Self-supervised medical image segmentation method and system based on quantum computing

Abstract

The invention discloses a self-supervised medical image segmentation method and system based on quantum computing. Accurate segmentation of a focus in a medical image is realized by designing a three-layer qutr-it neural network structure. According to the method, a quantum tri-state system is adopted as a basic calculation unit, feature transfer and aggregation are realized by using a dual quantum gate operation scheme, and the dual quantum gate operation scheme comprises a T transformation gate for feature mapping and a phase Hadamard gate for weight mapping. According to the method, a quantum fuzzy level concept and a qutr-it-based adaptive multi-class quantum Sigmoi d activation function are innovatively introduced, and a quantum state judgment mechanism based on imaginary part measurement is designed. According to the method, the dependence on the annotated data is remarkably reduced, the feature extraction capability and the segmentation precision are improved, the problem of data scarcity in medical image segmentation can be effectively solved, and a new technical scheme is provided for computer-aided diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and particularly relates to a medical image segmentation method and system based on quantum computing, especially a method and system for accurately segmenting brain lesions in medical images by using quantum three states and self-supervised learning. Background Art

[0002] With the rapid development of medical imaging, medical image analysis such as lesion segmentation tasks has become an important research direction in the field of computer-aided diagnosis (CAD). The accurate segmentation of lesions plays a crucial role in doctors' diagnosis and treatment decisions. However, the segmentation of lesions in medical images faces many challenges, such as the irregular morphology of brain lesions, the interference of image noise, the blurred boundaries between lesion tissues and healthy tissues, etc. These factors make the lesion segmentation task extremely complex and pose a severe challenge to segmentation algorithms.

[0003] Before the rise of deep learning technology, traditional image segmentation methods were difficult to effectively handle such complex medical image segmentation tasks. With the development of deep learning technology, deep learning methods represented by convolutional neural networks (CNNs) have made remarkable progress in the field of medical image segmentation. Traditional deep learning methods, such as convolutional neural networks and fully convolutional networks (FCNs), have been widely applied to medical image segmentation tasks. CNNs effectively extract image features through multiple convolutional operations and achieve high-precision segmentation tasks through layer-by-layer feature fusion. In particular, for the specific task of brain lesion segmentation, researchers have developed 3D convolutional neural networks (3D-CNNs) to process three-dimensional brain imaging data to better capture the spatial information of lesions.

[0004] Among many deep learning models, the U-Net network structure has been widely applied to medical image segmentation, especially in the field of brain lesion segmentation, due to its unique design. U-Net obtains feature information through an encoder-decoder structure and retains low-level image details through skip connections, thereby being able to accurately segment the lesion area. However, a significant drawback of these traditional deep learning methods is their extreme dependence on a large amount of labeled data, which constitutes a serious bottleneck in medical image analysis because the annotation of medical image data is usually very expensive and time-consuming.

[0005] In order to overcome the problem of scarce labeled data, self-supervised learning has gradually become an important research direction. As an unsupervised learning method, self-supervised learning is particularly suitable for situations where there is no labeled data. It pre-trains the input data by constructing proxy tasks, and then learns the potential structural features of the data through these pre-training tasks. In the field of medical imaging, self-supervised learning has been used to improve the feature learning ability of images and reduce the dependence on a large amount of labeled data. However, these self-supervised learning methods still have the problem of not being able to fully capture complex data relationships and insufficient generalization ability for complex tasks.

[0006] At the same time, quantum computing, as an emerging computing paradigm, has made significant progress in many fields in recent years. Quantum computing can process more complex data patterns than classical computers and provide more efficient computing capabilities by utilizing the superposition and entanglement characteristics of quantum states such as qubits or quantum triplets. Although quantum computing technology is still in the exploratory stage, it has shown great potential in fields such as medical image processing and deep learning acceleration.

