Quantum-inspired progressive focusing plant cell microtubule image segmentation method and system

Abstract

The present application belongs to the technical field of image processing, and particularly relates to a quantum heuristic progressive focusing plant cell microtubule image segmentation method and system. The method of the present application takes quantum heuristic progressive focusing Transformer as the backbone network, takes the feature texture extraction module as the bypass network, and loads the Hamilton growth layer at the output end of the segmentation head to form a plant cell microtubule segmentation network; each Transformer layer of the backbone network comprises, in sequence, a progressive attention layer, a full connection layer, a normalization layer, a global state cognition module, a Bell non-local module and a quantum interference gate. The multi-quantum feature coupling mechanism composed of the global state cognition module, the Bell non-local module and the quantum interference gate effectively solves the segmentation fracture and blur problem of the plant cell microtubule image under the condition of few samples due to low signal-to-noise ratio and complex structure, and improves the accuracy, topological integrity and robustness of segmentation.

Description

Technical Field

[0001] This invention belongs to the field of plant cell image processing technology, and relates to a quantum-inspired progressive focusing plant cell microtubule image segmentation method and system. Background Technology

[0002] In cellular and biomedical image analysis, the morphology and dynamic changes of plant cell microtubules are crucial for observing cellular life activities. Due to the elongated and easily broken structure of plant cell microtubules, and the challenges of weak signal, high background noise, and dense structural intersections in fluorescence microscopy, automatic segmentation has always faced significant challenges. Traditional image processing methods (such as tubular structure enhancement algorithms based on Frangi filtering) typically rely on manual parameter adjustments, are sensitive to noise, and are prone to misjudgment or breakage in complex intersection regions. With the development of deep learning, the U-Net network, with its symmetrical encoder-decoder structure and skip connections, has become a commonly used framework for microscopic image segmentation, but it still has significant shortcomings: its convolutional operations have a limited receptive field, making it difficult to capture long-distance dependencies; skip connections easily transmit background noise from the encoder to the decoder, affecting detail recovery; and the model converges slowly during training, resulting in high computational costs.

[0003] In recent years, Transformer-based architectures (such as Swin-UNet) have enhanced their ability to model global semantic information through self-attention mechanisms, which has alleviated the receptive field limitation of convolutional networks to some extent. However, in the segmentation of small and topologically complex structures such as plant cell microtubules, existing methods still generally suffer from local breaks and insufficient preservation of feature continuity, especially exhibiting poor robustness in the segmentation of plant cell microtubule images with significant noise and dense structures. Summary of the Invention

[0004] The purpose of this invention is to propose a quantum-inspired progressively focused plant cell microtubule image segmentation method and system to address the limitations of existing methods in accurately extracting and segmenting clear and continuous plant cell microtubule images in situations characterized by low signal-to-noise ratio, weak microtubule signals, and varying intracellular brightness. First, a dual-stream feature extraction framework is constructed. Multi-scale semantic features are extracted using a Progressive Focused Transformer (PFT) as the backbone network, while a separate feature texture extraction module is used to extract global high-frequency texture features of the image in parallel. Secondly, within each Transformer Layer of the PFT backbone network, a Global State Cognition Module (GCSB) is first introduced to perform global consistency verification and denoising on the output local features using the quantum phase interference principle. Subsequently, a Bell Nonlocal Module (BNLB) is connected to establish nonlocal entanglement associations based on Bell inequality on the calibrated feature map to capture the long-range topological dependence of plant cell microtubules. Then, through a quantum interference gate (QIG), the texture features extracted by the bypass are accurately injected into the deep network using a gating mechanism to compensate for the edge details lost as the network deepens. Finally, a Hamiltonian Growth Layer (HGL) is designed at the output end of the segmentation head of the plant cell microtubule segmentation network. Based on the Hamiltonian dynamics equation, the initially predicted mask is iteratively evolved and smoothed to enhance the continuity of the plant cell microtubule structure and fill in the breaks, ultimately outputting a high-precision plant cell microtubule segmentation map.

