Fluorescence boundary fused tumor image boundary identification method and system

Abstract

The invention relates to the technical field of medical image processing and tumor treatment, and provides a tumor image boundary identification method and system fused with a fluorescence boundary, and the method comprises the steps: collecting a tumor dynamic fluorescence image sequence and breathing waveform data through a fluorescence imaging device; performing multi-scale noise suppression processing on the image sequence to generate a preprocessed image without artifacts; extracting fluorescence intensity distribution characteristics through a multi-layer characteristic weighted fusion network and generating an intensity distribution matrix; constructing a motion compensation model according to the breathing waveform data, and correcting the boundary displacement of the intensity distribution matrix; fusing the corrected fluorescence features and the tumor tissue texture features to generate an initial tumor boundary contour; iterative topological optimization is carried out on the initial boundary, and boundary discontinuity is eliminated; and mapping the optimized boundary to a space coordinate system of the laser treatment equipment, generating a boundary mapping model and updating the boundary mapping model in real time. The precision and safety of laser treatment can be improved, and damage to normal tissue is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing and tumor treatment, and more specifically, to a tumor image boundary recognition method and system integrating fluorescence boundaries. Background Art

[0002] In the field of medical imaging technology and tumor treatment, fluorescence imaging technology is widely used in the identification and positioning of tumor boundaries due to its high sensitivity and specificity. Fluorescence imaging can reflect the distribution of fluorescence signals in the tumor area in real time by marking fluorescent probes of tumor tissue, providing a visual basis for surgery or laser treatment. However, fluorescence imaging faces many challenges in practical applications. First, the fluorescence signal in the tumor area is easily affected by respiratory movement, which causes displacement and blurring of boundary information in the image, affecting the accuracy of diagnosis and treatment. Secondly, the fluorescence imaging process is often accompanied by noise and artifacts, such as high-frequency noise, fluorescence leakage, etc., which further interfere with the clear identification of tumor boundaries. In addition, the existing technology lacks an effective multi-dimensional feature weight allocation mechanism when fusing fluorescence signals with the structural characteristics of tumor tissue, making it difficult to achieve accurate boundary extraction.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: it is impossible to effectively compensate for the influence of respiratory movement on the boundaries of fluorescence images, and it is difficult to maintain the stability of boundary recognition under dynamic conditions; the ability to suppress noise and eliminate artifacts in fluorescence images is insufficient, resulting in limited boundary extraction accuracy; there is a lack of mechanisms for dynamic optimization of tumor boundaries and design of safety buffer zones, and the high precision and safety requirements of laser treatment cannot be met. Summary of the invention

[0004] A first aspect of the present invention provides a tumor image boundary recognition method integrating fluorescence boundaries, comprising:

[0005] The dynamic fluorescence image sequence of the tumor area is continuously collected by the fluorescence imaging device, and the real-time respiratory waveform data output by the respiratory motion monitoring sensor is simultaneously obtained;

[0006] Performing multi-scale noise suppression processing on the dynamic fluorescence image sequence, adaptively adjusting the filtering parameters based on the image resolution, and generating a pre-processed fluorescence image with artifacts removed;

[0007] The spatial fluorescence intensity distribution features are extracted from the preprocessed fluorescence image through a multi-layer feature weighted fusion network to generate a fluorescence intensity distribution matrix that integrates multi-scale features.

[0008] A motion compensation model is constructed based on real-time respiratory waveform data, and the intensity distribution matrix is ​​corrected for boundary displacement through phase angle tracking and motion vector prediction to obtain the corrected fluorescence intensity distribution characteristics.

[0009] The corrected fluorescence intensity distribution features are fused with the texture structure features of the tumor tissue to generate an initial tumor boundary contour containing multi-dimensional feature weights;

[0010] Iterative topological optimization is performed on the initial tumor boundary contour, and boundary discontinuities are eliminated through dynamic threshold segmentation and morphological operations to obtain the optimized tumor boundary.

[0011] The optimized tumor boundary is mapped to the spatial coordinate system of the laser treatment device, and a boundary mapping model including a safety buffer zone is generated and updated in real time.

[0012] Furthermore, the spatial fluorescence intensity distribution features are extracted from the preprocessed fluorescence image through a multi-layer feature weighted fusion network, including:

[0013] Three layers of convolution kernels with different scales are used to extract local features, regional features and global features of the fluorescence image respectively;

[0014] The features of each layer are corrected by the feature attenuation compensation formula, which is expressed as:

[0015] ;

[0016] in, represents pixel coordinates, represents the fluorescence intensity feature extracted by the i-th convolution kernel, represents the fluorescence diffusion attenuation coefficient, represents the fluorescence attenuation compensation factor, represents the fluorescence intensity feature extracted by the i-th convolution kernel after correction;

[0017] Generating a fluorescence intensity distribution matrix that fuses multi-scale features includes: performing weighted fusion on the corrected features to generate a fluorescence intensity distribution matrix, which is expressed as:

[0018] ;

[0019] in, is the fluorescence intensity distribution matrix, is the weight coefficient of the i-th feature layer.

