Oral cancerous region boundary identification system and method based on photoacoustic spectroscopic imaging
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XUZHOU MEDICAL UNIVERSITY
- Filing Date
- 2026-05-12
- Publication Date
- 2026-06-26
Smart Images
Figure CN122286191A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing and intelligent diagnostic technology, specifically to a system and method for identifying the boundary of oral cancerous regions based on photoacoustic spectral imaging. Background Technology
[0002] Photoacoustic imaging technology has demonstrated unique advantages in the detection of oral cancer. It generates ultrasound signals by exciting tissue with a laser, combining optical absorption and acoustic propagation characteristics to achieve high-resolution imaging. However, current technologies still face the following challenges in practical clinical applications: Traditional photoacoustic imaging systems often use single-wavelength laser excitation, which makes it difficult to cover the differential absorption spectra between cancerous and normal tissues, resulting in limited optical contrast. At the same time, the complex anatomical structures of the oral cavity (such as saliva and blood vessels) introduce a large amount of random noise. Existing global threshold denoising methods often suffer from loss of feature information or noise residue because they cannot adapt to the non-stationary characteristics of the signal.
[0003] While numerical simulation methods based on the wave equation can improve imaging accuracy, the computational complexity of solving the full-dimensional sparse matrix increases exponentially. Current strategies that directly employ GPU parallel acceleration are limited by memory capacity and communication latency, making it difficult to meet the millisecond-level response requirements of intraoperative real-time imaging and severely restricting clinical applicability.
[0004] For intraoperatively acquired time-series data, mainstream tensor decomposition methods (such as CP decomposition) do not consider the causal constraints of the time dimension, allowing data from future time points to participate in the calculation of current features during the update process. This information leakage can lead to non-physical oscillations in boundary recognition results, especially in surgical scenarios with significant tissue deformation, resulting in misleading outputs. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a system and method for identifying the boundary of oral cancerous regions based on photoacoustic spectral imaging, which solves the core defects of existing photoacoustic imaging technology in oral cancer detection, such as low signal-to-noise ratio, large computational delay, temporal logic misalignment, and poor model generalization ability.
[0006] To achieve the above objectives, the present invention provides the following technical solution: an oral cancerous region boundary recognition system based on photoacoustic spectral imaging, comprising: The signal acquisition module is used to generate spatiotemporal-spectral three-dimensional data by exciting and receiving photoacoustic signals through multi-wavelength lasers. The physical modeling module is used to discretize the wave equation to generate a sparse matrix and compress the matrix dimension through subspace projection. The tensor decomposition module is used to perform dynamic tensor decomposition on preprocessed data and apply time causality constraints. The network training module is used to jointly optimize the physical constraints and segmentation network parameters through a meta-learning framework; The real-time recognition module is used to deploy the trained network to edge computing devices for intraoperative dynamic boundary correction. The output of the physical modeling module is connected to the tensor decomposition module, and the output of the tensor decomposition module is connected to the network training module.
[0007] Preferably, the signal acquisition module includes an adaptive wavelet threshold denoising unit, which separates noise components through discrete wavelet transform and dynamically adjusts the threshold according to the noise variance.
[0008] Preferably, the physical modeling module compresses the sparse matrix through the following steps: Constructing the basis matrix of the Krylov subspace ,in The sparse matrix generated by discretizing the wave equation. The vector of heat source terms; Projecting the sparse matrix onto a low-dimensional subspace generates a compressed matrix. , This is the matrix transpose operator.
[0009] Preferably, the tensor decomposition module uses a lower triangular mask matrix. Applying a time causality constraint, satisfying: The time dimension factor matrix is updated to ,in It represents the Hadamardi (or Hadama) stack; Mask matrix The lower triangular element is 1, and the rest are 0.
[0010] Preferably, the tensor decomposition module dynamically adjusts the weights of the factor matrix based on spectral-spatial correlation.
[0011] Preferably, the network training module uses an alternating direction optimization strategy to perform meta-learning task updates in the outer loop and combines projective gradient descent in the inner loop to ensure that the parameters satisfy the wave equation constraints.
[0012] Preferably, the network training module includes a dual-branch structure: The first branch takes the compressed sparse matrix as input and outputs the sound pressure field prediction. The second branch takes the core features obtained from tensor decomposition as input and outputs a boundary probability map.
[0013] Preferably, the real-time recognition module uses an FPGA chip to realize pipelined parallel computation of tensor decomposition and network inference.