[0007] In summary, the existing technologies have the following main problems in the task of lesion segmentation in medical images: first, traditional deep learning methods are overly dependent on a large amount of labeled data; second, the existing self-supervised learning methods have insufficient generalization ability when processing complex medical images; third, the advantages of quantum computing in medical image processing have not been fully utilized. Therefore, how to effectively solve these technical problems and improve the accuracy and efficiency of lesion segmentation in medical images has become a key scientific issue that needs to be solved in this field. Summary of the invention

[0008] In view of the problems existing in the above-mentioned prior art, the present invention adopts a quantum three-state system as the basic computing unit and utilizes a dual quantum gate operation scheme to realize feature transfer and aggregation, including a T transformation gate for feature mapping and a phase Hadamard gate for weight mapping. At the same time, the concept of quantum fuzzy hierarchy and a qutrit-based adaptive multi-class quantum Sigmoid activation function are innovatively introduced, and a quantum state determination mechanism based on imaginary part measurement is designed, thereby realizing accurate segmentation of lesions in medical images by designing a three-layer qutrit neural network structure.

[0009] Specifically, the present invention provides a self-supervised medical image segmentation method based on quantum computing. First, the input medical image data is mapped to the quantum state of a quantum three-level system, and a neural network including an input layer, an intermediate layer, and an output layer is constructed. Each layer contains multiple quantum three-level neurons, and local connections are established between adjacent neurons. Information transfer and weight mapping between layers are achieved through quantum transformation gates and quantum phase gates. A context-sensitive activation function based on quantum fuzzy theory is applied to the neurons, and the activation state is adjusted according to local quantum information. Finally, the segmentation result is obtained through quantum measurement in the output layer.

[0010] In a preferred embodiment of the present invention, the local connection adopts an 8-connected structure. Each neuron establishes symmetric connections with its eight nearest neighboring neurons around it, and the quantum state transfer of second-order neighborhood information is achieved through this structure. This structural design enables the network to capture both local fine features and context information in a larger range, improving the accuracy of segmentation.

[0011] In another preferred embodiment, the information transfer between layers includes self-forward propagation from the input layer to the output layer and self-backward propagation from the output layer to the intermediate layer. Among them, the self-backward propagation realizes parameter update through direct state transfer. Such a propagation mechanism significantly reduces the computational complexity and speeds up the convergence rate of the network.

[0012] The present invention also provides a preferred embodiment regarding the design of quantum gates. Among them, the quantum transformation gate is a T gate for realizing feature mapping, and the quantum phase gate is a Hadamard gate for mapping the connection weights between neurons. The angle parameter of the Hadamard gate is dynamically adjusted according to the relative difference of quantum information between adjacent neurons to further enhance the feature extraction ability and adaptability of the network.

[0013] Furthermore, the operation of the Hadamard gate can be expressed as:

[0014] where |θ ij > represents the quantum state relationship between the i-th neuron and the j-th neuron, and ω i,j is the rotation angle based on the local quantum information difference. This preferred example further ensures the accuracy of quantum operations.

[0015] In terms of the design of the activation function, the context-sensitive activation function accumulates the contributions of 8 fully internally connected spatially arranged neighborhood quantum three-level neurons as quantum fuzzy context-sensitive activation on the candidate neurons, and realizes feature transformation by suppressing the highest ground state as a temporary storage. This design improves the network's perception ability of complex medical image features.

[0016] Accordingly, the quantum fuzziness level information is determined by the angle of the Hadamard gate, and the angle value is equal to the sum of the quantum fuzziness measures of all adjacent neurons, which is used to adaptively adjust the sensitivity of the activation function to local features, enhancing the network's ability to recognize features in different regions.

[0017] In terms of quantum measurement, based on the quantum state evolution under the combined action of the T gate and the Hadamard gate, classification determination is performed by calculating the square value of the imaginary part of the quantum state of the output layer neurons. The imaginary part information reflects the projection characteristics of the quantum state in this basis, providing a generally more reliable classification basis.

[0018] In addition, during the information transfer process from the input layer to the middle layer, the connection weights are dynamically adjusted by the rotation angle ω, which is determined according to the complementary value of the difference in the quantum fuzziness level information of adjacent neurons, enhancing the network's adaptability to local feature differences.

[0019] Finally, the present invention also provides a system for implementing the above method. The system includes a processor and a memory. The memory is used to store computer programs, and the processor is used to execute the programs to implement any one of the above methods. The system also includes an image storage module for storing medical image data, a quantum computing module for implementing the quantum three-level neural network, and a display module for displaying the segmentation results. The processor controls the cooperation of these modules by executing programs to achieve self-supervised segmentation of medical images.