[0005] The present invention is achieved through the following technical solution.

[0006] A quantum-inspired progressive focusing plant cell microtubule image segmentation method is proposed, which uses a quantum-inspired progressive focusing Transformer as the backbone network, a feature texture extraction module as a bypass network, and loads a Hamiltonian growth layer (HGL) at the output end of the segmentation head to form a plant cell microtubule segmentation network.

[0007] Each Transformer layer of the backbone network (24 layers in total) sequentially includes a progressive attention layer, a fully connected layer, a normalization layer, a Global State Cognition Module (GCSB), a Bell Nonlocal Module (BNLB), and a Quantum Interference Gate (QIG). The input plant cell microtubule image undergoes two processing steps: one path passes through a feature texture extraction module to extract global texture features, which serve as the injection source for the QIG in the Transformer layers of the backbone network; the other path extracts shallow semantic features through convolution. In the Transformer layers, these shallow semantic features are processed sequentially through a progressive attention layer, a fully connected layer, and a normalization layer to obtain the input feature tensor, which is then used for global... The Global Consciousness Module (GCSB) uses the quantum phase interference principle to perform global consistency verification and denoising on the input feature tensor, outputting calibrated features. The Bell Nonlocal Module (BNLB) uses the CHSH inequality to obtain a Bell entangled correlation graph based on the calibrated features, capturing the long-range topological dependence of plant cell microtubules and obtaining deep semantic features. The global texture features and deep semantic features are processed by a quantum interference gate. The output features of the quantum interference gate are convolved and residually connected with the shallow semantic features. The resulting features are input to the segmentation head for processing to obtain the initial segmentation feature map. The initial segmentation feature map is processed by a Hamiltonian growth layer to obtain the final plant cell microtubule segmentation map.

[0008] Further preferably, the Global State Cognition Module (GCSB) divides the input feature tensor into a local cognitive phase branch and a global cognitive phase branch. The local cognitive phase branch first obtains local features through convolution, and then obtains the local cognitive phase through a phase scaling factor. The global cognitive phase branch first obtains global features through an average pooling layer and a multilayer perceptron, and then obtains the global cognitive phase through a phase scaling factor. Then, the global cognitive phase and the local cognitive phase are processed through a rotation gate and a measurement gate to calculate and measure the attention weight map of the two. Finally, the attention weight map is multiplied element-wise with the input feature tensor to obtain the calibrated features.

[0009] In a further preferred embodiment, the Bell nonlocal module first passes the calibrated features through three parallel convolutions to obtain the output features of three branches. Then, the output features of the first branch are transposed, and the transposed features are subjected to controlled operations on the output features of the second branch. The result is then passed through a rotation gate to obtain the original correlation degree between each pair of feature points. The original correlation degree is then normalized through a measurement gate to obtain the classical probability. The classical probability and the output features of the third branch are then used to simulate the construction form of the CHSH inequality to generate a Bell entangled correlation graph. The Bell entangled correlation graph is then convolved with the calibrated features to enhance the features. The enhanced features are then dimensionally adjusted through convolution. Finally, the features are added to the calibrated features to obtain the deep semantic features.

[0010] Further preferably, the processing procedure of the quantum interference gate includes: processing the deep semantic features and global texture features through convolution, ... The activation function and phase scaling factor are used to obtain the feature phases respectively. Then, the feature phases of the global texture features and the feature phases of the deep semantic features are passed through a rotation gate. Their phase difference is calculated in Hilbert space. The measurement is obtained by passing through a test gate to obtain the gate weights. The gate weights are multiplied element-wise with the global texture features to filter noise. Finally, they are added with the deep semantic features to realize texture injection and obtain the output features of the quantum interference gate.

