Axial resolution enhancement method and system for plant leaf chloroplast three-dimensional imaging

Abstract

The invention provides a plant leaf chloroplast three-dimensional imaging axial resolution enhancement method and system, and the method comprises the steps: constructing an in-vitro model sample through fluorescent microspheres, and the in-vitro model sample comprises a non-scattering true value image and an axial stretching distortion image caused by scattering of the same visual field imaging; pre-training the deep learning network by using the in-vitro model sample, and establishing a pre-training model; constructing a plant real sample of plant leaf chloroplast, and training the pre-training model to obtain an axial resolution enhancement model; and inputting a to-be-processed chloroplast three-dimensional fluorescence image collected in a complete leaf into the axial resolution enhancement model, and outputting a chloroplast image used for accurate three-dimensional reconstruction after axial resolution enhancement. According to the method, hardware does not need to be transformed, the axial deformation of the plant leaf chloroplast in three-dimensional imaging is specifically corrected through an innovative training data preparation strategy and supervised deep learning, and the real three-dimensional form of the chloroplast is recovered.

Description

Technical Field

[0001] This invention belongs to the field of computational optical imaging and plant phenotyping technology, specifically relating to a method and system for enhancing the axial resolution of three-dimensional imaging of chloroplasts in plant leaves. Background Technology

[0002] Plant leaves are key organs for photosynthesis, and the three-dimensional structure and spatial distribution of chloroplasts directly determine the photosynthetic efficiency of a plant. However, leaf tissues (such as the epidermal wax layer, stomatal complex, and deep mesophyll cells) have extremely strong light scattering and absorption characteristics, severely limiting the imaging depth and resolution of three-dimensional fluorescence microscopy (such as three-photon microscopy). During deep imaging, excitation light and fluorescence signals suffer severe scattering, resulting in a significantly lower axial resolution than lateral resolution, exhibiting obvious axial stretching, while the signal-to-noise ratio decreases rapidly with depth. This leads to distorted chloroplast morphology and inaccurate spatial distribution information, making precise three-dimensional quantitative analysis difficult.

[0003] Existing technologies offer both hardware and software solutions to address the aforementioned problems. Hardware solutions, such as 4PI microscopy, can improve isotropic resolution, but require complex and expensive optical path modifications and are only applicable to thin samples. Software solutions, such as computational deconvolution methods, typically assume the imaging medium is isotropic and cannot effectively handle tissues like leaves, which exhibit strong anisotropic scattering characteristics. Therefore, there is an urgent need in the field for a high-precision computational method that can specifically correct axial deformation in three-dimensional imaging of chloroplasts in plant leaves without requiring hardware modifications. Summary of the Invention

[0004] The purpose of this invention is to propose a method and system for enhancing the axial resolution of three-dimensional imaging of chloroplasts in plant leaves. Without modifying the hardware, it specifically corrects the axial deformation in three-dimensional imaging of chloroplasts in plant leaves and restores the true three-dimensional morphology of chloroplasts through an innovative training data preparation strategy and supervised deep learning.

[0005] To achieve the above objectives, the technical solution of the present invention is implemented as follows: A method for enhancing axial resolution in three-dimensional imaging of chloroplasts in plant leaves, comprising: S1. Construct an in vitro model sample using fluorescent microspheres, wherein the in vitro model sample includes a scatter-free true image and an axially stretched distortion image caused by scattering from the same field of view. S2. Pre-train the deep learning network using in vitro model samples to establish a pre-trained model that maps axially stretched distorted images caused by scattering to ground truth images without scattering. S3. Construct a real plant sample of chloroplasts in plant leaves. The real plant sample includes a ground truth image of chloroplasts without scattering and an axially stretched distortion image caused by scattering, imaged in the same field of view. The scattering conditions of the axially stretched distortion image caused by scattering are consistent with those of the in vitro model sample. S4. Train the pre-trained model using real plant samples to obtain an axial resolution enhancement model; S5. Input the 3D fluorescence image of chloroplasts collected in a complete leaf into the axial resolution enhancement model, and output the chloroplast image with enhanced axial resolution for accurate 3D reconstruction.

[0006] Furthermore, in step S1, the size of the fluorescent microspheres used is similar to that of the chloroplasts, and the diameter of the fluorescent microspheres falls within the range of chloroplast diameters.