[0020] Compared with the prior art, the shallow quantum learning network combining Qutrit-inspired quantum computing and self-supervised learning proposed by the present invention can have the following beneficial technical effects:

[0021] 1. Reduce the dependence on labeled data: Through a fully self-supervised learning method, the model can automatically learn from unlabeled medical image data, reducing the dependence on a large amount of labeled data, thereby improving the scalability and adaptability of the model.

[0022] 2. Improve the accuracy of medical image segmentation, such as brain lesion segmentation: Combining the advantages of the quantum three-state (Qutrit) calculation model, providing a larger state space than traditional quantum bits (Qubits), enabling the model to better capture the subtle features of lesions in the segmentation task and improving the segmentation accuracy.

[0023] 3. Enhance the computing efficiency: Through the characteristics of quantum computing, especially the high-dimensional state of Qutrit, the computing efficiency is improved, enabling the model to process high-dimensional medical image data and enhancing the computing power of the algorithm. Brief Description of the Drawings

[0024] Figure 1It is a schematic diagram of a quantum fully self-supervised neural network architecture. Detailed implementation manners

[0025] The technical solutions of this patent will be further described in detail below in combination with specific implementation manners. It should be noted that although the following takes the segmentation of brain lesions, such as brain tumor images, as an example for detailed description, these descriptions are exemplary. Those skilled in the art know that the technical solutions are also applicable to the field of complex medical image segmentation, rather than being only applicable to the segmentation of brain lesion images. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs.

[0026] Similar to the segmentation of lesions in most medical images, brain lesion segmentation, as an important task in medical image processing, usually also faces the following challenges: The shapes, sizes, and positions of brain lesions vary greatly. The complex lesion shapes and structures make it difficult for traditional segmentation methods (especially deep learning methods based on classical computing) to handle. Existing supervised learning methods require a large amount of labeled data, and labeling medical image data is usually very expensive and time-consuming. The high cost of medical image labeling limits the establishment of large-scale data sets, thus affecting the training and generalization ability of the segmentation model.

[0027] As an emerging computing paradigm, quantum computing has made remarkable progress in many fields in recent years. By utilizing the superposition and entanglement properties of quantum states such as qubits or qutrits, quantum computing can process more complex data patterns than classical computers and provide more efficient computing power. Although quantum computing technology is still in the exploration stage, it has shown potential in fields such as medical image processing and deep learning acceleration. Therefore, the combination of quantum computing and deep learning technologies, especially the combination of Qutrit (quantum trit) in quantum computing and self-supervised learning, provides a new idea for solving the data scarcity problem in brain lesion segmentation and improving the segmentation accuracy.

[0028] The present invention constructs a quantum fully self-supervised neural network (Quantum Fully Self-Supervised Neural Network, hereinafter simply referred to as QFS-Net), which is a fully self-supervised learning network designed based on the principles of quantum computing. Its main goal is to achieve efficient medical image segmentation, such as segmenting brain lesions. The network effectively captures complex image features through modules such as feature representation of quantum states, feature transfer and aggregation, activation functions, and quantum measurements, and overcomes the limitations of traditional neural networks in small-sample learning.

[0029] Quantum computing provides new ways of implementing algorithms in the field of computing through the properties of superposition, coherence, decoherence, and entanglement in quantum mechanics. Classical computing systems use binary logic, while quantum systems typically have multiple possible states (D-level). The states of these systems are called qudits. Different from qubits with only two ground states, multi-level quantum systems (D > 2) are represented by D orthogonal ground states. For a three-level system (D = 3), the basis of each quantum trit or qutrit can be represented as {|0>, |1>, |2>}. A general pure (coherent) quantum state is a superposition of all three ground states and can be represented as

[0030] |ψ i > = a0|0> + a1|1> + a2|2>

[0031] where, |a0| 2 + |a1| 2 + |a2| 2 = 1.