[0014] Preferably, the real-time recognition module dynamically updates the tensor decomposition results based on the newly acquired signals during the operation and re-solves the network output to correct the boundaries.
[0015] Methods for identifying the boundaries of oral cancerous regions based on photoacoustic spectroscopy imaging include: S1. Acquire multispectral photoacoustic signals and perform adaptive wavelet denoising; S2. Based on the discretization of the wave equation, a sparse matrix is generated and compressed by subspace projection. S3. Perform dynamic tensor decomposition on the denoised data and apply time causality constraints; S4. Jointly train the physical constraint and segmentation network using a meta-learning framework; S5. Deploy the network to edge devices for real-time intraoperative boundary identification and dynamic updates.
[0016] This invention provides an intelligent boundary recognition system and method for oral cancerous regions based on photoacoustic spectral imaging, which has the following beneficial effects: 1. This invention significantly enhances the contrast of light absorption characteristics in cancerous regions by combining multi-wavelength laser excitation with adaptive wavelet threshold denoising. Compared to traditional fixed-wavelength light sources and global threshold denoising, it effectively overcomes the signal-to-noise ratio fluctuation problem caused by tissue heterogeneity and eliminates random noise interference introduced by complex structures such as saliva and blood vessels.
[0017] 2. This invention introduces Krylov subspace compression technology to project the discrete matrix of the wave equation into a low-dimensional space for solution. Compared with existing full-dimensional matrix operation methods, it reduces the computational complexity by 1-2 orders of magnitude while ensuring the accuracy of sound field reconstruction, thus overcoming the technical problem that traditional photoacoustic imaging systems cannot meet the real-time requirements during surgery.
[0018] 3. This invention, based on dynamic Tucker decomposition and lower triangular mask constraints, forces the time dimension to rely solely on historical data to generate features. Addressing the shortcomings of existing dynamic tensor decomposition methods that suffer from future information leakage, this invention resolves the boundary misjudgment problem caused by temporal logic misalignment in intraoperative continuous acquisition scenarios.
[0019] 4. This invention employs a dual-branch network combined with a meta-learning framework, achieving for the first time joint optimization of physical modeling and data-driven features. Compared to a single data-driven model, it overcomes the problem of poor model generalization ability when training data is insufficient, enabling the system to quickly adapt to the differences in tissue characteristics among different patients. Attached Figure Description
[0020] Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] Example: Please refer to the appendix Figure 1 This invention provides a system for identifying the boundary of oral cancerous regions based on photoacoustic spectral imaging, comprising: The signal acquisition module is used to generate spatiotemporal-spectral three-dimensional data by exciting and receiving photoacoustic signals through multi-wavelength lasers. In this invention, the signal acquisition module is used to acquire and preprocess multispectral photoacoustic signals from oral tissues. This module includes a multi-wavelength laser excitation unit, an ultrasonic transducer array receiving unit, and an adaptive wavelet threshold denoising unit. Its core function is to generate high signal-to-noise ratio spatiotemporal-spectral three-dimensional data, providing a foundation for subsequent physical modeling and feature extraction. The technical implementation is described in detail below step by step: First, a multi-wavelength laser excitation model is established, and a laser beam covering the absorption spectrum of a specific biological tissue is emitted using a tunable pulsed laser. Preferably, the laser wavelength range includes 650 nm to 1300 nm, covering the light absorption characteristic peaks of biological tissues such as hemoglobin, water molecules, and lipids. The laser pulse width is set to less than or equal to 10 nanoseconds to balance signal excitation efficiency with the risk of tissue thermal damage. The pulse energy is preferably controlled below 10 millijoules to avoid interference from nonlinear optical effects. The laser beam is guided and focused onto the target area via an optical fiber, exciting the tissue to generate a photoacoustic signal.
[0023] The photoacoustic signal is then received via a ring-shaped ultrasonic transducer array. The transducer array preferably comprises 64 elements arranged uniformly in a ring to ensure multi-angle signal acquisition. The center frequency of each element is set to 5MHz, with a bandwidth coverage of at least 80% to accommodate the differences in acoustic characteristics of tissues at different depths. The time sampling rate of the received signal is preferably no less than 50MHz to ensure a time resolution on the order of 20 nanoseconds. An initial three-dimensional data cube is generated through spatial-temporal synchronous sampling. ,in Indicates the spatial dimension (number of array elements). Indicates the number of time sampling points. This indicates the number of spectral channels (consistent with the number of laser wavelengths).