[0011] In a further preferred embodiment, the Hamiltonian growth layer treats the initial segmentation feature map as a quantum system that evolves over time. The Hamiltonian time evolution operator drives the dynamic evolution iteration of the quantum system. After Hamiltonian evolution and rotation gate, the iteratively updated feature map is obtained. After the iteration ends, the final plant cell microtubule segmentation map is obtained through measurement gate.

[0012] This invention also provides a quantum-inspired progressive focusing plant cell microtubule image segmentation system, comprising:

[0013] Image input module, used to receive images of plant cell microtubules;

[0014] The segmentation module has a built-in plant cell microtubule segmentation network, which is used to execute the plant cell microtubule image segmentation method to extract and segment features from the plant cell microtubule image.

[0015] The output module is used to output the segmented image of plant cell microtubules.

[0016] The present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the plant cell microtubule image segmentation method described above.

[0017] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the plant cell microtubule image segmentation method as described above.

[0018] The present invention has the following advantages:

[0019] 1. It can achieve good segmentation results with only a small amount of plant cell microtubule image data for training, which can effectively alleviate the problem of small datasets and great difficulty in labeling plant cell microtubule image data.

[0020] 2. By accurately locking and suppressing incoherent background noise through the global state cognition module, an extremely low false positive rate and high-purity background segmentation are achieved.

[0021] 3. By using Bell nonlocal modules to establish semantic connections across the entire image space, long-distance dependencies between plant cell microtubules are captured.

[0022] 4. A phase consistency constraint is constructed using quantum interference gates, which only allow high-frequency texture features that are in phase with the core semantic features to pass through, thereby precisely blocking the transmission of background clutter. This results in extremely sharp edges on the microtubule skeleton of the plant cells generated by the model while preserving rich morphological details (such as subtle branches), ultimately improving the segmentation accuracy and topological integrity.

[0023] 5. The Hamiltonian growth layer transforms static segmentation prediction into a dynamic process that evolves over time. The Hamiltonian operator drives energy to diffuse along the microtubule structure of plant cells and morphologically enhances weak signals, thereby smoothing the microstructure and automatically repairing micro-fractures. Attached Figure Description

[0024] Figure 1 This is a network architecture diagram of a quantum-inspired progressive focusing plant cell microtubule image segmentation method.

[0025] Figure 2 This is a schematic diagram of the global state cognition module.

[0026] Figure 3 This is a schematic diagram of a Bell nonlocal module.

[0027] Figure 4 This is a schematic diagram of a quantum interference gate.

[0028] Figure 5 This is a schematic diagram of the Hamiltonian growth layer.

[0029] Figure 6 These are original plant cell microtubule image samples and corresponding manually labeled samples.

[0030] Figure 7 This is the segmentation result of the U-Net network on the microtubule image of the test plant cell.

[0031] Figure 8 The image shows the segmentation results of the plant cell microtubule segmentation network of the present invention on the microtubule image of the test plant cell. Detailed Implementation

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0033] Reference Figures 1-5This embodiment provides a quantum-inspired progressive focusing plant cell microtubule image segmentation method, which uses a quantum-inspired progressive focusing Transformer as the backbone network, a feature texture extraction module as a bypass network, and loads a Hamiltonian growth layer (HGL) at the output end of the segmentation head to form a plant cell microtubule segmentation network.

[0034] Each Transformer layer of the backbone network (24 layers in total) sequentially includes a progressive attention layer, a fully connected layer, a normalization layer, a Global State Cognition Module (GCSB), a Bell Nonlocal Module (BNLB), and a Quantum Interference Gate (QIG). The input plant cell microtubule image passes through a feature texture extraction module to extract global texture features containing rich high-frequency information; these global texture features serve as the injection source for the quantum interference gate in the Transformer layer of the backbone network. Another path extracts shallow semantic features through convolution. In the Transformer layer, these shallow semantic features sequentially pass through a progressive attention layer, a fully connected layer, and a normalization layer. The input feature tensor is obtained. The global state cognition module uses the quantum phase interference principle to perform global consistency verification and denoising on the input feature tensor, and outputs the calibrated features. The Bell nonlocal module obtains the Bell entanglement correlation graph based on the calibrated features through the CHSH inequality to capture the long-distance topological dependence of plant cell microtubules and obtain deep semantic features. The global texture features and deep semantic features are processed by the quantum interference gate. The output features of the quantum interference gate are convolved and residually connected with the shallow semantic features. The resulting features are input to the segmentation head for processing to obtain the initial segmentation feature map. The initial segmentation feature map is processed by the Hamiltonian growth layer to obtain the final plant cell microtubule segmentation map.

