A Method for Constructing a Regression Analysis Model Based on Deep Neural Networks

Abstract

This invention discloses a regression analysis model construction method based on deep neural networks and applies it to the field of optical molecular imaging. The main research results of this invention are as follows: (1) A regression analysis model based on deep neural networks is proposed, which overcomes the limitation of artificial intelligence methods that can only reconstruct targets on the same grid, and realizes the reconstruction of light distribution information on biological tissue surfaces generated by grids with different structures into light source information; (2) A research method based on point cloud technology is proposed, which overcomes the limitation of the reconstructed light source shape in traditional reconstruction methods that is limited by the grid subdivision, and realizes high-precision morphological reconstruction of the light source. This invention has important application value in the fields of medical molecular imaging and reconstruction methods.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision and molecular imaging, and relates to a regression analysis model based on deep neural networks and a research method based on point cloud technology, which is applicable to image processing and optical molecular imaging. Background Technology

[0002] Optical molecular imaging is a rapidly developing field of biomedical imaging that combines optical processes with specific molecular properties. Utilizing the optical properties of specific probes, it reflects changes at the molecular level in vivo, enabling early detection of pathological changes in tissues. This imaging method offers advantages such as being radiation-free, highly sensitive, and having high spatiotemporal resolution. It is frequently used in small animal research for molecular, cellular, and gene expression imaging, facilitating drug development, disease research, and therapeutic interventions.

[0003] Common optical molecular imaging modalities are categorized into direct optical molecular imaging and indirect optical molecular imaging, based on whether external energy excitation is required during the imaging process. Direct optical molecular imaging is an imaging technique that does not require external energy excitation. It primarily achieves imaging by detecting photons emitted by specific molecular probes or radioactive nuclides, including bioluminescence tomography and Cherenkov optical imaging. Indirect optical molecular imaging is an imaging technique that requires external energy excitation, such as fluorescence molecular tomography using lasers as the external energy source and X-ray luminescence tomography using X-rays as the excitation source.

[0004] In medicine, to obtain information about lesions within organisms non-invasively, imaging-friendly nanoluminescent probes can be combined with the lesions. Specific nanoprobes attached to the lesion site spontaneously or under stimulated emission of photons. During their transmission from the inside out, the light undergoes complex physical processes such as reflection, absorption, and scattering, ultimately forming light distribution information on the surface of the organism. Previous studies have yielded numerous fruitful results in simulating the light transmission process in biological tissues, typically employing a trade-off between accuracy and efficiency. This approach is not only cumbersome but also suffers from limitations in computational accuracy and efficiency, hindering the rapid and accurate reconstruction of light sources in optical molecular tomography. Artificial intelligence methods can only reconstruct on the same grid or grids with the same structure, a limitation that severely restricts their application. Traditional reconstruction methods are time-consuming and inefficient, and the reconstructed light sources are morphologically severely constrained by the grid structure, failing to accurately reflect the morphology of the real light source. Therefore, it is essential to explore a deep neural network-based regression analysis model for a fusion point cloud research method that is easy to implement and possesses good localization and morphological reconstruction capabilities, enabling reconstruction on grids with different structures. Summary of the Invention

[0005] To address the shortcomings and deficiencies in existing technologies, this invention provides a regression analysis model based on deep neural networks that integrates point cloud research methods, solving two defects of existing reconstruction methods: First, artificial intelligence methods can only reconstruct on the same grid or grids with the same structure; second, the shape of the light source reconstructed by traditional reconstruction methods is severely limited by the density of the grid subdivision.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for constructing a regression analysis model based on deep neural networks includes the following steps:

[0008] Step 1, Construction of training and test sets: Implant a light source with given coordinates and radius into the biological simulation model to obtain the actual light source coordinates and actual light source radius. The light source forms light distribution information on the surface of the biological organism, which includes light distribution coordinates and luminance. The light distribution coordinates and luminance are the sample data of the training set, and the actual light source coordinates and actual light source radius are the label data of the training set.

