A Deep Learning-Based Perspective 3D Reconstruction Method for Serous Epidermis

Abstract

This invention belongs to the field of bioinformatics, specifically relating to a deep learning-based method for perspective 3D reconstruction of serous epidermis. The method includes the following steps: First, stained continuous sections of the serous epidermis are acquired; second, the stained continuous sections are input into a reference point analysis module to obtain reference point descriptions; then, the reference point descriptions and stained continuous sections are processed by a continuous section registration module to obtain aligned continuous sections; subsequently, the aligned continuous sections are spatially filled in using a continuous section interpolation module to obtain filled continuous sections; finally, the filled continuous sections are spatially modeled using a continuous section stitching module to obtain an initial 3D reconstruction, and then a perspective 3D reconstruction is obtained using a section background removal module and a 3D data denoising module. This invention proposes a complete set of 3D reconstruction methods for serous epidermis based on deep learning technology, capable of performing perspective 3D reconstruction under a conventional microscope, possessing both color layer information and internal structure representation capabilities.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a deep learning-based method for three-dimensional reconstruction of serous epidermis. Background Technology

[0002] Given the high reproductive capacity and ecological adaptability of insects, frequent outbreaks of crop diseases and pests threaten food security and agricultural production. Against this backdrop, biological control, with its environmentally friendly and highly effective loss-mitigation advantages, has gradually become an important direction in integrated pest management. The egg stage, as the source of insect population replenishment and dispersal, is a crucial target for timely intervention, allowing for effective control before population establishment and reducing subsequent risks. The serous epidermis plays a core protective role during the embryonic development of insect eggs, such as resisting extreme environmental fluctuations, supporting tissue structure, and preventing pathogen invasion. Significant differences exist in the serous epidermis of different insect groups in terms of tissue thickness, cell arrangement, surface texture, and spatial configuration, which is closely related to the microenvironment and biological adaptation functions of their oviposition sites. Therefore, the structure of the serous epidermis is not only a key entry point for understanding the ecological adaptation and evolution of insects during the egg stage but also provides important morphological and functional evidence for screening egg-stage targeted biological control strategies.

[0003] Existing 3D reconstruction methods can be broadly categorized into imaging-based and algorithmic methods. The distinction lies in whether or not machine learning algorithms are incorporated: imaging-based methods rely on specialized equipment to acquire 3D data of the sample's surface or interior, while algorithmic methods infer the sample's 3D structure using algorithms after data acquisition. Imaging-based methods primarily include micro-computed tomography (Micro-CT) and transmission electron microscopy (TEM). Micro-CT is a crucial technique for high-resolution 3D imaging using X-rays. Its 3D reconstruction principle can be summarized as follows: first, two-dimensional X-ray projection images of the sample generated at different rotation angles are acquired; then, these two-dimensional images are reconstructed into a 3D model of the sample's internal structure. TEM, on the other hand, uses a high-energy electron beam to penetrate ultrathin slices; the differences in electron scattering and absorption in different tissue components create a contrast, thereby obtaining nanoscale resolution images of the ultrastructure. However, the above methods have certain defects in the three-dimensional reconstruction of the serous epidermis: (1) High equipment purchase and maintenance costs: High requirements for experimental conditions and operators, which is not conducive to large-scale, cross-species or multi-developmental stage systematic studies in conventional laboratories; (2) Limited ability to observe the three-dimensional structure inside the sample: Although Micro-CT can realize perspective three-dimensional imaging of the sample, it is often limited by spatial resolution and tissue contrast on the sample, and is prone to boundary blurring and detail loss. In contrast, TEM can provide nanoscale fine tissue ultrastructure information, but cannot realize continuous, closed perspective three-dimensional reconstruction and spatial correlation analysis; (3) Insufficient observation of tissues in various regions inside the sample: The imaging results are mainly based on grayscale and density contrast, and it is difficult to present information such as tissue layers, cell arrangement and interface layering at the same time, which is not conducive to fine identification and quantitative analysis of the internal tissue characteristics of the serous epidermis. Algorithm-based methods mainly include shape from focus (SFF) and shape from motion (SFM). Among them, SFF uses the mapping relationship between focus and depth to estimate the surface morphology of the sample from a fixed perspective, while SFM uses feature point matching and geometric constraint estimation to recover the surface point cloud model of the sample. However, the above methods have other defects in the three-dimensional reconstruction of the serous epidermis: (1) due to the principle of two-dimensional surface imaging, algorithmic methods cannot achieve perspective three-dimensional reconstruction; (2) the reconstruction quality is easily affected by error factors such as illumination changes and imaging noise, resulting in unstable three-dimensional reconstruction. Summary of the Invention

[0004] To overcome the shortcomings of existing 3D reconstruction methods for serous epidermis, the present invention aims to provide a deep learning-based perspective 3D reconstruction method for serous epidermis, comprising the following steps:

[0005] Step 1: Collect stained serial sections of the sample to be tested. , This represents the sequence number, and the size of a single frame slice is... , This indicates the height and width of a single slice. Indicates the number of channels in the slice;

[0006] Step 2: Process the stained serial sections obtained in Step 1 The reference point is input into the reference point analysis module, and the reference point description is obtained through equation (1). ,

[0007] (1)

[0008] in, The reference point analysis module is calculated as shown in equation (2).