[0007] Furthermore, in step S1, the method for obtaining the scatter-free true image of the in vitro model sample includes: dispersing fluorescent microspheres in a low-scattering medium and performing three-dimensional fluorescence imaging.

[0008] Furthermore, in step S1, the acquisition of the axial stretching distortion image caused by scattering of the in vitro model sample includes: replacing the low-scattering medium with a high-scattering medium, performing three-dimensional fluorescence imaging on the same field of view, and acquiring the axial stretching image.

[0009] Preferably, the high scattering medium is a lipid droplet solution, the concentration of which is configured such that the reduced scattering coefficient of the high scattering medium matches the reduced scattering coefficient of the intact plant leaf to be studied.

[0010] Furthermore, the deep learning network described in step S2 is based on the U-Net encoder-decoder architecture. The encoder downsamples the axially stretched distortion image caused by scattering of the input layer by layer. Each downsampled feature map is fed into the stacked Transformer module, which captures the three-dimensional long-range dependency through multi-head attention to enhance global context modeling. At the jump connection between the encoder and decoder, the CBAM attention module is integrated. First, channel attention is used to generate channel weights to enhance the channel response of key features. Then, spatial attention is used to generate spatial weights to locate the core region. After double weighting, the features are fused with the upsampled features of the decoder to suppress scattering noise. The decoder upsamples the deep features processed by the Transformer layer by layer and fuses them with the encoder features optimized by CBAM to restore resolution.

[0011] Furthermore, deep learning networks are trained using a hybrid loss function: ;in Indicates the mean square error loss. express Weighting coefficients; This represents the multi-scale structural similarity loss. express Weighting coefficients; The parameters are iteratively updated using the Adam optimizer and cosine annealing learning rate via backpropagation. Loss convergence.

[0012] Furthermore, the method for constructing real plant samples of chloroplasts in plant leaves in step S3 includes: The epidermis of plant leaves was peeled off to expose mesophyll cells, and three-dimensional fluorescence imaging was performed under conditions without epidermal scattering to obtain scatter-free true images of chloroplasts. Then, the plant leaves were immersed in the same high-scattering medium used when constructing the in vitro model sample, and three-dimensional fluorescence imaging was performed on the same area to obtain images of axial stretching distortion caused by chloroplast scattering.

[0013] In another aspect, the present invention proposes an application of the above method to enhance the axial resolution of other spherical intracellular structures labeled with fluorescent proteins in plant leaves, wherein the other spherical intracellular structures include cell nuclei, fluorescently labeled auxin vesicles, etc.

[0014] Another aspect of the present invention proposes an axial resolution enhancement system for three-dimensional imaging of chloroplasts in plant leaves, comprising: In vitro model sample module: In vitro model samples are constructed using fluorescent microspheres. The in vitro model samples include scatter-free true images and axially stretched distortion images caused by scattering from the same field of view. Pre-training module: Uses in vitro model samples to pre-train the deep learning network, establishing a pre-trained model that maps axially stretched distorted images caused by scattering to ground truth images without scattering. Plant Real Sample Module: Constructs real plant samples of chloroplasts in plant leaves. The real plant samples include scatter-free ground truth images of chloroplasts and scatter-induced axial stretching distortion images imaged in the same field of view. The scattering conditions of the scatter-induced axial stretching distortion images are consistent with those of the in vitro model samples. Training module: The pre-trained model is trained using real plant samples to obtain an axial resolution enhancement model; Enhancement module: Input the 3D fluorescence image of chloroplasts collected in a complete leaf into the axial resolution enhancement model, and output the chloroplast image with enhanced axial resolution for accurate 3D reconstruction.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention is the first to specifically target plant leaves, a high-scattering medium, and solves the unique problem of 3D imaging of chloroplasts. Without modifying the hardware, it specifically corrects the axial deformation in 3D imaging of chloroplasts in plant leaves and restores the true 3D morphology of chloroplasts through an innovative training data preparation strategy and supervised deep learning.

[0016] 2. This invention uses fluorescent microspheres to construct in vitro model samples for pre-training, which simplifies the training task, reduces data costs, avoids dependence on chloroplast data in the early stage of main task training, achieves effective initialization of model parameters, accelerates main task convergence, improves the model's generalization ability, and reduces overfitting to chloroplast data.

[0017] 3. This invention constructs real plant samples of chloroplasts and cleverly obtains paired, physically realistic training data by physically peeling off the epidermis and simulating external scattering media, providing a reliable foundation for deep learning.