[0035] Plant cell microtubule images were collected and manually labeled to create an initial dataset. This dataset was then divided into training, validation, and test sets. To prevent overfitting under limited training, online data augmentation strategies, including random rotation, horizontal flipping, and illumination contrast jitter, were introduced into the training stream. The plant cell microtubule segmentation network was trained end-to-end based on the training set. The initial learning rate was set to 1e-4, and a cosine annealing hot restart strategy was used to dynamically adjust the learning rate. The training process lasted for 200 training epochs. During this period, the Dessian coefficients on the validation set were monitored, and the best-performing model weights were saved to obtain the final plant cell microtubule segmentation network.

[0036] Then, to overcome the global brightness differences and high-frequency background noise interference caused by uneven illumination and sample preparation during the fluorescence microscopy of plant cell microtubules, which severely reduce the signal-to-noise ratio of the images and make it difficult to capture weak plant cell microtubule signals, the following measures were taken. Although a progressive attention layer was introduced into the backbone network to focus on salient features, its lack of a global consistency verification mechanism made it easy to misidentify bright background noise as plant cell microtubule structures. To address this problem, this invention optimizes the underlying logic of feature extraction by embedding a global state cognition module based on the quantum superposition principle after the normalization layer of the backbone network. Figure 2 As shown, the global state cognition module divides the input feature tensor into a local cognitive phase branch and a global cognitive phase branch. The local cognitive phase branch first obtains local features through convolution, and then applies a phase scaling factor. To obtain local cognitive phase The global cognitive phase branch first obtains global features through an average pooling layer and a multilayer perceptron (MLP), and then obtains the global cognitive phase through a phase scaling factor. Then the global cognitive phase and the local cognitive phase are passed through a revolving door. The attention weight map of the two is calculated and measured by the measurement gate. Finally, the attention weight map is weighted and fused with the input feature tensor to adjust the response intensity of local features. Figure 2 middle, H represents the initial quantum state, and H represents the Hadamard gate, used to create superposition states; combined with the rotation gate... And the attention weight map of the two is obtained by measuring the gate. The processing of the input feature tensor by the global state cognition module can be expressed by formula (1):

[0037] (1);

[0038] in, Represents the input feature tensor; This is a convolution operation for local phase branches; This represents the average pooling layer operation of the global phase branch; Multilayer perceptron computation representing global phase branches; This represents the phase scaling factor; the internal phase subtraction indicates that the phase difference is calculated using a rotating gate. This represents the cosine value of the phase difference calculated after the measurement gate, i.e., the interference effect; add 1 and multiply by Used to limit the weight size to between 0 and 1; This indicates element-wise multiplication; This indicates the calibrated features.