[0024] For the original signal Noise suppression is performed. First, a discrete wavelet transform model is established to decompose the three-dimensional data into multi-scale coefficients channel by channel along the time dimension. Preferably, the Daubechies 4 wavelet basis function is used, with a decomposition level of 3 to match the transient characteristics of the photoacoustic signal. For the decomposition coefficients W(X) of each channel, its noise variance estimate is calculated: ; in This represents the detail coefficients in the first horizontal and vertical directions. This is for median retrieval. The global threshold is dynamically calculated based on the noise variance. ; In the formula For signal length, Represents the natural logarithm. Apply a hard thresholding function to each wavelet coefficient: ; Finally, the denoised signal is reconstructed using inverse wavelet transform. : ; This algorithm can effectively separate noise components while preserving signal characteristics, providing high-quality input for subsequent processing.
[0025] The multi-wavelength laser excitation unit ensures that the photoacoustic signal contains the differentiated optical properties of cancerous and normal tissues by selecting wavelengths that cover the absorption spectrum of biological features; the high spatiotemporal sampling rate design of the ultrasonic transducer array can accurately capture the differences in sound wave propagation paths, providing spatial positioning basis for subsequent physical modeling; the adaptive wavelet threshold denoising eliminates random noise introduced by the complex anatomical structure of the oral environment (such as saliva and blood vessels) through multi-scale decomposition and dynamic threshold adjustment, while avoiding signal distortion caused by excessive smoothing.
[0026] Formula variable definition: : Original spatiotemporal-spectral 3D data cube; Discrete wavelet transform operator; : First layer horizontal-vertical detail coefficients; Noise variance estimate; Adaptive global threshold; Hard threshold function; X̃: The denoised 3D data cube; Implementation process sequence structure: First, a photoacoustic signal is generated through multi-wavelength laser excitation. Then, an ultrasonic transducer array performs spatiotemporal sampling to acquire the original three-dimensional data. Next, wavelet multi-scale decomposition is performed on the data to calculate the noise variance and dynamically determine the threshold. Finally, threshold processing and signal reconstruction are used to output the denoised data X̃. This process ensures that the data output by the signal acquisition module meets the stringent signal-to-noise ratio and resolution requirements of subsequent modules.
[0027] The above implementation method can achieve the following effects: Multidimensional spectral information is obtained by multi-wavelength laser excitation, thereby enhancing the light absorption contrast of cancerous areas; The adaptive denoising algorithm based on wavelet transform can effectively suppress environmental noise and inherent equipment noise; Dynamic threshold calculation adapts to the differences in tissue characteristics among different patients, avoiding overfitting or underfitting caused by fixed thresholds.
[0028] The denoised data X̃ is directly input into the physical modeling module for wave equation discretization and sparse matrix generation. The noise suppression capability of the signal acquisition module directly affects the numerical stability of the physical modeling module, and thus determines the accuracy of subsequent tensor decomposition and network training.
[0029] The physical modeling module is used to discretize the wave equation to generate a sparse matrix and compress the matrix dimension through subspace projection. In this invention, the physical modeling module transforms the physical laws governing the propagation of photoacoustic signals into a computable mathematical model and reduces computational complexity through matrix compression techniques. This module generates a sparse matrix by discretizing the wave equation and achieves dimensionality compression using a subspace projection method, providing physical constraints for subsequent tensor decomposition and network training. The technical implementation is described in detail below step by step: First, a non-homogeneous wave equation model is established to describe the propagation process of photoacoustic signals in tissue. For example, the equation takes the form: ; in For sound pressure field, To organize the speed of sound, The sign of the partial derivative. It is a spatial position vector. For time variables, The light absorption coefficient is... This is the heat source term. To transform the continuous equation into a discrete form, an implicit finite difference method is preferably used for space-time domain discretization. On a three-dimensional spatial grid, the Laplace operator is... Discretized into a second-order derivative matrix in space The second derivative term of time is discretized into a matrix. This yields a sparse linear system of equations: ; in Let be a sparse coefficient matrix, and b be a discretized vector of the heat source term. The boundary conditions are preferably embedded in the matrix using the Perfect Match Layer (PML) method. To simulate a non-reflective boundary.