[0020] Furthermore, the modified fluorescence intensity distribution characteristics are obtained by:

[0021] Perform fast Fourier transform on the respiratory waveform data, extract the dominant frequency component and calculate the respiratory phase angle;

[0022] Based on the respiratory phase angle, a respiratory motion compensation coefficient model is constructed to obtain a respiratory motion compensation coefficient. The respiratory motion compensation coefficient model is expressed as:

[0023] ;

[0024] in, represents the respiratory phase angle, represents the spatial displacement weight, represents the speed impact factor, represents the respiratory phase change rate, C represents the respiratory motion compensation coefficient;

[0025] The fluorescence intensity distribution matrix is ​​dynamically interpolated and corrected according to the respiratory motion compensation coefficient to obtain the corrected fluorescence intensity distribution characteristics, where the corrected matrix satisfies the following conditions:

[0026] ;

[0027] in, , is the tissue displacement reference value calibrated before surgery, and F is the fluorescence intensity distribution matrix function.

[0028] Furthermore, the corrected fluorescence intensity distribution feature is fused with the texture structure feature of the tumor tissue to generate an initial tumor boundary contour containing multi-dimensional feature weights, including:

[0029] The fusion weights of fluorescence features and tissue features are calculated as follows:

[0030] ;

[0031] in, represents the standard deviation of the fluorescence intensity distribution matrix F, Represents the standard deviation of the texture feature matrix T;

[0032] Based on the fusion weight and the corrected fluorescence intensity distribution characteristics, a fusion boundary feature map is generated, which is expressed as follows:

[0033] ;

[0034] in, represents the fused boundary feature map, represents the fusion weight, Represents the texture feature matrix; represents the corrected fluorescence intensity distribution matrix;

[0035] The fused boundary feature map is binarized to generate the initial tumor boundary contour.

[0036] Furthermore, the binarization processing of the fused boundary feature map to generate an initial tumor boundary contour includes:

[0037] The dynamic segmentation threshold is calculated based on the fused boundary feature map, expressed as:

[0038] ;

[0039] Among them, τ represents the dynamic segmentation threshold, N represents the total number of pixels, represents the mean of the feature map, is the sensitivity adjustment coefficient, Represents the fused boundary feature map;

[0040] Compare the fused boundary feature map with the dynamic segmentation threshold. If the pixel of the fused boundary feature map is greater than the segmentation threshold, the comparison result is 1; if the pixel of the fused boundary feature map is less than the segmentation threshold, the comparison result is 0.

[0041] generating an initial tumor boundary contour based on the comparison results;

[0042] Multi-scale morphological optimization is performed, and the size of the structure element decreases proportionally to the number of iterations to eliminate pseudo-boundary residues.

[0043] Furthermore, the fluorescence attenuation compensation factor is determined by the following method:

[0044] Use fluorescence calibration films at different depths to conduct benchmark attenuation experiments to obtain the initial value of the fluorescence attenuation compensation factor;

[0045] The fluorescence attenuation compensation factor is dynamically corrected according to the real-time detected tissue optical parameters, as shown below:

[0046] ;

[0047] in, represents the nominal absorbance, is the tissue variability coefficient, represents the fluorescence attenuation compensation factor, represents the tissue optical parameters, Indicates the initial value.

[0048] Furthermore, the tissue displacement reference amount is determined as follows:

[0049] Based on the preoperative four-dimensional CT images, a respiratory phase-tissue displacement mapping relationship table was constructed to record the tissue displacement corresponding to different respiratory phases;

[0050] Querying a reference displacement from a mapping relationship table based on the real-time respiratory phase;

[0051] The reference displacement is adjusted according to the real-time breathing amplitude A, and the adjustment formula is:

[0052] ;

[0053] ;

[0054] in, , is the reference displacement, is the breathing amplitude threshold.

[0055] Furthermore, the performing multi-scale noise suppression processing on the dynamic fluorescence image sequence includes:

[0056] Non-local mean filtering is used to eliminate high-frequency noise in dynamic fluorescence image sequences, and the filter window size is positively correlated with the image resolution.

[0057] Fluorescence leakage artifacts are suppressed by anisotropic diffusion algorithm, and the diffusion coefficient is negatively correlated with the local gradient amplitude;

[0058] The periodic noise is soft-thresholded in the wavelet transform domain to retain the low-frequency fluorescence features.