[0039] Subsequently, the features were calibrated. Entering the Bell nonlocal module. Although the backbone network employs progressive attention layers to reduce computational complexity, this also limits its receptive field to a fixed window, failing to capture long-distance pixel dependencies across windows. To address this issue, this invention simulates the nonlocal correlation phenomenon that violates Bell's inequality in quantum mechanics, constructing a Bell nonlocal module. The Bell nonlocal module is no longer limited to a local window but establishes long-range semantic connections across the entire image by calculating the quantum entanglement strength between global feature points. This enables the network to perceive spatially separated but topologically connected microtubule segments of plant cells, effectively repairing structural breaks caused by window partitioning. Figure 3 As shown, the Bell nonlocal module first calibrates the features The output features of the three branches are obtained through three parallel convolutions. , and Then, the output features of the first branch are... Perform transpose, then use the transposed features Output characteristics of the second branch Perform controlled operations before entering the revolving door. Obtain the original correlation degree between each pair of feature points ( Figure 3 middle, This indicates the operation of the revolving door on the first branch. (This represents the revolving door operation of the second branch). Then, the original correlation degree is normalized by measuring the gate to obtain the classical probability. Finally, the classical probability and the output features of the third branch are used. The simulated CHSH inequality is used to construct hidden strong entanglement relationships between pixels and generate a Bell entanglement graph. This Bell entanglement graph is then compared with... Feature enhancement is performed on features obtained solely through convolution, followed by dimensionality adjustment of the enhanced features using convolution, and finally, the enhanced features are compared with the calibrated features. Addition is performed to preserve the original information while enhancing the related parts, outputting deep semantic features. The Bell nonlocal module calculation process can be represented as follows (2), (3), and (4):

[0040] (2);

[0041] (3);

[0042] (4);

[0043] in, , , The output characteristics of the first, second, and third branches are represented respectively; in formula (2) express Transpose; This is an activation function used to... and The result after matrix multiplication, after normalization, is in the range of 0 to 1; Formula (3) represents classical probability; Formula (3) represents the classical probability. and The calculation of the quantized primitive correlation degree is based on the construction of the CHSH inequality, and the core is through... To amplify the classical probability Small but semantically relevant feature associations This represents a Bell entanglement graph. Represents the random error term; in formula (4) This represents subtracting a threshold from the Bell entanglement graph. We obtain the strongly correlated parts; This means that the strongly correlated components selected are mapped to [0-1] to become weights; It can also be used or Alternative, as long as it is We can use only the features generated by convolution; This indicates element-wise multiplication; This represents the convolution operation of a Bell nonlocal module; This represents the deep semantic features output by the Bell nonlocal module.

[0044] As the number of network layers increases, high-frequency texture details of plant cell microtubules are often gradually lost, leading to blurred segmentation boundaries. To address this issue, this invention incorporates a quantum interference gate after the Bell nonlocal module, implementing a bypass texture injection strategy. The Bell nonlocal module, drawing inspiration from the Mach-Zehnder interferometer principle in optical experiments, first convolves the deep semantic features of the backbone network and the global texture features extracted by the texture extraction module through convolution, ... Activation function and phase scaling factor To obtain the feature phases separately, and then combine the feature phases of the global texture features Feature phase of deep semantic features Through the revolving door Then, their phase difference is calculated in Hilbert space, and the gating weight is obtained by measurement through a test gate. This gating weight is then multiplied element-wise with the global texture features to filter noise. Finally, it is added with the deep semantic features to achieve texture injection, thus obtaining the output features of the quantum interference gate. This calculation process can be expressed by the following formula (5):

[0045] (5);

[0046] in Represents the global texture features of the feature texture extraction module; This represents the deep semantic features output by the Bell nonlocal module; This indicates a convolution operation on global texture features. This represents a convolution operation on deep semantic features; Indicates the activation function; Indicates the phase scaling factor; This represents the cosine value of the phase difference calculated through the measurement gate; add 1 and multiply by This means mapping the result to [0,1]. This indicates element-wise multiplication; This represents the output characteristics of a quantum interference gate.

[0047] The feature texture extraction module uses three consecutive double-layer convolutions to extract global texture features from plant cell microtubule images.