[0030] To reduce the sparsity of the matrix The solution complexity is reduced, and dimensionality compression is further achieved through Krylov subspace projection. First, the Krylov subspace basis matrix is constructed: ; in As the initial vector, Let be the subspace dimension (preferably not exceeding 50). Then, the original equation is projected onto the lower-dimensional subspace to obtain the compressed matrix: ; Simultaneously, the heat source term is projected as , This is the matrix transpose operator. The compressed equation is in the form of: ; A low-dimensional solution is obtained by solving this equation. Then through Reconstruct the original spatial sound pressure field.
[0031] Discretization of the wave equation transforms the physical laws of photoacoustic propagation into matrix operations, ensuring that subsequent network training conforms to the deterministic constraints of sound field propagation; Krylov subspace projection captures the matrix... The main characteristic modes are identified, significantly reducing the computational load and meeting real-time requirements; boundary condition embedding avoids acoustic reflection errors caused by traditional fixed boundaries, improving numerical stability.
[0032] Formula variable definition: : Sound pressure field vector; : Spatial Laplace discrete matrix; : Time-difference matrix; : Sparse coefficient matrix; : Discrete vector of heat source term; : Krylov subspace basis matrix; : The compressed low-dimensional matrix; : Solution vector of low-dimensional subspace; Implementation process sequence structure: First, the wave equation is discretized using the implicit finite difference method, generating a sparse matrix containing boundary conditions. With heat source vector Subsequently based on the initial vector Iterative generation of Krylov subspace basis matrices Next, the original equation is projected into a lower-dimensional space and the compression equation is solved; finally, the sound pressure field is reconstructed. The physical constraints are input to the tensor decomposition module.
[0033] The above implementation method can achieve the following effects: By discretizing the wave equation, physical laws are embedded into the mathematical model to ensure the physical rationality of the sound field prediction. Krylov subspace projection reduces the matrix dimension, thereby reducing the storage and computational overhead of large-scale sparse matrices. Boundary condition embedding effectively suppresses boundary reflection errors in numerical simulations and improves the accuracy of sound field reconstruction.
[0034] Compression matrix With projection basis matrix The output is fed to the network training module as input to the physical constraint branch; the sound pressure field is reconstructed. This was used to verify the feature extraction performance of the tensor decomposition module. The numerical stability of the physical modeling module directly affects the convergence of subsequent network training and the accuracy of boundary recognition.
[0035] The tensor decomposition module is used to perform dynamic tensor decomposition on preprocessed data and apply time causality constraints. In this invention, the tensor decomposition module is used to extract low-rank features from preprocessed photoacoustic data and ensure temporal logic consistency through causal constraints. This module achieves dimensionality reduction representation of high-dimensional data through dynamic Tucker decomposition, while applying mask constraints to eliminate interference from future information in the temporal dimension, providing robust feature input for subsequent network training. The technical implementation is described in detail below: First, a Tucker decomposition model is established, and the preprocessed 3D data cube is... It is decomposed into a combination of a core tensor and a factor matrix. For example, the decomposition takes the following form: ; in For the core tensor,
[0036] These are factor matrices representing the spatial, temporal, and spectral dimensions, respectively. Preferably, the decomposition results are iteratively optimized using alternating least squares (ALS) method, with the goal of minimizing the reconstruction error. ; Furthermore, to adapt to the non-steady-state characteristics of photoacoustic signals, the weights of the factor matrix are dynamically adjusted during the iteration process. For example, the weights are updated based on the correlation between spectral channels. matrix: ; In the formula For learning rate, This represents the gradient operator.
[0037] To prevent the leakage of future information in the time dimension, causal constraints are imposed on the time factor matrix V. First, a lower triangular mask matrix is constructed. Its lower triangular elements (including the main diagonal) are all 1s, and the rest are 0s. The mask matrix is then applied to the factor matrix update process: ; In the formula This represents the Hadamard product (element-by-element multiplication). To strengthen the causal constraint, a regularization term is added to the objective function: ; in , where is the balancing coefficient, used to control the strength of causality constraints. The optimized objective function is: ; Dynamic Tucker decomposition extracts the essential features of photoacoustic data through dimensionality reduction, reducing the interference of redundant information on subsequent network training; the causal constraint masking mechanism forces the time dimension to rely only on historical information, avoiding logical errors introduced by future data during intraoperative dynamic updates; dynamic weight adjustment enables the decomposition process to adapt to the differences in tissue characteristics of different patients, enhancing the model's generalization ability.