[0032] The Quantum Fully Self-Supervised Neural Network (QFS-Net) effectively captures complex image features through modules such as feature representation, feature transfer and aggregation, activation functions, and quantum measurement of quantum states. Its structure consists of three layers of qutrit neurons, serving as the input layer, intermediate layer, and output layer respectively. The schematic diagram of the QFS-Net architecture as a quantum neural network model is as Figure 1 shown. Each quantum neuron describes a quantum state, which can correspond to a pixel in the image, represented by a dot in Figure 1 and its quantum state is represented by |ψ ij >, where i can represent the layer index and j represents the neuron index in that layer.

[0033] In the design of the network structure, each layer of QFS-Net organizes qutrit neurons in a fully connected manner. Preferably, the connection strength is uniformly set to 2π / 3. The grid lines in the figure show the connection relationships between neurons, adopting an 8-connected second-order neighborhood structure, that is, each quantum neuron establishes connections with its 8 nearest neighbor neurons around it. This connection structure is organized layer by layer in the underlying system and realizes the step-by-step processing of information through inter-layer propagation.

[0034] In the inter-layer propagation mechanism, connections are established between the input layer, intermediate layer, and output layer through the self-forward propagation of qutrit states. In particular, a self-reflexive propagation mechanism is also established from the output layer to the intermediate layer. This design innovatively eliminates the traditional quantum backpropagation algorithm and significantly reduces the time complexity. The self-reflexive propagation mechanism realizes parameter update through direct state transfer, avoiding the complex quantum gradient calculation process.

[0035] This three - layer structure design not only maintains the simplicity of the network structure, but also ensures that the network has sufficient feature extraction and transformation capabilities through a reasonable connection mechanism and propagation method. The 8 - connected neighborhood structure is particularly suitable for dealing with local feature associations in medical images, providing a structural basis for the accurate segmentation of brain lesions.

[0036] As Figure 1 shown, QFS - Net adopts a double - quantum - gate operation scheme to achieve feature extraction and transformation. Specifically, the qutrit neurons in each layer use the T - transformation gate for feature mapping and the phase Hadamard gate (H) to implement the mapping of interconnection weights.

[0037] In the connection mechanism of the network, the setting of the rotation angle is a key innovation point. As Figure 1 shown by the arrow in, this angle is determined according to the relative difference in quantum information between each candidate qutrit neuron and its neighboring qutrit neurons in the same layer. This angle - setting mechanism based on local information differences enables the network to adaptively adjust the intensity of feature transformation.

[0038] In terms of quantum state measurement, the present invention adopts a quantum observation process to collapse the quantum state to one of the ground states |0> or |1>, and |2> is considered a temporary state. This measurement mechanism realizes the effective conversion of quantum information into a classical segmentation result. As Figure 1 shown on the right, when the network converges, the segmentation result can be directly obtained at the output layer of QFS - Net; if it does not converge, the quantum state will be further processed. This design ensures the reliability and accuracy of the segmentation result.

[0039] This design provides a new solution for medical image segmentation by transforming complex quantum computing processes into understandable segmentation labels. The quantum computing characteristics and adaptive mechanism of the network enable it to effectively process complex features in medical images, and are particularly suitable for dealing with challenging tasks such as medical image segmentation of brain lesions.

[0040] The following further details the preferred design of QFS - Net. The corresponding quantum feature representation module realizes feature mapping by using the T - transformation gate for each layer of qutrit neurons and uses the phase Hadamard gate (H) to implement the mapping of interconnection weights. The mathematical expression of the Hadamard gate operation is:

[0041]

[0042] where |θ ij > represents the interconnection weight between the i - th neuron and the j - th neuron, ω i,jis the rotation angle based on the local quantum information difference. This quantum gate operation enables the network to effectively capture the complex feature information of the image. The role of the relative measure of quantum fuzzy information is that after adopting the relative measure, the difference between the foreground and background image pixels is clearly visible. Assume that the quantum fuzzy level information of the i-th candidate neuron and its 8-connected second-order neighborhood neurons are μ i and μ i,k , and the angle of the Hadamard gate is determined by the following formula:

[0043] ω i,k = 1 - (μ i - μ i,k ); k ∈ {1, 2,..., 8}

[0044] where k represents the index of the 8 fully interconnected spatially arranged neighborhood qutrit neurons. μ i represents the quantum fuzzy level information of the central neuron, and μ i,k represents the corresponding information of the k-th neighborhood neuron. This angle setting mechanism based on local differences ensures the accuracy of feature extraction.