[0059] Furthermore, mapping the optimized tumor boundary to the spatial coordinate system of the laser treatment device comprises the following steps:

[0060] The optimized two-dimensional boundary segmentation result of the tumor boundary is projected into a three-dimensional space coordinate system to establish a voxel-level boundary mapping model;

[0061] A dynamic safety buffer is generated based on the kinematic parameters of the laser device. The buffer width is expressed as follows:

[0062] ;

[0063] Where, v represents the moving speed of the laser head, Indicates system response delay, is the safety margin factor, Indicates the buffer width;

[0064] A buffer layer with a thickness of W is generated on the outer surface of the boundary mapping model.

[0065] In a second aspect, the present invention provides a tumor image boundary recognition system integrating fluorescence boundaries, including a data acquisition module, a preprocessing module, a matrix generation module, a feature correction module, a contour generation module, a contour optimization module and a data mapping module;

[0066] A data acquisition module continuously acquires a dynamic fluorescence image sequence of the tumor area through a fluorescence imaging device, and synchronously acquires real-time respiratory waveform data output by a respiratory motion monitoring sensor;

[0067] A preprocessing module performs multi-scale noise suppression processing on the dynamic fluorescence image sequence, adaptively adjusts the filtering parameters based on the image resolution, and generates a preprocessed fluorescence image with artifacts removed;

[0068] The matrix generation module extracts the spatial fluorescence intensity distribution features from the preprocessed fluorescence image through a multi-layer feature weighted fusion network to generate an intensity distribution matrix containing the fluorescence diffusion attenuation characteristics;

[0069] The feature correction module builds a motion compensation model based on real-time respiratory waveform data, performs boundary displacement correction on the intensity distribution matrix through phase angle tracking and motion vector prediction, and obtains the corrected fluorescence intensity distribution feature;

[0070] The contour generation module fuses the corrected fluorescence intensity distribution features with the texture structure features of the tumor tissue to generate an initial tumor boundary contour containing multi-dimensional feature weights;

[0071] The contour optimization module performs iterative topological optimization on the initial tumor boundary contour, eliminates boundary discontinuities through dynamic threshold segmentation and morphological operations, and obtains the optimized tumor boundary;

[0072] The data mapping module maps the optimized tumor boundary to the spatial coordinate system of the laser treatment device, generates a boundary mapping model including a safety buffer zone, and updates it in real time.

[0073] The above-mentioned embodiments of the present invention have at least the following beneficial effects: the method of the present invention can effectively solve many problems of tumor boundary identification in the prior art. Through the dynamic feature extraction network and the respiratory motion compensation model, combined with the phase angle tracking and motion vector prediction technology, the boundary displacement of the fluorescence intensity distribution matrix caused by respiratory motion can be accurately corrected, thereby maintaining the stability and accuracy of tumor boundary identification under dynamic conditions. At the same time, by integrating the fluorescence intensity distribution characteristics with the texture structure characteristics of the tumor tissue, and through dynamic threshold segmentation and multi-scale morphological optimization, a high-precision tumor boundary contour can be generated, eliminating boundary discontinuities and pseudo-boundary residues, and providing a reliable basis for subsequent laser treatment.

[0074] In addition, the present invention can also map the optimized tumor boundary to the spatial coordinate system of the laser treatment device in real time, generate a boundary mapping model including a safety buffer zone, and dynamically adjust the buffer zone width according to the kinematic parameters of the laser device. At the same time, by establishing a multi-parameter safety interlocking mechanism, the treatment pause can be triggered in time when the boundary is blurred, the respiratory motion changes suddenly, or the device positioning deviation exceeds the safety range, thereby improving the safety and reliability of laser treatment and reducing the risk of accidental injury to normal tissues. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, in which:

[0076] Figure 1 A schematic flow chart of a tumor image boundary recognition method integrating fluorescence boundaries provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0077] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0078] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, device, apparatus, method or computer program product. Therefore, the present invention can be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0079] It should be noted that any number of elements in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.