[0048] Finally, the output characteristics of the quantum interference gate After shape adjustment via convolution, the features are residually connected with shallow semantic features. This result is input into the segmentation head for processing, yielding an initial segmentation feature map. To overcome morphological defects in microscopic imaging, such as blurred microtubules, missing tube wall pixels, and minute fractures caused by fluorescence signal attenuation, a Hamiltonian growth layer is added after the segmentation head. This layer treats the initial segmentation feature map as a time-evolving quantum system, using the Hamiltonian time evolution operator to drive the dynamic evolution and iteration of the quantum system. This simulates the natural growth mechanism of plant cell microtubules to enhance weak edge signals and automatically connect minute topological fractures. Figure 5 As shown, in the quantum system representation of the Hamiltonian growth layer, Corresponding initial segmentation feature map , Corresponding to the current step After Hamiltonian evolution and revolving door After obtaining the iteratively updated feature map, the final plant cell microtubule segmentation map is obtained through the measurement gate. This process can be represented by the following formula (6):

[0049] (6);

[0050] in Indicates the first Feature maps of step iterations Indicates the first Feature maps obtained through step-by-step iterations; Represents the Hamiltonian time evolution operator; The imaginary unit is used to introduce numerical values ​​into the complex number plane. Represents the Hamiltonian; Indicates the evolution step size; This represents the initial segmentation feature map. is the regularization coefficient, used to constrain semantic drift during the evolution process.

[0051] For ease of comparison, the original plant cell microtubule image samples and the corresponding manually labeled sample sets were analyzed as follows: Figure 6 The results show the segmentation of the U-Net network on the test plant cell microtubule images, using the weights trained on a dataset of five plant cell microtubule images. Figure 7 As shown, the plant cell microtubule segmentation network of this invention was finally trained using the same dataset of five plant cell microtubule images, and the weights were obtained. The segmentation results of the trained plant cell microtubule segmentation network on the test plant cell microtubule images are as follows. Figure 8 As shown in the comparison of several images, it can be seen that the images segmented by the U-Net network have many discontinuous and varying-thickness plant cell microtubule fragments. In contrast, although the plant cell microtubule segmentation network proposed in this invention does not achieve the same clarity as the labels, it is a visible improvement over U-Net. While it also has some discontinuities, these are significantly fewer than in the U-Net network, and the plant cell microtubule lines are thicker and smoother. This morphological continuity and smoothness has important practical significance for subsequent tasks such as plant cell microtubule counting.

[0052] Furthermore, to further verify the effectiveness of the quantum-inspired plant cell microtubule segmentation network proposed in this invention, a comparative experiment was designed. The proposed plant cell microtubule segmentation network and four other segmentation networks were trained on the same mini-batch dataset and tested on the same test set for comparison. Among them, the U-Net network, as a cornerstone of medical image segmentation, adopts a symmetrical encoder-decoder architecture and fuses deep semantics with shallow texture through skip connections, effectively preserving spatial details of the image and serving as a standard reference for CNN architectures. UPerNet (Unified Perceptual Parsing Network) is a unified perceptual framework based on Feature Pyramid (FPN). It introduces a pyramid pooling module (PPM) to aggregate contextual information at different scales, demonstrating excellent performance in handling multi-scale targets and complex scene parsing. MA-Net (Multi-scale Attention Net) focuses on capturing rich contextual dependencies through attention mechanisms, utilizing a positional attention module (PAB) and a multi-scale fusion attention module (MFAB) to adaptively integrate local and global features, significantly improving the discriminative power of feature representations. Finally, SegFormer represents the latest advancement based on Transformer. It abandons positional encoding and uses a hierarchical self-attention mechanism to directly capture long-range dependencies of images. Combined with a lightweight MLP decoder, it breaks through the limitations of the local receptive field of CNNs while maintaining high computational performance. Seven commonly used image segmentation metrics were used for evaluation: Intersection over Union (IoU), Dice, Accuracy, Precision, Negative Prediction Value (NPV), Markup Value (MK), and Matthews Correlation Coefficient (MCC). Higher values ​​of these metrics indicate better performance. Their calculation process is shown in Equations (7)-(13).

[0053] (7);

[0054] (8);

[0055] (9);

[0056] (10);

[0057] (11);

[0058] (12);

[0059] (13);

[0060] All evaluation metrics are calculated based on a pixel-level confusion matrix, including true positive (TP), true negative (TN), false positive (FP), and false negative (FN). A smoothing term is also introduced to prevent division by zero errors. .