[0009] (2)

[0010] in, For input data, This is a 2D convolution operation. For max pooling operation, It is a non-linear activation function;

[0011] Step 3: Describe the reference point obtained in Step 2 and the stained serial sections obtained in step 1 The data is input into the continuous slice registration module, and the continuous slices are spatially aligned using equation (3) to obtain aligned continuous slices. ,

[0012] (3)

[0013] in, For the continuous slice registration module, its calculation method is shown in equation (4).

[0014] (4)

[0015] in, Therefore Affine parameters for image Perform the transformation. This is the maximum response filtering operation. It is a normalized exponential function. It is a linear mapping operation. ;

[0016] Step 4: Align the continuous slices obtained in Step 3 The input is fed into the continuous slice interpolation module, and the original spatial position is filled in by iterative interpolation according to equation (5) to obtain the filled continuous slice. ,

[0017] (5)

[0018] in, This is a continuous slice frame interpolation module. To complete the total number of slices and , This indicates that adjacent slices in a continuous slice are aligned and , Indicates the frame insertion position, when This indicates a slice that fills in the middle of adjacent slices. The calculation method for the continuous slice interpolation module is shown in equation (6).

[0019] (6)

[0020] in, It is an image feature encoding operation. It is a feature concatenation operation. It is an image feature decoding operation;

[0021] Step 5: Complete the continuous slices obtained in Step 4. The data is input into the continuous slice stitching module, and the initial three-dimensional volume data is constructed using equation (7) to obtain the initial three-dimensional reconstruction. ,

[0022] (7)

[0023] in, This is a continuous slice splicing module;

[0024] Step 6: Apply the initial 3D model obtained in Step 5. The input is fed into the slice background removal module, and the threshold for removing the background color is set by equation (8) to obtain the perspective 3D reconstruction. ,

[0025] (8)

[0026] in, For the slice background removal module, This indicates the background threshold that is set;

[0027] Step 7: Obtain the perspective 3D from Step 6 By removing slice noise and interpolation noise using equation (9), the final perspective 3D reconstruction is obtained. ,

[0028] (9)

[0029] in, This is a combined Gaussian smoothing and median filtering denoising function.

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] (1) This invention proposes a complete three-dimensional reconstruction method for serous epidermis based on deep learning technology, which realizes robust registration of sections, spatial interpolation and completion, and volume data denoising and reconstruction. This method can realize a visualized perspective three-dimensional model of the serous epidermal region and quantifiable morphological indicators based on conventional stained sections, thereby supporting cross-species and cross-developmental stage structural comparison and functional correlation analysis.

[0032] (2) This invention does not rely on expensive special devices such as Micro-CT and TEM, and can complete three-dimensional reconstruction under conventional microscope and histological slide preparation conditions, which significantly reduces equipment costs and experimental thresholds. At the same time, compared with algorithms such as SFF and SFM that can only restore the morphology of the outer surface, this invention realizes three-dimensional observation of the serous epidermis based on continuous slice sequences, which has both histological color layer information and internal structure expression capabilities.

[0033] Therefore, this patent utilizes deep learning to construct a three-dimensional perspective reconstruction method for the serous epidermis during the insect egg stage, aiming to balance the representation of tissue details and internal three-dimensional reconstruction. Specifically, this method uses stained serial sections as data foundation and constructs a three-dimensional perspective structure of the serous epidermis that can be used for quantitative analysis through steps such as keypoint registration, section interpolation, and stereo stitching. This patent reduces equipment dependence and cost while preserving the details of colored histological layers and achieving continuous, closed three-dimensional representation of the serous epidermis, providing technical support for subsequent morphological difference comparison and index quantification. Attached Figure Description

[0034] Figure 1 A flowchart of a deep learning-based three-dimensional reconstruction method for serous epidermis;

[0035] Figure 2 This is a network structure diagram of a deep learning-based three-dimensional reconstruction method for serous epidermis.

[0036] Figure 3 These are the first 5 samples from the stained serial sections in step 1 of this invention;

[0037] Figure 4 This is a sample of continuous slice registration in step 3 of embodiment 1 of the present invention;

[0038] Figure 5 This is an example of continuous slice frame interpolation in step 4 of embodiment 1 of the present invention;

[0039] Figure 6 These are four view examples of perspective 3D reconstruction in step 7 of Embodiment 1 of the present invention. Detailed Implementation

[0040] Example 1

[0041] like Figure 1 , Figure 2 As shown, a deep learning-based three-dimensional reconstruction method for serous epidermis includes the following steps:

[0042] Step 1: Collect stained serial sections of the sample to be tested. , This represents the sequence number, and the size of a single frame slice is... , This indicates the height and width of a single slice. This represents the number of channels in a slice. Specifically, in the implementation example... Figure 3 , sequence number Image high Image width and image channels ;

[0043] Step 2: Process the stained serial sections obtained in Step 1 The reference point is input into the reference point analysis module, and the reference point description is obtained through equation (1). ,

[0044] (1)

[0045] in, The reference point analysis module is calculated as shown in equation (2).