[0080] Reference below Figure 1 , Figure 1 FIG. 1 is a flow chart of a tumor image boundary recognition method fused with fluorescence boundary provided by an embodiment of the present invention. Figure 1 As shown, a tumor image boundary recognition method 100 integrating fluorescence boundaries includes:

[0081] S1: Continuously collect dynamic fluorescence image sequences of the tumor area through the fluorescence imaging device, and synchronously obtain the real-time respiratory waveform data output by the respiratory motion monitoring sensor;

[0082] S2: performing multi-scale noise suppression processing on the dynamic fluorescence image sequence, adaptively adjusting filtering parameters based on image resolution, and generating a pre-processed fluorescence image with artifacts removed;

[0083] S3: Extract spatial fluorescence intensity distribution features from the preprocessed fluorescence image through a multi-layer feature weighted fusion network to generate an intensity distribution matrix containing fluorescence diffusion attenuation characteristics;

[0084] S4: A motion compensation model is constructed based on real-time respiratory waveform data, and boundary displacement correction is performed on the intensity distribution matrix through phase angle tracking and motion vector prediction;

[0085] S5: Fusing the corrected fluorescence intensity distribution features with the texture structure features of the tumor tissue to generate an initial tumor boundary contour containing multi-dimensional feature weights;

[0086] S6: Perform iterative topology optimization on the initial tumor boundary contour and eliminate boundary discontinuities through dynamic threshold segmentation and morphological operations;

[0087] S7: Map the optimized tumor boundary to the spatial coordinate system of the laser treatment device, generate a boundary mapping model including a safety buffer zone, and update it in real time.

[0088] It should be noted that when implementing the tumor image boundary recognition method, it is first necessary to continuously collect the dynamic fluorescence image sequence of the tumor area through the fluorescence imaging device, and synchronously obtain the real-time respiratory waveform data output by the respiratory motion monitoring sensor. The fluorescence imaging device is a device that can capture fluorescence signals. It uses a light source of a specific wavelength to excite the fluorescent probe to obtain a fluorescent image of the tumor area. The dynamic fluorescence image sequence refers to multiple frames of images collected continuously over a period of time. These images can reflect the changes in the fluorescence signal of the tumor area over time. The respiratory motion monitoring sensor is used to monitor the patient's respiratory state in real time. The respiratory waveform data it outputs can reflect information such as the respiratory cycle and amplitude, which is crucial for subsequent motion compensation processing.

[0089] Specifically, the fluorescence imaging device can use a highly sensitive camera and a light source of a specific wavelength, such as a near-infrared light source, to improve the efficiency of capturing fluorescence signals. The acquisition frequency of the dynamic fluorescence image sequence can be adjusted according to the respiratory cycle, usually set to 5 to 10 frames per second to ensure that the changes in the fluorescence signal caused by respiratory movement can be captured. The respiratory movement monitoring sensor can be a device based on pressure, capacitance or optical principles, and the respiratory waveform data it outputs can be parameters such as respiratory cycle and respiratory amplitude. These parameters can be converted into digital signals by the signal processing unit of the sensor for subsequent analysis and processing. During the image acquisition process, it is necessary to ensure the synchronization of the fluorescence imaging device and the respiratory movement monitoring sensor to ensure the temporal consistency of the dynamic fluorescence image sequence and the respiratory waveform data.

[0090] Preferably, in order to improve the accuracy and stability of image acquisition, automatic exposure and automatic focusing functions can be introduced into the fluorescence imaging device to adapt to the imaging requirements of different patients and tumor locations. At the same time, the respiratory motion monitoring sensor can select a high-precision fiber optic sensor, and the respiratory waveform data output by it can be filtered to remove noise to improve signal quality. In addition, in order to further enhance the dynamic information of the image sequence, time resolution technology can be introduced in the acquisition process, such as controlling the acquisition time window of the fluorescence signal through time gating, so as to more clearly capture the changes of the fluorescence signal during the respiratory cycle.

[0091] In some embodiments, step S3 includes:

[0092] S31: Multi-layer convolution kernels are used to extract fluorescence intensity features of different scales. The output feature map of each layer is weighted fused by the following formula:

[0093] ;

[0094] in, represents pixel coordinates, represents the fluorescence intensity feature extracted by the i-th convolution kernel, represents the fluorescence diffusion attenuation coefficient, is the fluorescence attenuation compensation factor, is the weight coefficient of each feature layer and satisfies .

[0095] It should be noted that when constructing a dynamic feature extraction network, multi-layer convolution kernels are required to extract fluorescence intensity features of different scales. The feature map output by each layer is weighted fused to generate the final fluorescence intensity distribution matrix. In this process, the fluorescence intensity feature refers to the feature information related to the fluorescence signal extracted from the fluorescence image, which is used to reflect the distribution of the fluorescence signal in the tumor area. The fluorescence diffusion attenuation coefficient is a parameter used to describe the intensity attenuation of the fluorescence signal when it propagates in the tissue, and the fluorescence attenuation compensation factor is to correct this attenuation difference to make the extracted features more accurate. The weight coefficients of each feature layer are used to balance the importance of features of different scales to ensure that the final fused feature map can fully reflect the fluorescence characteristics of the tumor area.