[0018] 4. The deep learning network used in this invention is based on the imaging characteristics of chloroplasts. It uses the three-dimensional long-range dependency modeling of Transformer to adapt to the correlation of its spatial distribution. For axial stretching distortion, Transformer is used to capture global blur patterns, and U-Net's upsampling is used to achieve physical restoration of axial resolution. At the same time, CBAM's dual attention is used for noise suppression, and the encoder's deep feature extraction is used to weaken invalid signals. This can achieve a better axial resolution enhancement effect.

[0019] 5. This invention provides a high-precision imaging tool for plant photosynthetic physiology research, chloroplast dynamic observation and three-dimensional phenotypic analysis, achieving near-isotropic three-dimensional resolution without the need for expensive hardware upgrades. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the training sample preparation in Embodiment 1 of the present invention; in, Figure 1 (a) shows the preparation of a scatterless true image of fluorescent microspheres and chloroplasts; Figure 1 (b) is the image preparation of axial stretching distortion caused by scattering of fluorescent microspheres and chloroplasts; In the figure: 1. Fluorescent microspheres; 2. Leaf; 3. Glass slide; 4. Low scattering medium; 5. High scattering medium.

[0021] Figure 2 This is a schematic diagram of deep learning network training in Embodiment 1 of the present invention.

[0022] Figure 3 These are three-dimensional images and axial half-width comparison images of the fluorescent microspheres of Embodiment 1 of the present invention before enhancement, after enhancement, and under a real label; in, Figure 3 (a1) is a three-dimensional image of the fluorescent microspheres before enhancement; Figure 3 (a2) is the full width at half maximum (FWHM) of the fluorescent microspheres before enhancement; Figure 3 (b1) is the enhanced 3D image of the fluorescent microspheres output by the pre-trained model; Figure 3 (b2) is the enhanced axial full width at half maximum (FWHM) plot of the fluorescent microspheres output by the pre-trained model; Figure 3 (c1) is the three-dimensional imaging truth map of the fluorescent microspheres; Figure 3 (c2) is the true axial full width at half maximum (FWHM) plot of the fluorescent microspheres; Figure 3 (d) shows the intensity distribution curves of the fluorescent microspheres before enhancement, after enhancement, and under the real label; Figure 3 (e) is a comparison of the axial half-width of the fluorescent microspheres before enhancement, after enhancement, and under the real label.

[0023] Figure 4 This is a comparison chart of the imaging effects and intensity distribution curves of chloroplasts in Embodiment 1 of the present invention before enhancement, after enhancement, and under a real label. in, Figure 4 (a1) is a three-dimensional image of chloroplasts before enhancement; Figure 4 (a2) is the axial full width at half maximum (FWHM) plot of chloroplasts before enhancement; Figure 4 (b1) is the enhanced 3D image of chloroplasts output by the enhancement model; Figure 4 (b2) is the full width at half maximum (FWHM) plot of the chloroplasts after enhancement, output by the enhancement model; Figure 4 (c1) is the three-dimensional imaging ground truth map of chloroplasts; Figure 4 (c2) is the true axial half-width plot of the chloroplast; Figure 4 (d) is a graph showing the intensity distribution of chloroplasts before enhancement, after enhancement, and under the true label; Figure 4 (e) is a comparison of the axial half-width of chloroplasts before enhancement, after enhancement, and under the true label.

[0024] Figure 5 This is an image showing the enhanced axial resolution of chloroplast three-dimensional imaging according to Embodiment 2 of the present invention; in, Figure 5 (a1) is a three-dimensional image of the chloroplast input enhancement model; Figure 5 (a2) is the axial full width at half maximum (FWHM) plot of the chloroplast input enhancement model; Figure 5 (b1) is a 3D chloroplast diagram output by the enhanced model; Figure 5 (b2) is the chloroplast axial full width at half maximum (FWHM) plot output by the enhanced model; Figure 5 (c) shows the intensity distribution curves before and after enhancement. Detailed Implementation

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0026] Example 1: The axial resolution enhancement method for three-dimensional imaging of plant leaf chloroplasts used in this embodiment specifically includes: I. Pre-training based on fluorescent microspheres.

