Search

CN-116823989-B - Tomographic imaging and reconstruction method based on learning

CN116823989BCN 116823989 BCN116823989 BCN 116823989BCN-116823989-B

Abstract

The invention provides a tomographic imaging and reconstruction method based on learning, which measures scene density distribution in an illumination multiplexing mode. The light source emits light according to the intensity obtained by pre-learning during imaging, the light from different directions is absorbed and attenuated by the scene and then reaches the sensor, and the measured value is calculated and reconstructed to obtain the density information of the scene. The luminous intensity and the reconstruction algorithm are learned by a neural network. The CT imaging process is modeled as a linear full-connection layer, the weight corresponds to the luminous intensity of a light source during imaging, the reconstruction algorithm is modeled as a nonlinear neural network, and the method can be optimized in a targeted manner according to the characteristics of scanning geometry. The method has the advantages that the quantity of acquired data is small, strong priori assumption is not needed in calculation reconstruction, high-efficiency and high-quality CT acquisition reconstruction is realized, and the method can be applied to high-speed dynamic scene three-dimensional reconstruction.

Inventors

  • WU HONGZHI
  • ZHOU KUN
  • Kang Kaizhang

Assignees

  • 浙江大学

Dates

Publication Date
20260508
Application Date
20230627

Claims (10)

  1. 1. A tomographic imaging and reconstruction method based on learning is characterized in that scene density distribution is measured in an illumination multiplexing mode, and the method comprises the following steps: (1) Generating training data, namely acquiring parameters of scanning equipment, including the spatial positions of a light source and a sensor and the readings of all sensors when any light source emits light with maximum intensity under an empty scene; (2) Training a neural network according to the data generated in the step (1), wherein the neural network is characterized by comprising the following steps: a. Inputting CT images of all light sources of the scanning equipment; b. Outputting the corresponding density field; c. the first layer of the neural network is a linear full-connection layer, and a parameter matrix of the linear full-connection layer is obtained through the following formula: W l ＝f w (W raw ) W raw is a parameter to be trained, W l corresponds to an illumination intensity matrix in imaging, the size is k multiplied by n s ,n s , the number of light sources of the scanning device is k is the number of samples, and f w is a mapping for converting W raw so that the generated illumination intensity matrix W l can correspond to the possible luminous intensity of the light sources; d. The second layer and the later layers are nonlinear neural networks, and the output of the last layer is a density field reconstruction result; (3) And (3) the light source of the scanning equipment emits light row by row according to the illumination intensity matrix extracted in the step (2), sequentially irradiates a target scene, obtains a measured value matrix M through a sensor, wherein the size of the measured value matrix M is k multiplied by n d ,n d , and is the number of the sensors, and the M is used as the output of a first linear full-connection layer of the neural network to calculate and obtain a reconstructed density field.
  2. 2. The method for learning-based tomographic imaging and reconstruction of claim 1, wherein in the step (1), the specific method for generating the CT image is to randomly place a plurality of objects with different densities in the effective area of the scene, and generate the CT image based on the selected light model according to the calibrated light source and sensor positions.
  3. 3. The learning-based tomographic imaging and reconstruction method according to claim 2, wherein the light model is a linear absorption model, and the formula is as follows: Wherein I is a matrix composed of CT images of different light sources, the size is n s ×n d , element I ij in I is the reading of the jth sensor when the ith light source emits light with maximum intensity under a given density field, x is a vector after the density field is discretized into voxels, the length is the number of voxels n v , K is the Radon transform (Radon transform) represented by a third-order tensor, x 3 is the 3-factorial operation (mode-3 product) of tensor and vector, and as-is-by-element multiplication operation between matrices, I when the density field is everywhere 0.
  4. 4. The learning-based tomographic imaging and reconstruction method according to claim 1, wherein in the step (2), the relationship between the linear full-link layer and the input is as follows: M=W l I Wherein I is a matrix formed by CT images of different light sources, and the size is n s ×n d .
  5. 5. The learning-based tomographic imaging and reconstruction method according to claim 1, wherein in the step (2), the neural network for reconstruction is represented as follows: x nn ＝f recon (D nn )＝f recon (f nn (M)) Where f nn maps M to Sinogram D nn and then uses tomographic reconstruction method f recon to obtain a density field reconstruction result x nn .
  6. 6. The learning-based tomographic imaging and reconstruction method according to claim 5, wherein the tomographic reconstruction method is implemented using a filtered back projection method (filtered back projection).
  7. 7. The learning-based tomographic imaging and reconstruction method according to claim 1, wherein in the step (2), a loss function for training The expression is as follows: Wherein the method comprises the steps of For evaluating the quality of the reconstruction of the density field, For providing a specific property of the luminous intensity, g w is a function of the property of the actually estimated luminous intensity, and lambda r and lambda p are used for balancing the weights between the different loss functions.
  8. 8. The learning-based tomographic imaging and reconstruction method according to claim 7 wherein the loss function Where x is the vector after discretizing the density field into voxels and x nn is the density field reconstruction result.
  9. 9. The learning-based tomographic imaging and reconstruction method according to claim 7, wherein g w (W l )＝-∑|W l is set to be i for a dynamic scene requiring high-speed scanning, so that the value of W l is urged to be binarized.
  10. 10. The learning-based tomographic imaging and reconstruction method according to claim 7, wherein g w (W l )＝∑|W l is set to cause W l to take a value that tends to be minimized for a scene requiring a low dose scan.