[0009] The process of constructing the test set is the same as that of constructing the training set;

[0010] Step 2: Model Training

[0011] The training set obtained in step one is input into the regression analysis model of the deep neural network for training, to obtain the nonlinear mapping relationship between the sample data and the label data of the training set. The test set is used for model detection.

[0012] Optionally, the regression analysis model of the deep neural network includes linear layers, convolutional layers, max pooling layers, and activation layers.

[0013] Optionally, the regression analysis model of the deep neural network uses MSEloss as the loss function to evaluate the error between the predicted value and the true value.

[0014] Optionally, training and test sets are constructed on the USC digital mouse simulation model. The USC digital mouse simulation model is a digital mouse model obtained by extracting tissue information from CT imaging slice data. The model size is 38.0mm×20.8mm×35.0mm, and it is divided into six tissue organs, from top to bottom: heart, lung, liver, stomach, kidney and muscle.

[0015] Optionally, in order to quantitatively evaluate the regression analysis performance of the deep neural network regression analysis model, the localization error (LE) and volume deviation (VD) are used as evaluation indicators to measure the coordinate and volume information of the reconstructed light source.

[0016] The positioning error LE is used to measure the Euclidean distance between the center of the reconstructed light source and the center of the real light source, and its formula is as follows:

[0017]

[0018] The center coordinates of the reconstructed light source are (x) cons ,y cons ,z cons The center coordinates of the real light source are (x... true ,y true ,z true );

[0019] Volume deviation (VD) measures the deviation between the reconstructed light source and the actual light source volume, and its formula is as follows:

[0020]

[0021] Where V cons To reconstruct the volume of the light source, V true This represents the volume of a real light source.

[0022] The neural network model constructed in this invention directly processes point cloud data containing light distribution information on the surface of biological tissues, ultimately outputting point cloud data reconstructed from the light source. By using several reconstructed light sources to fit the real light source, the morphological information of the light source is reconstructed more accurately. The regression analysis model based on deep neural networks overcomes the limitations of different mesh structures on reconstruction in artificial intelligence methods, realizing the reconstruction of light distribution information on the surface of biological tissues from meshes with different structures into light source information. Compared with traditional reconstruction methods, this invention has the unique advantage of parallel processing, which can shorten reconstruction time and improve reconstruction efficiency. While achieving good positioning accuracy, the application of the point cloud technology-based research method overcomes the limitation of the reconstructed light source shape being affected by mesh partitioning in traditional reconstruction methods, making it more consistent with the shape of the real light source. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate the present disclosure and form part of the specification. They are used together with the following detailed description to explain the present disclosure, but do not constitute a limitation thereof. In the drawings:

[0024] Figure 1 This is a digital mouse simulation model from the University of Southern California.

[0025] Figure 2 Example of light distribution information on the surface of biological tissue;

[0026] Figure 3 The local mesh structure of the digital mouse model in the real light source and inverse reconstruction in the traditional reconstruction method based on finite element mesh;

[0027] Figure 4 This is a schematic diagram illustrating how traditional reconstruction methods based on finite element meshes are limited by the density of the inverse mesh.

[0028] Figure 5 This is a schematic diagram illustrating the application of point cloud methods in reverse reconstruction.

[0029] Figure 6 The results are from a single-source numerical simulation experiment. The black spheres represent the actual light source, and the cluster of gray spheres represents the reconstructed light source.

[0030] Figure 7 The results are from a numerical simulation experiment using two light sources. The black spheres represent the actual light source, while the cluster of gray spheres represents the reconstructed light source.

[0031] Figure 8 This is the overall flowchart of the present invention.

[0032] Specific implementation examples

[0033] The invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0034] This regression analysis model, based on deep neural networks, directly processes point cloud data. It takes point cloud data containing light distribution information from the surface of biological tissue as input and outputs point cloud data of the light source within the organism. Therefore, this model overcomes the limitation of artificial intelligence methods, which can only reconstruct targets on a single mesh, and can learn and reconstruct target information from meshes with different structures. Furthermore, it overcomes the limitation of traditional finite element methods, where the shape of the reconstructed light source is severely restricted by the mesh division, resulting in higher accuracy and generalization in reconstructing the shape of the light source.