[0096] Specifically, the design of multi-layer convolution kernels may include convolution kernels of different sizes, such as 3×3, 5×5, etc., to extract fluorescence intensity features from local to global. The fluorescence intensity features extracted by each layer of convolution kernels are nonlinearly processed by the activation function and then weighted fused. In the weighted fusion process, the fluorescence diffusion attenuation coefficient can be obtained by experimental calibration, such as measuring the attenuation of fluorescence signals in tissues at different depths. The fluorescence attenuation compensation factor can be dynamically adjusted according to the optical properties of the tissue to compensate for the differences in fluorescence signal attenuation caused by different depths or different tissue types. The weight coefficients of each feature layer can be assigned according to the importance of the feature. These weights are usually automatically learned by training a neural network to ensure that their sum is 1, thereby ensuring the rationality of feature fusion.

[0097] Preferably, deep learning techniques can be introduced in the feature extraction process, such as using a convolutional neural network (CNN) to automatically learn the extraction and fusion of fluorescence intensity features. By training a network model containing a large amount of annotated data, the parameters and weight coefficients of the convolution kernel can be optimized, thereby improving the accuracy and robustness of feature extraction. In addition, in order to further enhance the expressiveness of features, an attention mechanism can be added to the feature extraction network so that the network can automatically focus on the key features of the tumor area. In the weighted fusion process, an adaptive weight adjustment mechanism can also be introduced to dynamically adjust the weight coefficient according to the statistical characteristics of the feature map to better adapt to the distribution of fluorescence signals in different images.

[0098] In some embodiments, step S4 includes:

[0099] S41: Analyze respiratory waveform data to calculate real-time respiratory phase angle and respiratory amplitude A;

[0100] S42: Construct a respiratory motion compensation coefficient C, which is expressed as:

[0101] ;

[0102] in, represents the spatial displacement weight, represents the speed impact factor, Represents the respiratory phase change rate;

[0103] S43: Dynamically interpolate and correct the fluorescence intensity distribution matrix according to the compensation coefficient C, and the corrected matrix satisfies:

[0104] ;

[0105] in, , It is the tissue displacement reference value calibrated before surgery.

[0106] It should be noted that when constructing a motion compensation model, it is necessary to parse the respiratory waveform data to calculate the real-time respiratory phase angle and respiratory amplitude, and then construct the respiratory motion compensation coefficient, and use this coefficient to dynamically interpolate and correct the fluorescence intensity distribution matrix. The respiratory waveform data is a signal obtained from the respiratory motion monitoring sensor, which reflects the periodicity and amplitude changes during the patient's breathing process. The real-time respiratory phase angle refers to the phase position in the current respiratory cycle, while the respiratory amplitude indicates the range of motion of the chest and abdomen during breathing. The respiratory motion compensation coefficient is a parameter used to correct the displacement of the fluorescence image caused by respiratory motion. The coefficient can be used to correct the fluorescence intensity distribution matrix, thereby eliminating the interference of respiratory motion on boundary recognition.

[0107] Specifically, when analyzing respiratory waveform data, the period and amplitude information of the respiratory signal can be extracted by Fourier transform or wavelet transform, and then the real-time respiratory phase angle can be calculated. The respiratory amplitude can be determined by measuring the peak-to-valley difference of the respiratory waveform. The construction of the respiratory motion compensation coefficient involves parameters such as spatial displacement weight, speed influence factor, and respiratory phase change rate. The spatial displacement weight is used to adjust the degree of influence of the compensation coefficient on the displacement, the speed influence factor takes into account the influence of the respiratory motion speed on the image displacement, and the respiratory phase change rate reflects the dynamic characteristics of the respiratory motion. In the dynamic interpolation correction process, the preoperatively calibrated tissue displacement reference is used to determine the specific degree of influence of the respiratory motion on the fluorescence intensity distribution matrix, thereby achieving accurate correction of the boundary displacement.

[0108] Preferably, when parsing the respiratory waveform data, a high-precision signal processing algorithm, such as a Kalman filter, can be used to improve the calculation accuracy of the respiratory phase angle and the respiratory amplitude. For the construction of the respiratory motion compensation coefficient, a machine learning method, such as a support vector machine or a neural network, can be used to dynamically adjust the spatial displacement weight and the velocity influencing factor according to the individual patient's respiratory characteristics, thereby improving the adaptability and accuracy of the compensation coefficient. In addition, the determination of the tissue displacement reference amount can be more accurately calibrated by multimodal image fusion technology, combined with preoperative CT or MRI data, to calibrate the effect of respiratory motion on tissue displacement, and provide a more reliable reference for dynamic interpolation correction.