[0027] (1) Training sample preparation: Prepare in vitro model samples, such as Figure 1 As shown in (a), fluorescent microspheres 1 are dispersed in a low-scattering medium 4, and a high-resolution three-dimensional image is obtained as a true label. The size of the fluorescent microspheres 1 is similar to that of chloroplasts, and their diameter should be within the typical diameter range of chloroplasts (2μm-10μm). For example, in this embodiment, green fluorescent polystyrene microspheres with a diameter of 3 μm are uniformly dispersed in an ultrasonic coupling agent, and a non-scattering image is obtained using a three-photon microscope (excitation wavelength 1300 nm, 60× water immersion objective) and a glass slide 3 as a non-scattering true image.

[0028] Subsequently, as Figure 1 As shown in (b), the medium is replaced with a high-scattering medium 5. In this embodiment, the high-scattering medium 5 uses a 20% lipid droplet solution, whose reduced scattering coefficient matches the reduced scattering coefficient of the complete plant leaf to be studied. After the replacement, the same field of view is imaged, and the axially stretched image is obtained as the axially stretched distortion image caused by scattering.

[0029] (2) Pre-training: First, the ground truth image of the fluorescent microspheres without scattering and the axial stretching distortion image caused by scattering are paired one-to-one as input-label pairs for training. After normalization preprocessing, they are divided into training set / validation set.

[0030] Constructing deep learning models based on U-Net architecture, such as Figure 2As shown, the encoder of this model downsamples the axially stretched distortion image caused by scattering layer by layer, and then feeds the feature map into a Transformer module composed of 12 stacked layers. Each layer of the Transformer module follows the process of "Layer Norm → Multi-head attention → Residual connection → Layer Norm → Multilayer Perceptron (MLP) → Residual connection". Multi-head attention captures the long-range dependencies of the three-dimensional morphology of the fluorescent microspheres, enhancing the global context modeling capability. At the same time, a convolutional block attention module (CBAM) is introduced in the jump connection part between the encoder and the decoder. First, channel attention (global average pooling + shared fully connected layer + sigmoid activation) is used to generate channel weights, amplifying the channel response of key features such as the edge of the fluorescent microsphere. Then, spatial attention (channel dimension pooling + convolutional layer + sigmoid activation) is used to generate spatial weights, locating the core region of the fluorescent microsphere. After double weighting, the weights are fused with the upsampled features of the decoder, effectively suppressing scattering background noise.

[0031] Training uses a hybrid loss function: ;in Indicates the mean square error loss. express Weighting coefficients; This represents the multi-scale structural similarity loss. express Weighting coefficients; In this embodiment, If we take the value 0.5 respectively, then the loss function is: (1); in, ; N is the number of pixels. , These are the ground truth pixels and the model-predicted pixels, respectively. ; K represents the scale; in this embodiment, three scales are used. , , These represent brightness, contrast, and structural similarity at the k-th scale, respectively. Scale weights are used. End-to-end training is performed using the Adam optimizer, with an initial learning rate of 0.001, dynamically adjusted using a cosine annealing learning rate strategy; the batch size is set to 16, and training continues until the training set loss fluctuates less than 1×e for 5 consecutive epochs. -4 If the PSNR and SSIM metrics on the validation set remain stable without decreasing, then the pre-trained model is obtained.

[0032] (3) Validation of pre-trained model results: such as Figure 3The image shows a comparison of the three-dimensional imaging and axial full width at half maximum (FWHM) of the fluorescent microspheres before enhancement, after enhancement, and under a real label. in, Figure 3 (a1) shows the three-dimensional imaging of the fluorescent microspheres before enhancement and Figure 3 (a2) shows the axial full width at half maximum (FWHM) of the fluorescent microspheres before enhancement; after inputting into the pre-trained model, the following is obtained: Figure 3 (b1) shows the enhanced 3D image of the fluorescent microspheres output by the pre-trained model and... Figure 3 (b2) shows the enhanced axial full width at half maximum (FWHM) plot of the fluorescent microspheres output by the pre-trained model; and then compared with... Figure 3 (c1) shows the three-dimensional imaging truth map of the fluorescent microspheres and Figure 3 By comparing the true axial half-width and full width at half-maximum plots of the fluorescent microspheres shown in (c2), we obtain... Figure 3 (d) shows the intensity distribution curves before enhancement, after enhancement, and under the true label; and Figure 3 (e) shows a comparison of the axial half-height and full width before enhancement, after enhancement, and under the true label.

