training sample and the three-dimensional reconstruction sample of the current iteration comprises: Calculating current projection data of the current iteration based on the three-dimensional reconstruction sample of the current iteration; calculating a projection tag based on the three-dimensional reconstruction volume tag; and calculating the current projection loss based on the current projection data and the projection tag.
7. A computing device comprising a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the multi-view projection-based three-dimensional CT image reconstruction method of any one of claims 1 to 6.
8. A computer program product comprising a computer program which, when executed by a processor, implements a multi-view projection based three-dimensional CT image reconstruction method as claimed in any one of claims 1 to 6.

Description

Three-dimensional CT image reconstruction method, program product and device based on multi-view projection Technical Field The invention relates to the technical field of computer tomography, in particular to a three-dimensional CT image reconstruction method and computing equipment based on multi-view projection. Background The computed tomography (Computed Tomography, CT) technology has wide application space in industry by virtue of the characteristics of non-destructive, high-efficiency, non-contact and the like. The CT technique can accurately describe the internal information of the measured object. At present, the CT imaging technology is mainly divided into two categories, namely traditional reconstruction and deep learning reconstruction. Conventional reconstruction methods can be broadly classified into 2 categories, analytical reconstruction algorithms and iterative reconstruction algorithms. With hundreds of X-ray projections, conventional reconstruction algorithms can approximately reconstruct a CT volume. Due to the limitation of the nyquist sampling theorem, the key of high-quality reconstruction of the traditional reconstruction algorithm is to acquire complete projections, which limits the reconstruction speed and the application scene to a great extent. After that, a learner introduces compressed sensing into CT reconstruction to develop a regularized reconstruction method, and integrates prior information into a reconstruction model, so that the number of required projection images is successfully reduced. However, in some application scenarios with limited conditions, only very limited projection images, such as on-line detection of industrial products, dynamic test of transient damage, etc., can be obtained, and high-quality results cannot be obtained by means of the conventional method and the regularization method. Therefore, reconstructing images from ultra-sparsely sampled projections to expedite the CT imaging process is a hot spot problem of research. Recently, with the rapid development of deep learning, the method is applied to sparse view and ultra-sparse view CT reconstruction. The mapping function of X-ray projection to the 3D CT volume is directly learned by using a neural network designed manually, and good effect is obtained. Compared with the traditional regularization-based method, the deep learning method extracts prior information of the prediction CT through a data-driven end-to-end network, so that CT reconstruction can be realized more practically. However, depth learning based reconstruction methods also face ubiquitous limitations in application. First, it is difficult to obtain large-scale training data sets, which can constitute a significant bottleneck in certain application scenarios. In addition, the CT reconstruction technology based on deep learning is poor in generalization capability, is difficult to adapt to diversified imaging objects, and limits the wide applicability of the CT reconstruction technology. Furthermore, the three-dimensional CT reconstruction algorithm generally takes three-dimensional convolution as a core operation unit, and realizes feature mapping from projection data to voxel space through the space dimension perception characteristic of the three-dimensional convolution. However, three-dimensional convolution in cascade form can cause the quantity of parameters to be increased sharply, so that the consumption of the video memory is remarkable, the dimension and the reconstruction rate of the reconstruction voxels are severely restricted, the consumption of calculation resources is also increased, and higher requirements are put on the memory. Disclosure of Invention In order to solve the existing technical problems, the invention provides a three-dimensional CT image reconstruction method and a computing device based on multi-view projection, which can reconstruct a 3D CT image with high quality rapidly through projection images with fewer view angles, thereby improving the three-dimensional reconstruction speed and the three-dimensional reconstruction quality. The three-dimensional CT image reconstruction method based on multi-view projection comprises the steps of obtaining a plurality of projection images collected under different view angles of a three-dimensional object to be reconstructed and initial volume data corresponding to the three-dimensional object to be reconstructed, expanding channel dimensions of the projection images by utilizing a dimension expansion network in a pre-trained three-dimensional reconstruction model to obtain expanded projection images corresponding to the projection images, respectively taking each expanded projection image as input of a feature extraction network in the three-dimensional reconstruction model to extract multi-scale feature data corresponding to each projection image, obtaining multi-scale implicit neural representation data corresponding to the three-dimensional object to be reco