[0109] In some embodiments, step S5 includes:

[0110] S51: Extract the vascular density distribution map and cellular heterogeneity map of the current case from the tumor tissue database;

[0111] S52: Calculate the fusion weight of fluorescence features and tissue features :

[0112] ;

[0113] in, represents the standard deviation of the fluorescence intensity distribution matrix F, Represents the standard deviation of the texture feature matrix T;

[0114] S53: Generate fusion boundary feature map , which satisfies:

[0115] ;

[0116] It should be noted that when fusing the corrected fluorescence intensity distribution features with the texture structure features of the tumor tissue, it is necessary to generate an initial tumor boundary contour containing multidimensional feature weights. The fluorescence intensity distribution feature refers to the distribution of the fluorescence signal after motion compensation correction, which reflects the fluorescence characteristics of the tumor area; while the texture structure features of the tumor tissue describe the spatial distribution, shape and texture information of the tumor tissue on the image. The fusion weight is used to balance the importance of these two features to generate a more accurate initial tumor boundary contour. In this way, the fluorescence signal and tissue structure information can be comprehensively considered to improve the accuracy and robustness of boundary recognition.

[0117] Specifically, the standard deviation of the fluorescence intensity distribution matrix ( ) is used to measure the intensity variation range of the fluorescence signal, and the standard deviation of the texture feature matrix ( ) reflects the complexity of texture information. The calculation of fusion weight is based on the ratio of these two standard deviations, which is used to dynamically adjust the contribution ratio of fluorescence features and texture features in boundary generation. When the fluorescence signal changes greatly, the weight of the fluorescence feature will increase; when the texture information is richer, the weight of the texture feature will dominate. When generating a fused boundary feature map, by weighted fusion of the corrected fluorescence intensity distribution features and texture features, a boundary contour that comprehensively considers multi-dimensional features can be obtained, providing a basis for subsequent boundary optimization.

[0118] Preferably, in order to further optimize the fusion process, an adaptive mechanism can be introduced when calculating the fusion weight, and the parameters in the weight calculation formula can be dynamically adjusted according to different tumor types or image quality. For example, for tumors with weak fluorescence signals, the weight of texture features can be appropriately increased to improve the reliability of boundary recognition. In addition, texture features can be extracted using a variety of methods, such as gray-level co-occurrence matrix (GLCM) or local binary pattern (LBP) to extract richer texture information. In the process of generating the fused boundary feature map, a regularization term can also be introduced to reduce the impact of noise on the boundary contour. For example, the fused boundary feature map can be preprocessed using methods such as Gaussian smoothing or wavelet denoising, thereby further improving the accuracy of boundary recognition.

[0119] In some embodiments, step S6 includes:

[0120] S61: Calculate the dynamic segmentation threshold τ based on the fused boundary feature map, and its expression is:

[0121] ;

[0122] Where N represents the total number of pixels, represents the mean of the feature map, is the sensitivity adjustment coefficient;

[0123] S62: Perform multi-scale morphological optimization, and the size of the structural element decreases proportionally with the number of iterations to eliminate pseudo-boundary residues.

[0124] It should be noted that when iterative topological optimization is performed on the initial tumor boundary contour, boundary discontinuities need to be eliminated through dynamic threshold segmentation and morphological operations. Dynamic threshold segmentation is a method that dynamically adjusts the segmentation threshold according to image features to separate the tumor boundary from the background. Morphological operations optimize the boundary through mathematical morphology to remove pseudo-boundaries and fill boundary gaps, thereby generating a smoother and more continuous tumor boundary. This process aims to improve the accuracy and robustness of boundary recognition and provide a more reliable basis for subsequent treatment.

[0125] Specifically, the threshold calculation of dynamic threshold segmentation is based on the statistical characteristics of the fused boundary feature map, such as the mean and standard deviation of the feature map. By introducing a sensitivity adjustment coefficient, the threshold can be dynamically adjusted according to the complexity of the image to meet the segmentation requirements of different tumor boundary features. Morphological operations include corrosion, dilation, opening and closing operations, etc. These operations process the boundaries through structural elements. The size of the structural element can be adjusted according to the complexity of the boundary features, and is usually optimized in a geometric decreasing manner during the iteration process to gradually refine the boundary. In each iteration, these operations are used to remove noise points and pseudo-boundaries on the boundary, and small gaps on the boundary are filled, thereby eliminating boundary discontinuities.

[0126] Preferably, in the dynamic threshold segmentation process, a local adaptive threshold segmentation algorithm can be introduced, such as a threshold segmentation algorithm based on the Otsu method or a local area, to better adapt to local changes in boundary features. For morphological operations, multi-scale structural elements can be used, combined with structural elements in different directions for optimization, to more comprehensively handle discontinuities on the boundaries. In addition, deep learning methods, such as convolutional neural networks (CNNs), can be introduced to further refine and correct the optimized boundaries to improve the accuracy and robustness of boundary recognition. For example, CNNs can be used to classify the optimized boundaries and distinguish between real boundaries and pseudo boundaries, thereby further improving the accuracy of boundary recognition.