[0045] In the feature transfer module, the system establishes a quantum channel for interlayer information transfer through the transformation gate (T) and the parameter γ. The contribution of the candidate quantum neurons (denoted as i') in adjacent layers, whose state can be expressed as:

[0046]

[0047] In addition, the contributions of the 8 fully interconnected spatially arranged neighborhood qutrit neurons are accumulated on the candidate qutrit neurons as quantum fuzzy context-sensitive activation (ξ i ) and represented using the Hadamard gate:

[0048]

[0049] where the angle of the Hadamard gate is defined as:

[0050]

[0051] Integrates the quantum fuzzy information of the local neighborhood.

[0052] The self-supervised forward and backward propagation of QFS-Net is guided by a novel qutrit-based adaptive multi-class quantum Sigmoid (QSig) activation function with quantum fuzzy context-sensitive threshold characteristics. The dynamic basis of the network is the two-way self-organizing propagation of qutrit states between the intermediate layer and the output layer through the update of interconnected links. The mapping is implemented using transformation gates (T) combined into a sequence, which can represent the basic input-output relationship of the network. The overall dynamic characteristics of the network are described by the following state evolution equation:

[0053]

[0054] where |ψ k l > is the output of the k-th constituent neuron in the l-th layer, and ψ k,i I-1 represents the contribution of each 8-connected neighborhood neuron of the k-th candidate neuron, that is, the layer-by-layer extraction of features is achieved by integrating information from 8 neighborhood neurons. The evolution process of features is achieved through the following formula:

[0055]

[0056] where δ k l is the rotation gate parameter, i.e., the phase factor, which is used to regulate the fine process of quantum state evolution, and its value range is from 0 to 2π. When performing a quantum observation on a quantum neuron, the quantum state is converted into a ground state. When the true result (|1>) is measured on the quantum neuron, consider the imaginary part of:

[0057]

[0058] When the imaginary part of the quantum state in the output layer satisfies the condition, the quantum neuron is measured as |1> (representing the foreground, such as the lesion area), otherwise it is |0> representing the background. This result is used for the final segmentation task. Specifically, based on the actual distribution of the data, a threshold is automatically determined using the imaginary part distributions of the foreground and background in the training data. For example, the imaginary part square value distributions of the lesion area (foreground) and non-lesion area (background) in the training set can be analyzed, and a suitable segmentation point can be selected to ensure a clear boundary between the foreground and the background.

[0059] The above has been described with respect to the segmentation of brain lesions. However, in fact, the network architecture of the present invention can be used in any medical image segmentation task. For example, in the applications of tuberous sclerosis complex (TSC) and fundus vascular segmentation, QFS-Net demonstrates significant advantages through its quantum-inspired architecture, especially in multi-modal image fusion, feature extraction, and segmentation accuracy. Tuberous sclerosis complex is a rare autosomal dominant genetic disorder caused by mutations in TSC1 or TSC2 genes. Patients often form benign tumors in multiple organs, especially in the brain. MRI imaging is usually used for the diagnosis of TSC, where T1, T2, and fluid-attenuated inversion recovery (FLAIR) sequences respectively reveal different signal characteristics of the lesions. Due to the large variability in the morphology, size, and location of the lesions caused by TSC, and the fact that the boundaries of the nodules are usually blurred, traditional manual segmentation methods often cannot provide sufficient accuracy and efficiency. However, excellent segmentation results can be obtained through the network of the present invention.

[0060] Through the deep fusion of multi-modal images, QFS-Net can effectively integrate the information from different MRI sequences (T1, T2, FLAIR), improving the recognition and segmentation accuracy of TSC lesions. The network utilizes quantum state evolution, quantum interference, and self-supervised learning mechanisms, and realizes the fine-grained extraction of lesions in different imaging modalities through quantum-inspired methods during the training process. Specifically, QFS-Net adjusts the phase factors of quantum neurons layer by layer, thereby achieving precise control of the quantum state, which helps the network avoid the accuracy loss of traditional methods when dealing with the blurred boundaries and variable morphologies of nodules. In addition, due to the characteristics of quantum interference, quantum neurons in QFS-Net can better capture local features in complex image structures, thus improving the stability and accuracy of lesion segmentation. The imaginary part of the quantum state plays a key role in the segmentation process. Through the calculation of the imaginary part of the quantum state, the model can accurately distinguish the foreground (such as the tumor area) and the background (such as the normal tissue area), effectively avoiding the problems of missed segmentation or mis-segmentation.