[0033] As can be seen, the average full width at half maximum (FWHM) of the fluorescent microspheres in the unenhanced image was 5.3 μm, which was reduced to 4.0 μm after prediction by the pre-trained model. This is basically consistent with the true value under the real label, which proves the effective correction ability of this method and the pre-trained model for axial stretching of quasi-spherical structures and the effective enhancement ability for axial resolution.

[0034] II. Enhanced axial resolution in chloroplast imaging.

[0035] (1) Training sample preparation: Prepare real plant samples; select healthy leaves, such as... Figure 1 As shown in (a), the upper epidermis of leaf 2 was carefully peeled off using fine tweezers to expose the underlying mesophyll cells. A clear three-photon image of the chloroplasts was obtained using a three-photon microscope (excitation wavelength 1600 nm, 60× water immersion objective) and slide 3 as the scatter-free true image of the chloroplasts. Subsequently, the leaf 2 sample was immersed in the same 20% lipid drop solution, and the same field of view was imaged again to obtain an image of axial stretching distortion caused by scattering of the chloroplasts under scattering conditions.

[0036] (2) Training the augmentation model: First, the ground truth image of chloroplasts without scattering and the axially stretched distortion image caused by scattering are paired one-to-one as input-label pairs for training. After normalization preprocessing, they are divided into training set / validation set of chloroplast images.

[0037] The pre-trained model is trained on chloroplast images using the training / validation set of the chloroplast images to obtain the final enhanced model.

[0038] (3) Enhanced model result validation: like Figure 4 The image shows a comparison of 3D images and axial half-width at half-height of chloroplasts before enhancement, after enhancement, and under the real label. in, Figure 4 (a1) shows a three-dimensional image of chloroplasts before enhancement and Figure 4 (a2) shows the axial full width at half maximum (FWHM) of the chloroplasts before enhancement; after inputting the enhancement model, the following is obtained: Figure 4 The 3D graph output by the enhanced model shown in (b1) is the same as... Figure 4 (b2) shows the axial half-height full width plot output by the enhanced model; then compared with... Figure 4 (c1) shows the three-dimensional imaging ground truth map of chloroplasts and Figure 4 By comparing the true axial half-width and full width at half-height plot of the chloroplast shown in (c2), we obtain... Figure 4 (d) shows the intensity distribution curves before enhancement, after enhancement, and under the true label; and Figure 4 (e) shows a comparison of the axial half-width at half-height before enhancement, after enhancement, and under the true label. The comparison shows that in the image output by the enhancement model, adjacent chloroplasts that were originally blurred and stuck together due to axial stretching are clearly distinguishable, the axial morphology is restored to reality, and its fluorescence intensity distribution curve is highly consistent with the true label.

[0039] Example 2: This embodiment is an application example of the enhanced model obtained in Embodiment 1.

[0040] (1) Sample preparation: The leaf was cut off and treated by solution-assisted immersion to increase the imaging depth. The treated leaf was fixed on the stage and three-dimensional three-photon fluorescence imaging was performed to obtain a three-dimensional image.

[0041] (2) Use the enhancement model trained in Example 1 to perform axial resolution enhancement processing on the entire three-dimensional image stack.

[0042] (3) The results are as follows Figure 5 As shown: Figure 5 The three-dimensional imaging diagram shown in (a1) and Figure 5 The axial half-height full width plot input enhancement model shown in (a2) yields the following results: Figure 5 (b1) shows the 3D chloroplast plot output by the enhanced model and Figure 5 (b2) shows the chloroplast axial full width at half maximum (FWHM) plot output by the enhanced model; comparison yields... Figure 5(c) shows the intensity distribution curves before and after enhancement. It can be seen that the axial resolution of the 3D distribution map of chloroplasts is improved after the enhancement model is processed, which can accurately distinguish axially adjacent chloroplasts and provide reliable data for downstream chloroplast phenotypic analysis.

[0043] Example 3: This embodiment proposes an axial resolution enhancement system for three-dimensional imaging of chloroplasts in plant leaves, comprising: In vitro model sample module: In vitro model samples are constructed using fluorescent microspheres. The in vitro model samples include scatter-free true images and axially stretched distortion images caused by scattering from the same field of view. Pre-training module: Uses in vitro model samples to pre-train the deep learning network, establishing a pre-trained model that maps axially stretched distorted images caused by scattering to ground truth images without scattering. Plant Real Sample Module: Constructs real plant samples of chloroplasts in plant leaves. The real plant samples include scatter-free ground truth images of chloroplasts and scatter-induced axial stretching distortion images imaged in the same field of view. The scattering conditions of the scatter-induced axial stretching distortion images are consistent with those of the in vitro model samples. Training module: The pre-trained model is trained using real plant samples to obtain an axial resolution enhancement model; Enhancement module: Input the 3D fluorescence image of chloroplasts collected in a complete leaf into the axial resolution enhancement model, and output the chloroplast image with enhanced axial resolution for accurate 3D reconstruction.