[0127] In some embodiments, the fluorescence attenuation compensation factor Determined by:

[0128] S61: Use fluorescent calibration films at different depths to perform baseline attenuation experiments to obtain Initial value of ;

[0129] S62: Based on real-time detection of tissue optical parameters right Perform dynamic correction, the correction formula is:

[0130] ;

[0131] in, represents the nominal absorbance, is the tissue variability coefficient.

[0132] It should be noted that when determining the fluorescence attenuation compensation factor, it is first necessary to conduct a baseline attenuation experiment using fluorescence calibration films at different depths to obtain the initial value, and then dynamically correct the initial value based on the real-time detected tissue optical parameters. The fluorescence attenuation compensation factor is used to correct the attenuation difference caused by depth and tissue characteristics when the fluorescence signal propagates in the tissue, thereby improving the accuracy of the fluorescence intensity distribution characteristics. The fluorescence calibration film is a standard sample with known fluorescence characteristics, which is used to simulate the attenuation of fluorescence signals at different depths, while the tissue optical parameters reflect the absorption and scattering characteristics of the tissue to light. The real-time detection of these parameters can provide a basis for the dynamic correction of the fluorescence attenuation compensation factor.

[0133] Specifically, the benchmark attenuation experiment places the fluorescent calibration piece at different depths and measures the change in the intensity of the fluorescent signal to obtain the initial attenuation compensation factor. This initial value reflects the attenuation law of the fluorescent signal under standard conditions. The tissue optical parameters detected in real time include the absorbance and scattering coefficient of the tissue, which can be obtained through optical sensors or imaging devices. The absorbance reflects the degree of light absorption by the tissue, while the scattering coefficient describes the scattering of light in the tissue. Dynamic correction of the initial attenuation compensation factor based on these parameters can more accurately compensate for the attenuation differences of the fluorescent signal in the actual tissue, thereby improving the accuracy of the fluorescence intensity distribution characteristics.

[0134] Preferably, when conducting a baseline attenuation experiment, fluorescent calibration films of various depths can be used to cover the different depth ranges where the tumor may exist. For example, calibration films of different depths such as 0.5 cm, 1.0 cm and 1.5 cm can be selected for experiments to obtain more comprehensive initial attenuation compensation factor data. For the detection of tissue optical parameters, multispectral imaging technology or photoacoustic imaging technology can be used, which can provide more accurate tissue optical property information. In the dynamic correction process, machine learning algorithms such as linear regression or neural networks can be introduced to automatically adjust the correction formula of the attenuation compensation factor according to the real-time detected tissue optical parameters, thereby further improving the accuracy and adaptability of the compensation.

[0135] In some embodiments, the tissue displacement reference amount , The determination includes:

[0136] S71: Construct a respiratory phase-tissue displacement mapping relationship table based on preoperative four-dimensional CT images;

[0137] S72: Adjust the reference amount according to the real-time breathing amplitude A, and the adjustment formula is:

[0138] ;

[0139] ;

[0140] in, , is the static calibration reference quantity, is the breathing amplitude threshold.

[0141] It should be noted that when determining the tissue displacement baseline, it is necessary to construct a respiratory phase-tissue displacement mapping relationship table based on the preoperative four-dimensional CT image, and adjust the baseline according to the real-time respiratory amplitude. Preoperative four-dimensional CT imaging is an imaging technology that can reflect tissue movement during the respiratory cycle. It records the position changes of tissues under different respiratory phases. The respiratory phase-tissue displacement mapping relationship table is established by analyzing four-dimensional CT image data to establish a quantitative relationship between respiratory phase and tissue displacement, which is used to predict the impact of respiratory movement on tissue position. Real-time respiratory amplitude refers to the actual amplitude of the patient's respiratory movement during surgery, which is obtained through monitoring equipment and used to adjust the tissue displacement baseline to adapt to the patient's current respiratory state.

[0142] Specifically, when constructing a respiratory phase-tissue displacement mapping relationship table, it is necessary to extract the displacement data of the tissue under each respiratory phase from the preoperative four-dimensional CT image. These data can be obtained through image registration technology, that is, aligning the CT images at different time points and calculating the displacement vector of the tissue during the respiratory cycle. The static calibration reference is a tissue displacement reference value pre-calculated based on the preoperative image data, usually including the amplitude and direction of the displacement. The respiratory amplitude threshold is a preset reference value used to determine whether the real-time respiratory amplitude is within the normal range. When adjusting the reference according to the real-time respiratory amplitude, a nonlinear function can be used to describe the relationship between the respiratory amplitude and tissue displacement, such as a logarithmic function or an exponential function, to more accurately reflect the impact of respiratory movement on tissue displacement.