[0046] (2)

[0047] in, For input data, This is a 2D convolution operation. For max pooling operation, It is a non-linear activation function;

[0048] Step 3: Describe the reference point obtained in Step 2 and the stained serial sections obtained in step 1 The data is input into the continuous slice registration module, and the continuous slices are spatially aligned using equation (3) to obtain aligned continuous slices. ,

[0049] (3)

[0050] in, For the continuous slice registration module, its calculation method is shown in equation (4).

[0051] (4)

[0052] in, Therefore Affine parameters for image Perform the transformation. This is the maximum response filtering operation. It is a normalized exponential function. It is a linear mapping operation. Serial slice registration sample, for example Figure 4 As shown, after pairwise registration, aligned continuous slices are obtained;

[0053] Step 4: Align the continuous slices obtained in Step 3 The input is fed into the continuous slice interpolation module, and the original spatial position is filled in by iterative interpolation according to equation (5) to obtain the filled continuous slice. ,

[0054] (5)

[0055] in, This is a continuous slice frame interpolation module. To complete the total number of slices and , This indicates that adjacent slices in a continuous slice are aligned and , Indicates the frame insertion position, when This indicates a slice that fills in the middle of adjacent slices. The calculation method for the continuous slice interpolation module is shown in equation (6).

[0056] (6)

[0057] in, It is an image feature encoding operation. It is a feature concatenation operation. It is an image feature decoding operation, such as continuous slice frame interpolation. Figure 5 As shown;

[0058] Step 5: Complete the continuous slices obtained in Step 4. The data is input into the continuous slice stitching module, and the initial three-dimensional volume data is constructed using equation (7) to obtain the initial three-dimensional reconstruction. ,

[0059] (7)

[0060] in, This is a continuous slice splicing module;

[0061] Step 6: Apply the initial 3D model obtained in Step 5. The input is fed into the slice background removal module, and the threshold for removing the background color is set by equation (8) to obtain the perspective 3D reconstruction. ,like Figure 4 As shown,

[0062] (8)

[0063] in, For the slice background removal module, This indicates the background threshold that is set;

[0064] Step 7: Obtain the perspective 3D from Step 6 By removing slice noise and interpolation noise using equation (9), the final perspective 3D reconstruction is obtained. ,

[0065] (9)

[0066] in, For example, the four views of the perspective 3D reconstruction are given by the joint denoising function of Gaussian smoothing and median filtering. Figure 6 As shown.

Claims

1. A deep learning-based three-dimensional reconstruction method for serous epidermis, characterized in that, Includes the following steps: Step 1: Collect stained serial sections of the sample to be tested. , This represents the sequence number, and the size of a single frame slice is... , This indicates the height and width of a single slice. Indicates the number of channels in the slice; Step 2: Process the stained serial sections obtained in Step 1 The reference point is input into the reference point analysis module, and the reference point description is obtained through equation (1). , （1） in, The reference point analysis module is calculated as shown in equation (2). （2） in, For input data, This is a 2D convolution operation. For max pooling operation, It is a non-linear activation function; Step 3: Describe the reference point obtained in Step 2 and the stained serial sections obtained in step 1 The data is input into the continuous slice registration module, and the continuous slices are spatially aligned using equation (3) to obtain aligned continuous slices. , （3） in, For the continuous slice registration module, its calculation method is shown in equation (4). （4） in, Therefore Affine parameters for image Perform the transformation. It is the maximum response filtering operation. It is a normalized exponential function. It is a linear mapping operation. ; Step 4: Align the continuous slices obtained in Step 3 The input is fed into the continuous slice interpolation module, and the original spatial position is filled in by iterative interpolation according to equation (5) to obtain the filled continuous slice. , （5） in, This is a continuous slice frame interpolation module. To complete the total number of slices and , This indicates that adjacent slices in a continuous slice are aligned and , Indicates the frame insertion position, when This indicates a slice that fills in the middle of adjacent slices. The calculation method for the continuous slice interpolation module is shown in equation (6). （6） in, It is an image feature encoding operation. It is a feature concatenation operation. It is an image feature decoding operation; Step 5: Complete the continuous slices obtained in Step 4. The data is input into the continuous slice stitching module, and the initial three-dimensional volume data is constructed using equation (7) to obtain the initial three-dimensional reconstruction. , （7） in, This is a continuous slice splicing module; Step 6: Apply the initial 3D model obtained in Step 5. The input is fed into the slice background removal module, and the threshold for removing the background color is set by equation (8) to obtain the perspective 3D reconstruction. , （8） in, For the slice background removal module, This indicates the background threshold that is set; Step 7: Obtain the perspective 3D from Step 6 By removing slice noise and interpolation noise using equation (9), the final perspective 3D reconstruction is obtained. , （9） in, This is a combined Gaussian smoothing and median filtering denoising function.

Images