[0044] In the in vitro model sample module, the size of the fluorescent microspheres used is similar to that of chloroplasts, and the diameter of the fluorescent microspheres falls within the range of chloroplast diameter.

[0045] In the in vitro model sample module, the method for obtaining the scatter-free true image of the in vitro model sample includes: dispersing fluorescent microspheres in a low-scattering medium and performing three-dimensional fluorescence imaging.

[0046] In the in vitro model sample module, the acquisition of axial stretching distortion images caused by scattering of in vitro model samples includes: replacing the low-scattering medium with a high-scattering medium, performing three-dimensional fluorescence imaging on the same field of view, and acquiring axial stretching images.

[0047] The high scattering medium is a lipid droplet solution, the concentration of which is configured such that the reduced scattering coefficient of the high scattering medium matches the reduced scattering coefficient of the intact plant leaf to be studied.

[0048] The deep learning network described in the pre-training module is based on the U-Net encoder-decoder architecture. The encoder downsamples the axially stretched distortion image caused by scattering of the input layer by layer. Each downsampled feature map is fed into a stacked Transformer module, which captures long-range 3D dependencies through multi-head attention to enhance global context modeling. At the jump connection between the encoder and decoder, a CBAM attention module is integrated. First, channel attention is used to generate channel weights to enhance the channel response of key features. Then, spatial attention is used to generate spatial weights to locate the core region. After double weighting, the features are fused with the upsampled features of the decoder to suppress scattering noise. The decoder upsamples the deep features processed by the Transformer layer by layer and fuses them with the encoder features optimized by CBAM to restore resolution.

[0049] The training of deep learning networks in the pre-training module uses a hybrid loss function: ;in Indicates the mean square error loss. express Weighting coefficients; This represents the multi-scale structural similarity loss. express Weighting coefficients; The parameters are iteratively updated using the Adam optimizer and cosine annealing learning rate via backpropagation. Loss convergence.

[0050] The construction of real plant samples, specifically plant leaf chloroplasts, in the real plant sample module includes: The epidermis of plant leaves was peeled off to expose mesophyll cells, and three-dimensional fluorescence imaging was performed under conditions without epidermal scattering to obtain scatter-free true images of chloroplasts. Then, the plant leaves were immersed in the same high-scattering medium used when constructing the in vitro model sample, and three-dimensional fluorescence imaging was performed on the same area to obtain images of axial stretching distortion caused by chloroplast scattering.

[0051] The axial resolution enhancement system for three-dimensional imaging of chloroplasts in plant leaves proposed in this embodiment can realize the axial resolution enhancement method for three-dimensional imaging of chloroplasts in plant leaves described in Embodiment 1, and has the same technical effect as Embodiment 1.

[0052] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for enhancing axial resolution in three-dimensional imaging of chloroplasts in plant leaves, characterized in that, include: S1. Construct an in vitro model sample using fluorescent microspheres, wherein the in vitro model sample includes a scatter-free true image and an axially stretched distortion image caused by scattering from the same field of view. S2. Pre-train the deep learning network using in vitro model samples to establish a pre-trained model that maps axially stretched distorted images caused by scattering to ground truth images without scattering. S3. Construct a real plant sample of chloroplasts in plant leaves. The real plant sample includes a ground truth image of chloroplasts without scattering and an axially stretched distortion image caused by scattering, imaged in the same field of view. The scattering conditions of the axially stretched distortion image caused by scattering are consistent with those of the in vitro model sample. S4. Train the pre-trained model using real plant samples to obtain an axial resolution enhancement model; S5. Input the 3D fluorescence image of chloroplasts collected in a complete leaf into the axial resolution enhancement model, and output the chloroplast image with enhanced axial resolution for accurate 3D reconstruction.