[0035] To address the first shortcoming mentioned above—that "artificial intelligence methods can only reconstruct on the same grid or grids with the same structure"—a regression analysis model based on deep neural networks is proposed. This model comprises linear layers, convolutional layers, max-pooling layers, and ReLU layers. The max-pooling layer plays a crucial role, extracting and retaining the most typical data features from the feature map. For example, if three number sequences [1,2,3], [2,3,1], and [3,1,2] are input into the max-pooling layer, the output will always be 3. Clearly, in this example, the max-pooling layer ignores the subtle features of the three number sequences, extracting and retaining only the most prominent feature "3". Therefore, the inconsistency of the three number sequences does not affect the result of the max-pooling process. Similarly, by migrating to different grids and arranging the points on each grid according to their numbers, we obtain a series of number sequences [node_1, node_2, node_3, ..., node_n], similar to the one described above. Since the construction of the training and test sets in step one above used Amira software for grid partitioning, the grid spatial topology of each sample data in the dataset and test sets is different, resulting in different spatial positions of points on the grid. Likewise, by inputting the large amount of sample data [node_1, node_2, node_3, ..., node_n] from the dataset and test sets into the MaxPooling layer, just as with the number sequences [1,2,3], [2,3,1], and [3,1,2], the MaxPooling layer only extracts and retains the most salient features of different number sequences, ignoring subtle features. Therefore, the inconsistency of different number sequences does not affect the results of the MaxPooling process. In summary, this model overcomes the limitations imposed by the different grid spatial topologies of the sample data in the dataset and test sets. Therefore, this model is able to process the light distribution information on the surface of biological tissues obtained by forward simulation of simulation models with different mesh structures.

[0036] Figure 8 This is the overall flowchart of the invention. Sub-network one is used for input transformation, converting the point cloud input into a 4×4 matrix. Sub-network two is used for feature transformation, converting the extracted features into a 64×64 matrix. Convolutional layers and linear layers are used for feature extraction from the point cloud data. Max pooling layers fuse the features to obtain the final global features.

[0037] Traditional reconstruction methods based on finite element meshes divide the digital mouse model into meshes of a certain density and then reconstruct the target light source based on these meshes. To briefly describe these traditional finite element mesh-based reconstruction methods, we will illustrate them in two-dimensional space.

[0038] like Figure 3 As shown, the left side represents the real light source, and the right side represents the local mesh structure of the digital mouse model used in the reverse reconstruction (the expected reconstruction result area).

[0039] Figure 4 The images show the local mesh structures (expected reconstruction result areas) of digital mouse models with different densities used in the inverse reconstruction. From (a) to (c), the mesh structures are sparse, medium-density, and dense, respectively, showing a gradual increase in the number of mesh cells. Traditional reconstruction methods reconstruct based on each small mesh cell. Assuming the inverse reconstruction method is determined... Figure 4 The red dashed circle represents the expected reconstruction result area. It can be seen that on (a) a sparse mesh, the reconstructed result area far exceeds the expected reconstruction result area; on (b) a medium-density mesh, the reconstructed result area exceeds the expected reconstruction result area; and on (c) a dense mesh, the reconstructed result area is very close to the expected reconstruction result area. Therefore, the shape of the light source reconstructed by the traditional finite element mesh-based reconstruction method is limited by the density of the inverse mesh. The above describes the dependence of the traditional finite element mesh-based reconstruction method on mesh density, which intuitively demonstrates the limitation that the density of the inverse mesh has on the reconstructed light source shape in the traditional reconstruction method.