Description

Tomographic imaging and reconstruction method based on learning Technical Field The invention relates to the technical field of computed tomography (Computed Tomography, CT), in particular to an imaging and reconstruction method. Background Tomographic scanning (Computed Tomography, CT) is an important research direction in the field of medical imaging and computer vision graphics. The technology can reconstruct the internal structure of an object by measuring the absorption of light rays by a scene in different directions. The technology has wide application fields such as medical imaging, industrial monitoring, aviation security inspection and cultural relic protection. In particular, the application of the tomography technology to dynamic scenes has great scientific and application value. Such as mechanical detection and medical diagnostics, require three-dimensional reconstruction of high-speed dynamic scenes. However, expanding the traditional tomography to dynamic scenes has a critical problem that because high-quality reconstruction results are often based on dense sampling in different directions, when the scene changes rapidly, the dense sampling process must be completed in a short time, so that the problem of ghost in reconstruction can be avoided, and the reconstruction quality is ensured. This makes dynamic scene oriented tomography far beyond the high sampling capability of conventional methods. Over the past decades, there have been many studies that have proposed different algorithms for tomographic acquisition reconstruction for dynamic scenarios. The properties of the observed scene may be exploited for solving for specific dynamic phenomena. Such as Chen et al, for periodic cardiac motion, propose a scan reconstruction algorithm (Chen Guang-Hong,Theriault-Lauzier Pascal,Tang Jie,Nett Brian,Leng Shuai,Zambelli Joseph,Qi Zhihua,Bevins Nicholas,Raval Amish,Reeder Scott.2011.Time-resolved interventional cardiac C-arm cone-beam CT:An application of the PICCS algorithm.IEEE transactions on medical imaging.31,4,907-923). that is limited to certain scene characteristics and is not ubiquitous. There is also research to increase the sampling speed by reducing the number of point light source samples and making stronger a priori assumptions to calculate the reconstruction, which limits the application scenario (Zang Guangming,Idoughi Ramzi,Wang Congli,Bennett Anthony,Du Jianguo,Skeen Scott,Roberts William L.,Wonka Peter,Heidrich Wolfgang.2020.TomoFluid:reconstructing dynamic fluid from sparse view videos.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.1870-1879). of the method, so that it is highly desirable to propose a tomographic acquisition reconstruction algorithm for general dynamic scenarios. Disclosure of Invention The invention aims to provide a method for greatly improving acquisition capacity and facing a general scene, aiming at the problems that the acquisition capacity of the existing tomography method is insufficient and strong priori assumption is needed for reconstruction under a dynamic scene. In order to achieve the above object, the present invention provides a tomographic imaging and reconstruction method based on learning, which measures scene density distribution in an illumination multiplexing manner, comprising the following steps: (1) Generating training data, namely acquiring parameters of scanning equipment, including the spatial positions of a light source and a sensor and the readings of all sensors when any light source emits light with maximum intensity under an empty scene; (2) Training a neural network according to the data generated in the step (1), wherein the neural network is characterized by comprising the following steps: a. Inputting CT images of all light sources of the scanning equipment; b. Outputting the corresponding density field; c. the first layer of the neural network is a linear full-connection layer, and a parameter matrix of the linear full-connection layer is obtained through the following formula: Wl＝fw(Wraw) W raw is a parameter to be trained, W l corresponds to an illumination intensity matrix in imaging, the size is k multiplied by n s,ns, the number of light sources of the scanning device is k is the number of samples, and f w is a mapping for converting W raw so that the generated illumination intensity matrix W l can correspond to the possible luminous intensity of the light sources; d. The second layer and the later layers are nonlinear neural networks, and the output of the last layer is a density field reconstruction result; (3) And (3) the light source of the scanning equipment emits light row by row according to the illumination intensity matrix extracted in the step (2), sequentially irradiates a target scene, obtains a measured value matrix M through a sensor, wherein the size of the measured value matrix M is k multiplied by n d,nd, and is the number of the sensors, and the M is used as the output