2. The method for enhancing axial resolution in three-dimensional imaging of chloroplasts in plant leaves according to claim 1, characterized in that, In step S1, the size of the fluorescent microspheres used is similar to that of chloroplasts, and the diameter of the fluorescent microspheres falls within the range of chloroplast diameters.

3. The method for enhancing axial resolution in three-dimensional imaging of chloroplasts in plant leaves according to claim 1, characterized in that, In step S1, the method for obtaining the scatter-free true image of the in vitro model sample includes: dispersing fluorescent microspheres in a low-scattering medium and performing three-dimensional fluorescence imaging.

4. The method for enhancing axial resolution in three-dimensional imaging of chloroplasts in plant leaves according to claim 3, characterized in that, In step S1, the acquisition of the axial stretching distortion image caused by scattering of the in vitro model sample includes: replacing the low-scattering medium with a high-scattering medium, performing three-dimensional fluorescence imaging on the same field of view, and acquiring the axial stretching image.

5. The method for enhancing axial resolution in three-dimensional imaging of chloroplasts in plant leaves according to claim 4, characterized in that, The high scattering medium is a lipid droplet solution, the concentration of which is configured such that the reduced scattering coefficient of the high scattering medium matches the reduced scattering coefficient of the intact plant leaf to be studied.

6. The method for enhancing axial resolution in three-dimensional imaging of chloroplasts in plant leaves according to claim 1, characterized in that, The deep learning network described in step S2 is based on the U-Net encoding and decoding architecture. The encoder downsamples the axially stretched distortion image caused by scattering of the input layer by layer. The downsampled feature map of each layer is fed into the stacked Transformer module, which captures the three-dimensional long-range dependency through multi-head attention and strengthens the global context modeling. At the jump connection between the encoder and decoder, a CBAM attention module is integrated to first generate channel weights with channel attention, thereby enhancing the channel response of key features. Spatial attention is then used to generate spatial weights to locate the core region. After double weighting, these weights are fused with the upsampled features from the decoder to suppress scattering noise. The decoder upsamples the deep features processed by the Transformer layer by layer and fuses them with the encoder features optimized by CBAM to restore the resolution.

7. The method for enhancing axial resolution in three-dimensional imaging of chloroplasts in plant leaves according to claim 6, characterized in that, Deep learning networks are trained using a hybrid loss function: ;in Indicates the mean square error loss. express Weighting coefficients; This represents the multi-scale structural similarity loss. express Weighting coefficients; The parameters are iteratively updated using the Adam optimizer and cosine annealing learning rate via backpropagation. Loss convergence.

8. The method for enhancing axial resolution in three-dimensional imaging of chloroplasts in plant leaves according to claim 1, characterized in that, The method for constructing real plant samples of chloroplasts in plant leaves in step S3 includes: The epidermis of plant leaves was peeled off to expose mesophyll cells, and three-dimensional fluorescence imaging was performed under conditions without epidermal scattering to obtain scatter-free true images of chloroplasts. Then, the plant leaves were immersed in the same high-scattering medium used when constructing the in vitro model sample, and three-dimensional fluorescence imaging was performed on the same area to obtain images of axial stretching distortion caused by chloroplast scattering.

9. The application of the axial resolution enhancement method for three-dimensional imaging of chloroplasts in plant leaves as described in any one of claims 1-8, applied to enhance the axial resolution of other spheroidal intracellular structures in plant leaves labeled with fluorescent proteins.

10. An axial resolution enhancement system for three-dimensional imaging of chloroplasts in plant leaves, characterized in that, include: In vitro model sample module: In vitro model samples are constructed using fluorescent microspheres. The in vitro model samples include a scatter-free true image and an axially stretched distortion image caused by scattering, imaged in the same field of view. Pre-training module: Uses in vitro model samples to pre-train the deep learning network, establishing a pre-trained model that maps axially stretched distorted images caused by scattering to ground truth images without scattering. Plant Real Sample Module: Constructs real plant samples of chloroplasts in plant leaves. The real plant samples include scatter-free ground truth images of chloroplasts and scatter-induced axial stretching distortion images imaged in the same field of view. The scattering conditions of the scatter-induced axial stretching distortion images are consistent with those of the in vitro model samples. Training module: The pre-trained model is trained using real plant samples to obtain an axial resolution enhancement model; Enhancement module: Input the 3D fluorescence image of chloroplasts collected in a complete leaf into the axial resolution enhancement model, and output the chloroplast image with enhanced axial resolution for accurate 3D reconstruction.

Images