[0061] In the application of fundus vessel segmentation, QFS-Net also demonstrates its unique advantages. This application is based on the CHASE_DB1 dataset, which contains fundus images from healthy individuals and patients with diabetic retinopathy. The challenge of the fundus vessel segmentation task lies in the fine structure and complex morphology of the vessels. QFS-Net utilizes the 8-connected neighborhood information of quantum neurons to extract the local features of vessels layer by layer, thereby effectively capturing the complex distribution of vessels in the image. Through the quantum interference effect, QFS-Net can enhance the features of the vessel region, making the segmentation process more robust and accurate. In addition, QFS-Net uses the square value of the imaginary part as the criterion for determining the foreground and background. When processing fundus images, it can accurately segment the fundus vessel region, which is crucial for the diagnosis of ophthalmic diseases such as diabetic retinopathy.

[0062] Through application examples in various medical images, QFS-Net demonstrates its strong potential in the field of medical image segmentation, especially its advantages in multi-modal image fusion, boundary blur processing, and small sample learning. Its self-supervised learning framework based on quantum heuristic methods can achieve efficient feature extraction and segmentation tasks in complex medical images, showing the application prospects of quantum computing in medical image processing.

[0063] The present invention also provides a self-supervised medical image segmentation system based on quantum computing, including a processor, a memory, an image storage module, a quantum computing module, and a display module. Among them, the memory is used to store the computer program for implementing the method of the present invention, and this program can be compiled executable program code.

[0064] Specifically, the image storage module can be an independent storage unit for storing medical image data to be processed, such as brain lesion image data. This module can also include an image preprocessing unit for preprocessing the input medical images, such as operations like size normalization and noise removal.

[0065] The quantum computing module is the core processing unit of the system, used to implement the construction and calculation of the quantum three-level neural network. This module includes a quantum state encoding unit, a quantum gate operation unit, and a quantum measurement unit. The quantum state encoding unit is responsible for mapping the preprocessed image data into the quantum state of the quantum three-level system. The quantum gate operation unit is responsible for implementing T-gate and Hadamard gate operations to complete feature extraction and information transfer. The quantum measurement unit is responsible for performing the final quantum state measurement to obtain the segmentation result.

[0066] The display module is used to visually display the segmentation result, and various display methods can be adopted, such as displaying the segmentation result by overlaying it on the original image in different colors, or generating the contour of the segmented region for display, etc.

[0067] The processor coordinates the work of each functional module by executing the program code in the memory. First, it reads the medical image data from the image storage module; then, it controls the quantum computing module to execute the construction and calculation process of the quantum neural network; finally, it transmits the calculated segmentation result to the display module for visual display.

[0068] Through the above system implementation scheme, an efficient self-supervised medical image segmentation system is realized, which can effectively handle complex medical image processing tasks such as lesion segmentation.

[0069] In summary, the technical solution of the present invention, through the innovative design and mutual cooperation of quantum state modeling, quantum activation function design, quantum gate design, and quantum measurement mechanism, jointly constructs an efficient quantum self-supervised learning framework.

[0070] The present invention first proposes an innovative method of directly mapping image pixels to quantum states. By using qutrits (quantum three-level neurons) as basic network units, it breaks through the limitations of traditional binary neurons (0 / 1 or activated / non-activated). The system utilizes the three-level system (|0>, |1>, |2>) to significantly enhance the information expression ability, enabling the network to capture richer image features.

[0071] A unique quantum activation function is designed, that is, the present invention proposes a novel quantum fuzzy context-sensitive activation function (QSig). The uniqueness of this activation function lies in combining 8-connected neighborhood information and dynamically adjusting the activation state according to fuzzy set theory, which enhances the robustness of the segmentation result by integrating local features.