[0038] Formula variable definition: : Preprocessed 3D data cube; Core tensor; Spatial, temporal, and spectral factor matrices; Lower triangular mask matrix; Causality constraint regularization coefficient; The learning rate is dynamically adjusted based on weights. Implementation process sequence structure: First, analyze the input data. Tucker decomposition initialization is performed to generate the initial core tensor and factor matrix; then, the decomposition results are iteratively optimized under an alternating least squares framework, during which the spectral factor matrix is dynamically adjusted. The weights are then applied; next, temporal causality constraints are imposed using a mask matrix. Corrected time factor matrix The final output is the core tensor. Add the factor matrix to the network training module.
[0039] The above implementation method can achieve the following effects: The Tucker decomposition is used to extract low-rank representations of the data, thereby reducing the input dimensionality for subsequent network training. The causal constraint masking mechanism eliminates interference from future information in the time dimension, ensuring the logical consistency of dynamic updates; Dynamic weight adjustment enhances the adaptability of the decomposition model to individual differences and improves the robustness of feature representation.
[0040] Core Tensor As a feature input to the segmentation network, it directly determines the accuracy of boundary recognition; the time factor matrix The causal constraints ensure temporal logic consistency during network training, preventing non-causal oscillations in prediction results. The output of the tensor decomposition module and the sparse matrix of the physical modeling module jointly constrain network training, forming a multimodal feature fusion.
[0041] The network training module is used to jointly optimize the physical constraints and segmentation network parameters through a meta-learning framework; In this invention, the network training module is used to jointly optimize physical constraints and segmentation network parameters. It achieves multi-task generalization capability through a meta-learning framework and ensures that the network output conforms to the wave equation constraints. This module adopts a dual-branch architecture, processing the compression matrix of the physical modeling module and the core features of the tensor decomposition module respectively, ultimately outputting a cancer boundary probability map. The technical implementation is explained in detail step by step below: First, a two-way structure is established, consisting of a physical constraint branch and a boundary partitioning branch. The input to the physical constraint branch is a compression matrix. and subspace basis Preferably, the predicted sound pressure field is mapped to the fully connected layer. : ; in The parameter is The neural network. The boundary segmentation branch takes the core tensor as input. The flattened feature is preferably achieved by generating boundary probability maps using a U-Net-type convolutional network. : ; in For parameters The segmentation network, This indicates the tensor vectorization operation.
[0042] To accommodate individual differences among patients, a model-independent meta-learning framework is employed. The outer loop updates the initial parameters across multiple tasks (patient datasets). : ; in , The balancing factor. The inner loop performs projected gradient descent (PGD) on a single task: ; in Physically constrained weights ensure the accuracy of the sound pressure field predicted by the network. Satisfying the compressed wave equation .
[0043] To embed the wave equation constraints into network training, it is preferable to adjust the parameters after each gradient update. Perform projection correction. For example, solve the constraints using the least squares method: ; And project the solution set to the parameter space to ensure Physical consistency is always maintained.
[0044] The dual-branch architecture separates physical modeling from data-driven features, avoiding overfitting of a single network to multimodal information; the meta-learning framework improves the model's generalization ability through multi-task pre-training, adapting to the tissue heterogeneity of different patients; projective gradient descent transforms the wave equation constraint into a hard constraint on the parameter space, ensuring that the network prediction conforms to the laws of sound field propagation.
[0045] Formula variable definition: : Predicted sound pressure field value; Boundary probability diagram; Network parameters for physical constraint branches and split branches; : Segmentation loss (binary classification cross-entropy); Physical constraint loss (mean squared error); Loss balance coefficient; : Internal and external learning rates.
[0046] Implementation process sequence structure: First, initialize the parameters of the dual-branch network and load the compression matrix. With core tensor Subsequently, the meta-parameters are updated based on multi-task data in the outer loop. Next, gradient descent is performed on a single task within the inner loop, with corrections made through projection operations. Finally, the trained network parameters are output to the real-time recognition module.
[0047] The above implementation method can achieve the following effects: The dual-branch architecture improves the interpretability of boundary recognition by separating physical constraints from data features; Meta-learning frameworks enhance the model's adaptability to new patients and reduce the overhead of repetitive training. Projected gradient descent ensures that the network output strictly satisfies the wave equation constraints, avoiding violations of physical laws.