[0040] To address the second drawback mentioned above, namely that "the shape of the light source reconstructed by traditional methods is severely limited by the density of the mesh," a research method based on point cloud technology is proposed. Point cloud data of light distribution information on the surface of biological tissue is used as input, and point cloud data of the light source within the organism is output. To more accurately reconstruct the morphological information of the light source, several reconstructed light sources are used to fit the real light source. The geometry formed by all the reconstructed light sources is the final reconstructed light source. This overcomes the limitation of mesh spatial structure information on reconstruction, achieving reconstruction entirely in numerical space, and achieving the goal of "overcoming the excessive dependence of the reconstructed target shape on the finite element mesh," as shown below. Figure 5 As shown.

[0041] Specifically, the regression analysis model construction method based on deep neural networks of this invention includes the following steps:

[0042] Step 1, Construction of Training and Test Sets: A light source with given coordinates (X, Y, and Z) and radius (r) is implanted into the biological simulation model to obtain the actual light source coordinates (X, Y, and Z) and actual light source radius (r). The light source forms light distribution information on the biological surface, including light distribution coordinates and luminance. The light distribution coordinates and luminance are the sample data for the training set, while the actual light source coordinates and actual light source radius are the label data for the training set.

[0043] The process of constructing the test set is the same as that of constructing the training set;

[0044] Step 2: Model Training

[0045] The training set obtained in step one is input into the regression analysis model of the deep neural network for training, to obtain the nonlinear mapping relationship between the sample data and the label data of the training set. The test set is used for model detection. The regression analysis model of the deep neural network includes linear layers, convolutional layers, max pooling layers and activation layers. MSEloss is used as the loss function to evaluate the error between the predicted value and the true value.

[0046] like Figure 1 As shown, the training and test sets were constructed on the USC digital mouse simulation model. The USC digital mouse simulation model is a digital mouse model obtained from tissue information extracted from CT imaging slice data. The model size is 38.0mm × 20.8mm × 35.0mm, divided into six organs: heart, lung, liver, stomach, kidney, and muscle, from top to bottom. Specific nanoprobes inside biological tissues spontaneously or under stimulated emission emit photons. During the transmission of light from the inside out, it undergoes various physical and optical processes such as reflection, absorption, and scattering, forming light distribution information on the surface of the biological tissue, such as... Figure 2 As shown.

[0047] To quantitatively evaluate the regression analysis performance of the deep neural network regression analysis model, the localization error (LE) and volume deviation (VD) are used as evaluation indicators to measure the coordinate and volume information of the reconstructed light source.

[0048] The positioning error LE is used to measure the Euclidean distance between the center of the reconstructed light source and the center of the real light source, and its formula is as follows:

[0049]

[0050] The center coordinates of the reconstructed light source are (x) cons ,y cons ,z cons The center coordinates of the real light source are (x... true ,y true ,z true );

[0051] Volume deviation (VD) measures the deviation between the reconstructed light source and the actual light source volume, and its formula is as follows:

[0052]

[0053] Where V cons To reconstruct the volume of the light source, V true This represents the volume of a real light source.

[0054] The calculation formula is as follows:

[0055]

[0056]

[0057] Among them, P cons P true Let r be the number of nodes in the reconstructed light source and the number of nodes in the actual light source under the same sampling interval, where r is the actual radius of the light source.

[0058] Example 1:

[0059] To verify the good performance of the proposed model, we set up a single-light source numerical simulation experiment. We used the USC digital mouse simulation model, with dimensions of 38.0mm × 20.8mm × 35.0mm. In this experiment, the light source coordinates of the training set were distributed throughout the entire space of the digital mouse simulation model, with x-coordinate ranging from 0mm to 38.0mm, y-coordinate from 0mm to 20.8mm, and z-coordinate from 0mm to 35.0mm. The light source radius in the training set ranged from 0.8mm to 1.1mm. The light source center coordinates of the test sample were set at (18.0mm, 6.0mm, 18.0mm), and the light source radius was set to 1.0mm.