[0072] The present invention adopts two specially designed quantum gates: the transformation gate (T gate) and the phase Hadamard gate (H gate). The design of these two quantum gates simulates the interference effect in quantum mechanics, where: the T gate is responsible for completing information transformation and feature extraction; the H gate is used to adjust the connection weights between neurons to enhance the non-linear expression ability of the network. The specific design of the quantum gate operation follows the mathematical expressions described above and realizes the weight update rule through the 8-connected neighborhood structure. This design enables the network to effectively handle the correlation between local and global features.

[0073] In the output segmentation layer, the present invention designs a unique quantum state collapse and observation mechanism. The system distinguishes the lesion (foreground) and non-lesion (background) regions by calculating the square value of the imaginary part. This measurement scheme based on the imaginary part enhances the physical interpretability of the segmentation result, making the classification decision have a clear physical meaning.

[0074] The above four aspects of technological innovation are closely combined to form a complete quantum computing framework, realizing high-precision medical image segmentation while maintaining a low computational complexity.

[0075] The above are only some embodiments of the present invention. For those skilled in the art, without departing from the inventive concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A self-supervised medical image segmentation method based on quantum computing, characterized in that, Including: Mapping the input medical image data to the quantum state of a quantum three-level system, constructing a neural network with an input layer, an intermediate layer, and an output layer, where each layer contains multiple quantum three-level neurons, and establishing local connections between adjacent neurons; Implementing inter-layer information transfer and weight mapping through quantum transformation gates and quantum phase gates; Applying a context-sensitive activation function based on quantum fuzzy theory to the neurons and adjusting the activation state according to local quantum information; Obtaining the segmentation result through quantum measurement in the output layer.

2. The method according to claim 1, characterized in that: The local connection adopts an 8-connected structure, and each neuron establishes symmetric connections with its eight nearest neighboring neurons around it, and realizes the quantum state transfer of second-order neighborhood information through this structure.

3. The method according to claim 1, characterized in that: The inter-layer information transfer includes self-forward propagation from the input layer to the output layer and self-backward propagation from the output layer to the intermediate layer, where the self-backward propagation realizes parameter update through direct state transfer.

4. The method according to claim 1, wherein: The quantum transformation gate is a T gate for realizing feature mapping, and the quantum phase gate is a Hadamard gate for mapping the connection weights between neurons, where the angle parameter of the Hadamard gate is dynamically adjusted according to the relative difference of quantum information between adjacent neurons.

5. The method according to claim 4, wherein: The operation of the Hadamard gate is expressed as: where |θ ij > represents the quantum state relationship between the i-th neuron and the j-th neuron, ω i,j is the rotation angle based on the local quantum information difference.

6. The method according to claim 1, characterized in that: The context-sensitive activation function accumulates the contributions of 8 fully internally connected spatially arranged neighborhood quantum three-level neurons as quantum fuzzy context-sensitive activation on the candidate neurons, and realizes feature transformation by suppressing the highest ground state as a temporary storage.

7. The method according to claim 6, wherein: The quantum fuzzy level information is determined by the angle of the Hadamard gate, and this angle value is equal to the sum of the quantum fuzzy metrics of all adjacent neurons, which is used to adaptively adjust the sensitivity of the activation function to local features.

8. The method according to claim 1, characterized in that: The quantum measurement is based on the quantum state evolution under the joint action of the T gate and the Hadamard gate, and the classification determination is carried out by calculating the square value of the imaginary part of the quantum state of the neurons in the output layer, where the imaginary part information reflects the projection characteristics of the quantum state in this basis.

9. The method according to claim 1, characterized in that: During the information transfer process from the input layer to the intermediate layer, the connection weights are dynamically adjusted by the rotation angle ω, and this rotation angle is determined according to the complementary value of the difference in quantum fuzzy level information between adjacent neurons.

10. A self-supervised medical image segmentation system based on quantum computing, comprising a processor and a memory, the memory is used to store computer programs, and the processor is used to implement the method according to any one of claims 1 to 9 when executing the computer programs, characterized in that, Also including: An image storage module for storing medical image data; A quantum computing module for implementing the quantum three-level neural network; A display module for displaying the segmentation result; Wherein, the processor is further used to control the collaborative work of the image storage module, the quantum computing module, and the display module when executing the computer program, so as to realize the self-supervised segmentation of medical images.

Images