[0048] Compression matrix of the physical modeling module subspace basis The input to the physical constraint branch and the core tensor of the tensor decomposition module to the segmentation branch are both fed into the network's output boundary probability graph. The data is transmitted to the real-time recognition module for intraoperative visualization and dynamic updates. During training, the loss calculation of the physical constraint branch depends on the true value of the sound pressure field. The true value is provided by the sound field reconstruction results from the physical modeling module.
[0049] The real-time recognition module is used to deploy the trained network to edge computing devices for intraoperative dynamic boundary correction. In this invention, the real-time recognition module is used to deploy the trained network model to an edge computing device and dynamically update the recognition results during surgery. This module achieves pipelined processing of tensor decomposition and network inference through hardware acceleration, and combines a real-time signal triggering mechanism to correct boundary predictions, ensuring the system meets clinical real-time requirements. The technical implementation is described in detail below step by step: First, a hardware mapping model is established to convert the tensor decomposition algorithm and network inference process into programmable logic circuits. For example, the matrix multiplication operation of tensor decomposition is implemented through a parallel multiplier array, preferably configured as a 16×16 array, with each multiplier having a 32-bit fixed-point width. The convolution operation in network inference is implemented through pre-compiled IP cores, preferably using the Winograd algorithm to optimize computational efficiency. The preferred resource allocation strategy is Dynamic Partial Reconfiguration (DPR), which adjusts the ratio of logic units to memory blocks in real time according to task requirements.
[0050] When a new photoacoustic signal is acquired during the procedure, an update condition is first triggered by detecting changes in signal energy. For example, the trigger threshold is defined as 10% of the average energy of historical signals. When the energy of the new signal exceeds this threshold, a dynamic correction process is initiated. The update process includes: re-performing tensor decomposition to generate the core tensor. The corrected boundary probability map is calculated using the network parameters embedded in the FPGA. For example, the update cycle is controlled within 100ms to ensure the continuity of clinical operations.
[0051] Visual interface and dynamic overlay Boundary probability diagram The data is transmitted to the display terminal via a DMA channel and fused with the endoscope video stream in real time. Preferably, an Alpha blending algorithm is used to achieve transparency overlay. ; in The original endoscopic image. A transparency factor (preferably 0.6-0.8) is used for superposition. The fused image is output to the surgical navigation display via an HDMI interface.
[0052] FPGA hardware acceleration addresses the real-time bottleneck of traditional CPUs / GPUs through parallel computing and resource optimization; dynamic triggering mechanisms avoid resource waste caused by invalid data updates, ensuring system response speed; and visualization overlay algorithms transform abstract probability maps into intuitive visual guidance, assisting doctors in accurately locating cancerous boundaries.
[0053] Formula variable definition: : The updated core tensor; : Corrected boundary probability map; : Fuse the output image; Transparency blending factor.
[0054] Implementation process sequence structure: First, the network parameters are loaded and the pipeline is initialized via FPGA; after the intraoperative real-time signal is triggered, tensor decomposition is performed to update the core tensor. Subsequently, the network inference IP core was called to generate... Finally, through Alpha mixing, It is superimposed onto the endoscope video stream and output to the display terminal.
[0055] The above implementation method can achieve the following effects: Accelerated by FPGA pipeline, the time constraints for real-time intraoperative processing are met; The dynamic triggering mechanism reduces redundant calculations while ensuring update accuracy; Visual overlays provide intuitive guidance on cancerous boundaries, assisting in surgical decisions.
[0056] New data from the signal acquisition module triggers the update process of the real-time recognition module; the tensor decomposition module generates... As network input, with the physical modeling module Commonly constrained reasoning process; network output Feedback is sent to the display terminal, forming a closed-loop interaction. The hardware resource allocation strategy of the real-time recognition module directly affects the signal processing throughput and must be matched with the computational complexity of the physical modeling module.
[0057] To ensure real-time performance, the delays at each stage must meet the following requirements: ; in For tensor decomposition time, For network inference time, For rendering overlay time, preferably, delay control is achieved by adjusting the clock frequency and balancing it with pipeline stages.
[0058] When FPGA resource usage exceeds a preset threshold (preferably 90%), a degradation mode is triggered: non-core logic units (such as visualization rendering) are suspended, prioritizing resource allocation for tensor decomposition and network inference. After the fault is recovered, it automatically switches back to full-function mode. The output of the physical modeling module is connected to the tensor decomposition module, and the output of the tensor decomposition module is connected to the network training module.