[0060] Experimental results show that our model has good target localization and morphological reconstruction performance. The reconstructed center coordinates of the light source of the test sample are (18.083mm, 6.158mm, 17.379mm), the localization error LE is 0.646mm, and the volume deviation VD is 0.11. The experimental results of the single-light source numerical simulation are shown in the table below. Figure 6 As shown.

[0061]

[0062] Example 2:

[0063] The training set for the dual-source numerical simulation experiment is the same as that for the single-source numerical simulation experiment in Example 1.

[0064] The center coordinates of light source 1 in the test sample are set at (13.0mm, 15.0mm, 18.0mm), and the radius of the light source is set to 1.0mm; the center coordinates of light source 2 are set at (13.0mm, 15.0mm, 24.0mm), and the radius of the light source is set to 1.0mm.

[0065] Experimental results show that our model has good target localization and morphological reconstruction performance. The reconstructed center coordinates of light source 1 in the test sample are (13.398mm, 14.762mm, 17.571mm), with a localization error LE of 0.632mm and a volume deviation VD of 0.07; the reconstructed center coordinates of light source 2 are (12.262mm, 15.334mm, 22.768mm), with a localization error LE of 1.474mm and a volume deviation VD of 0.53. The experimental results of the dual-light source numerical simulation are shown in the table below. Figure 7 As shown.

[0066]

[0067] The two examples above demonstrate that our model can accurately reconstruct the location of the target in different grid structures, and also has the ability to perform good morphological reconstruction of the target.

[0068] The preferred embodiments of the present disclosure have been described in detail above with reference to the accompanying drawings. However, the present disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present disclosure, various simple modifications can be made to the technical solutions of the present disclosure, and these simple modifications all fall within the protection scope of the present disclosure.

[0069] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0070] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A method for constructing a regression analysis model based on a deep neural network, characterized in that, Includes the following steps: Step 1, Construction of training and test sets: Implant a light source with given coordinates and radius into the biological simulation model to obtain the actual light source coordinates and actual light source radius. The light source forms light distribution information on the surface of the biological organism, which includes light distribution coordinates and luminance. The light distribution coordinates and luminance are the sample data of the training set, and the actual light source coordinates and actual light source radius are the label data of the training set. The process of constructing the test set is the same as that of constructing the training set; Step 2: Model Training The training set obtained in step one is input into the regression analysis model of the deep neural network for training, to obtain the nonlinear mapping relationship between the sample data and the label data of the training set. The test set is used for model detection. The regression analysis model of the deep neural network includes linear layers, convolutional layers, max pooling layers, and activation layers. The regression analysis model of the deep neural network uses MSEloss as the loss function to evaluate the error between the predicted value and the true value. The model directly processes point cloud data, taking the point cloud data of light distribution information on the surface of biological tissue as input and outputting the point cloud data of light sources in the biological body. To quantitatively evaluate the regression analysis performance of the deep neural network regression analysis model, the localization error (LE) and volume deviation (VD) are used as evaluation indicators to measure the coordinate and volume information of the reconstructed light source. The positioning error LE is used to measure the Euclidean distance between the center of the reconstructed light source and the center of the real light source, and its formula is as follows: ； The center coordinates of the reconstructed light source are: The center coordinates of the real light source are ; Volume deviation (VD) measures the deviation between the reconstructed light source and the actual light source volume, and its formula is as follows: ； in To reconstruct the volume of the light source, The volume of a real light source; The calculation formula is as follows: ； ； in, , Let r be the number of nodes in the reconstructed light source and the number of nodes in the actual light source under the same sampling interval, where r is the actual radius of the light source.

2. The method for constructing a regression analysis model based on a deep neural network according to claim 1, characterized in that, The training and test sets were constructed on the digital mouse simulation model from the University of Southern California. The USC digital mouse simulation model is a digital mouse model obtained by extracting tissue information from CT imaging slice data. The model measures 38.0mm × 20.8mm × 35.0mm and is divided into six tissue organs, from top to bottom: heart, lungs, liver, stomach, kidneys, and muscles.

Images