[0061] The numerical results of each model under the same experimental conditions for each evaluation index are shown in Table 1. In order to clearly compare the performance of the models on these evaluation indexes, the values ​​of the models that performed best in the evaluation indexes were bolded, and the values ​​of the models that performed second best were underlined.

[0062] Table 1 Comparison of Model Segmentation Results

[0063]

[0064] As can be clearly seen from the data in Table 1, the model obtained by this invention outperforms the comparison model in all seven apple metrics. In the most critical metrics for segmentation quality—Dyess coefficient and Intersection over Union (IoU)—the model of this invention achieves IoU values ​​of 0.8447 and 0.7328, respectively. This indicates that the plant cell microtubule mask generated by the model of this invention has a high spatial overlap with the ground truth. Compared to SegFormer, the significant improvement in IoU demonstrates the effectiveness of the Hamiltonian growth layer in restoring the topological continuity and edge details of plant cell microtubules, overcoming the smoothing tendency of Transformer in local details. Furthermore, the difficulty in plant cell microtubule segmentation lies in high-frequency noise interference. The model of this invention achieves high precision and accuracy, while the negative prediction value reaches 0.9369. This result verifies that the noise reduction mechanism of the global cognitive state module and the quantum interference gate accurately removes incoherent background fluorescence, ensuring the purity of the segmentation results. The model's performance in the Matthews correlation coefficient demonstrates that it effectively overcomes the severe class imbalance problem. Considering these evaluation metrics, the model obtained by this invention exhibits excellent performance in both background and foreground aspects.

[0065] Another embodiment of the present invention provides a quantum-inspired progressive focusing plant cell microtubule image segmentation system, comprising:

[0066] Image input module, used to receive images of plant cell microtubules;

[0067] The segmentation module has a built-in plant cell microtubule segmentation network and is used to execute the plant cell microtubule image segmentation method of the aforementioned embodiment to extract features and segment plant cell microtubule images.

[0068] The output module is used to output the segmented image of plant cell microtubules.

[0069] Another embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the plant cell microtubule image segmentation method of the foregoing embodiments.

[0070] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the plant cell microtubule image segmentation method of the foregoing embodiments.

[0071] The above-described invention merely illustrates implementation methods of the present invention and should not be construed as limiting the scope of the invention patent, nor as imposing any form of limitation on the structure of the embodiments of the present invention. It should be noted that those skilled in the art can make various changes and improvements without departing from the concept of the embodiments of the present invention, and these all fall within the protection scope of the embodiments of the present invention.

Claims

1. A quantum-inspired progressive focusing method for microtubule image segmentation in plant cells, characterized in that, A plant cell microtubule segmentation network is formed by using a quantum-inspired progressive focusing Transformer as the backbone network, a feature texture extraction module as a bypass network, and loading a Hamiltonian growth layer at the output of the segmentation head. Each Transformer layer of the backbone network sequentially includes a progressive attention layer, a fully connected layer, a normalization layer, a global state cognition module, a Bell nonlocality module, and a quantum interference gate. The input plant cell microtubule image undergoes two processing steps: one path passes through a feature texture extraction module to extract global texture features, which serve as the injection source for the quantum interference gate in the backbone network's Transformer layer; the other path extracts shallow semantic features through convolution. In the Transformer layer, these shallow semantic features are processed sequentially through a progressive attention layer, a fully connected layer, and a normalization layer to obtain an input feature tensor. The global state cognition module uses the quantum phase interference principle to perform global consistency verification and denoising on the input feature tensor, outputting calibrated features. The Bell nonlocality module first passes the calibrated features through three parallel convolutions to obtain the output features of three branches, and then processes the output features of the first branch... The features are transposed, and then the transposed features and the output features of the second branch are subjected to controlled operations. The results are then passed through a rotation gate to obtain the original correlation between each pair of feature points. The original correlation is then normalized using a measurement gate to obtain classical probabilities. These classical probabilities and the output features of the third branch are used to simulate the construction of the CHSH inequality, generating a Bell entangled correlation graph. This graph is then convolved with the calibrated features to enhance the features. The enhanced features are then dimensionally adjusted using convolution, and finally added to the calibrated features to obtain deep semantic features. Global texture features and deep semantic features are processed through a quantum interference gate. The output features of the quantum interference gate are convolved and then residually connected with the shallow semantic features. The resulting features are input into the segmentation head for processing to obtain the initial segmentation feature map. This initial segmentation feature map is then processed through a Hamiltonian growth layer to obtain the final plant cell microtubule segmentation map.