[0059] Please see the appendix Figure 2 This invention provides a method for identifying the boundary of oral cancerous regions based on photoacoustic spectral imaging, including: S1. Acquire multispectral photoacoustic signals and perform adaptive wavelet denoising; S2. Based on the discretization of the wave equation, a sparse matrix is generated and compressed by subspace projection. S3. Perform dynamic tensor decomposition on the denoised data and apply time causality constraints; S4. Jointly train the physical constraint and segmentation network using a meta-learning framework; S5. Deploy the network to edge devices for real-time intraoperative boundary identification and dynamic updates.
[0060] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A system for identifying the boundary of oral cancerous regions based on photoacoustic spectral imaging, characterized in that, include: The signal acquisition module is used to generate spatiotemporal-spectral three-dimensional data by exciting and receiving photoacoustic signals through multi-wavelength lasers. The physical modeling module is used to discretize the wave equation to generate a sparse matrix and compress the matrix dimension through subspace projection. The tensor decomposition module is used to perform dynamic tensor decomposition on preprocessed data and apply time causality constraints. The network training module is used to jointly optimize the physical constraints and segmentation network parameters through a meta-learning framework; The real-time recognition module is used to deploy the trained network to edge computing devices for intraoperative dynamic boundary correction. The output of the physical modeling module is connected to the tensor decomposition module, and the output of the tensor decomposition module is connected to the network training module.
2. The oral cancer region boundary recognition system based on photoacoustic spectral imaging according to claim 1, characterized in that, The signal acquisition module includes an adaptive wavelet threshold denoising unit, which separates noise components through discrete wavelet transform and dynamically adjusts the threshold according to the noise variance.
3. The oral cancer region boundary recognition system based on photoacoustic spectral imaging according to claim 1, characterized in that, The physical modeling module compresses sparse matrices through the following steps: Constructing the basis matrix of the Krylov subspace ,in The sparse matrix generated by discretizing the wave equation. The vector of heat source terms; Projecting the sparse matrix onto a low-dimensional subspace generates a compressed matrix. , This is the matrix transpose operator.
4. The oral cancer region boundary recognition system based on photoacoustic spectral imaging according to claim 1, characterized in that, The tensor decomposition module uses a lower triangular mask matrix. Applying a time causality constraint, satisfying: The time dimension factor matrix is updated to ,in It represents the Hadamardi (or Hadama) stack; Mask matrix The lower triangular element is 1, and the rest are 0.
5. The oral cancer region boundary recognition system based on photoacoustic spectral imaging according to claim 1, characterized in that, The tensor decomposition module dynamically adjusts the weights of the factor matrix based on spectral-spatial correlation.
6. The oral cancer region boundary recognition system based on photoacoustic spectral imaging according to claim 1, characterized in that, The network training module employs an alternating direction optimization strategy, performing meta-learning tasks to update parameters in the outer loop and combining projective gradient descent in the inner loop to ensure that the parameters satisfy the wave equation constraints.
7. The oral cancer region boundary recognition system based on photoacoustic spectral imaging according to claim 1, characterized in that, The network training module includes a dual-branch structure: The first branch takes the compressed sparse matrix as input and outputs the sound pressure field prediction. The second branch takes the core features obtained from tensor decomposition as input and outputs a boundary probability map.
8. The oral cancer region boundary recognition system based on photoacoustic spectral imaging according to claim 1, characterized in that, The real-time recognition module uses an FPGA chip to implement pipelined parallel computation of tensor decomposition and network inference.
9. The oral cancer region boundary recognition system based on photoacoustic spectral imaging according to claim 1, characterized in that, The real-time recognition module dynamically updates the tensor decomposition results based on the newly acquired signals during the operation and re-solves the network output to correct the boundaries.
10. A method for identifying the boundary of oral cancerous regions based on photoacoustic spectral imaging, characterized in that, Includes the following steps: S1. Acquire multispectral photoacoustic signals and perform adaptive wavelet denoising; S2. Based on the discretization of the wave equation, a sparse matrix is generated and compressed by subspace projection. S3. Perform dynamic tensor decomposition on the denoised data and apply time causality constraints; S4. Jointly train the physical constraint and segmentation network using a meta-learning framework; S5. Deploy the network to edge devices for real-time intraoperative boundary identification and dynamic updates.