2. The plant cell microtubule image segmentation method according to claim 1, characterized in that, The global state cognition module divides the input feature tensor into a local cognitive phase branch and a global cognitive phase branch. The local cognitive phase branch first obtains local features through convolution, and then obtains the local cognitive phase through a phase scaling factor. The global cognitive phase branch first obtains global features through an average pooling layer and a multilayer perceptron, and then obtains the global cognitive phase through a phase scaling factor. Then, the global cognitive phase and the local cognitive phase are processed through a rotation gate and a measurement gate to calculate and measure the attention weight map of the two. Finally, the attention weight map is multiplied element-wise with the input feature tensor to obtain the calibrated features.

3. The plant cell microtubule image segmentation method according to claim 1, characterized in that, The feature texture extraction module uses three consecutive double-layer convolutions to extract global texture features from plant cell microtubule images.

4. The plant cell microtubule image segmentation method according to claim 3, characterized in that, The calculation method for the Bell entanglement graph is as follows: ； in, This represents a Bell entanglement graph. Represents classical probability. Represents the random error term. This represents the output characteristics of the third branch.

5. The plant cell microtubule image segmentation method according to claim 1, characterized in that, The quantum interference gate process includes: processing deep semantic features and global texture features through convolution, ... The activation function and phase scaling factor are used to obtain the feature phases respectively. Then, the feature phases of the global texture features and the feature phases of the deep semantic features are passed through a rotation gate. Their phase difference is calculated in Hilbert space, and then the measurement is obtained through a test gate to obtain the gate weights. The gate weights are multiplied element-wise with the global texture features to filter noise. Finally, they are added with the deep semantic features to realize texture injection and obtain the output features of the quantum interference gate.

6. The plant cell microtubule image segmentation method according to claim 1, characterized in that, The Hamiltonian growth layer treats the initial segmentation feature map as a quantum system that evolves over time. It uses the Hamiltonian time evolution operator to drive the dynamic evolution iteration of the quantum system. After Hamiltonian evolution and rotation gate, the feature map is obtained after iterative updates. After the iteration ends, the final plant cell microtubule segmentation map is obtained through the measurement gate.

7. The plant cell microtubule image segmentation method according to claim 6, characterized in that, The processing procedure for the Hamiltonian growth layer is as follows: (6) in Indicates the first Feature maps of step iterations, Indicates the first Feature maps obtained through step-by-step iterations; Represents the Hamiltonian time evolution operator; Represents the imaginary unit; Represents the Hamiltonian; Indicates the evolution step size; This represents the initial segmentation feature map. is the regularization coefficient.

8. A quantum-inspired progressive focusing plant cell microtubule image segmentation system, characterized in that, include: Image input module, used to receive images of plant cell microtubules; The segmentation module has a built-in plant cell microtubule segmentation network for executing the plant cell microtubule image segmentation method as described in any one of claims 1-7, so as to extract features and segment plant cell microtubule images; The output module is used to output the segmented image of plant cell microtubules.

9. An electronic device comprising a processor, a memory, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the program, it implements the plant cell microtubule image segmentation method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the plant cell microtubule image segmentation method as described in any one of claims 1-7.

Images