[0143] Preferably, when constructing the respiratory phase-tissue displacement mapping relationship table, a more sophisticated image registration algorithm, such as feature point-based registration or voxel-based registration, can be used to improve the accuracy of the displacement data. For real-time respiratory amplitude monitoring, high-precision respiratory sensors, such as fiber optic sensors or capacitive sensors, can be used to obtain more stable and accurate respiratory signals. When adjusting the baseline, an adaptive adjustment mechanism can be introduced, such as dynamically adjusting the parameters of the nonlinear function according to the patient's breathing pattern through a machine learning algorithm, so as to better adapt to the respiratory characteristics of different patients. In addition, the adjustment formula of the baseline can be optimized in combination with the patient's physiological characteristics, such as age, weight and lung function, to further improve the accuracy of tissue displacement prediction.

[0144] In some embodiments, the multi-scale noise suppression in step S2 includes:

[0145] S81: Use non-local mean filtering to eliminate high-frequency noise, and the filter window size is positively correlated with the image resolution;

[0146] S82: Fluorescence leakage artifacts are suppressed by anisotropic diffusion algorithm, and the diffusion coefficient is negatively correlated with the local gradient amplitude;

[0147] S83: Soft thresholding is performed on periodic noise in the wavelet transform domain to retain low-frequency fluorescence features.

[0148] It should be noted that multiscale noise suppression processing is a preprocessing method for dynamic fluorescence image sequences, which aims to eliminate noise and artifacts in the image through a variety of technical means to improve image quality and provide clearer input for subsequent tumor boundary identification. Non-local mean filtering is a denoising method based on image similarity. It removes high-frequency noise by calculating the weighted average of similar pixels in the image while retaining the image's detail information. Anisotropic diffusion algorithm is an image processing technique based on partial differential equations. It suppresses fluorescence leakage artifacts by controlling the direction and intensity of diffusion while retaining the edge information of the image. Soft threshold processing in the wavelet transform domain decomposes the image signal into different frequency levels and performs threshold processing on the high-frequency noise component to remove periodic noise and retain low-frequency fluorescence features.

[0149] Specifically, the filter window size of the non-local mean filter is positively correlated with the image resolution, that is, for high-resolution images, a larger window is needed to cover more similar pixels to achieve better noise reduction. The diffusion coefficient of the anisotropic diffusion algorithm is negatively correlated with the local gradient amplitude, which means that diffusion is reduced in areas with large image gradients (such as edges) and increased in areas with small gradients (such as smooth areas), thereby suppressing artifacts while retaining edge information. Soft thresholding in the wavelet transform domain removes noise components corresponding to small coefficients by thresholding the wavelet coefficients, while retaining low-frequency fluorescence features corresponding to large coefficients. The parameter settings of these methods need to be optimized according to the specific application scenarios and image characteristics, such as the filter window size, the initial value of the diffusion coefficient, and the attenuation rate.

[0150] Preferably, in non-local mean filtering, an adaptive window size method can be used to dynamically adjust the window size according to the complexity of the local image to achieve a better balance between noise reduction and detail retention. For the anisotropic diffusion algorithm, an adaptive adjustment mechanism of the time step can be introduced to accelerate the convergence speed of the algorithm and improve processing efficiency. In the soft threshold processing in the wavelet transform domain, the threshold can be dynamically selected in combination with the statistical characteristics of the image, such as using a histogram-based threshold selection method to better remove noise and retain useful information. In addition, a hybrid noise suppression algorithm can be designed by combining the advantages of multiple noise suppression methods to further improve the effect of image preprocessing. For example, high-frequency noise is first removed by non-local mean filtering, and then the anisotropic diffusion algorithm is used to suppress artifacts. Finally, the low-frequency features are enhanced by soft threshold processing in the wavelet transform domain, thereby providing higher quality image input for subsequent tumor boundary identification.

[0151] In some embodiments, step S7 includes:

[0152] S91: Projecting the two-dimensional boundary segmentation result to a three-dimensional space coordinate system to establish a voxel-level boundary mapping model;

[0153] S92: Generate a dynamic safety buffer zone according to the kinematic parameters of the laser device, and the buffer zone width W satisfies:

[0154] ;

[0155] Where, v represents the moving speed of the laser head, Indicates system response delay, is the safety margin factor.

[0156] It should be noted that when mapping the two-dimensional boundary segmentation results to the spatial coordinate system of the laser treatment device, a boundary mapping model including a safety buffer zone will be generated. This process aims to ensure that the optimized tumor boundary can accurately fit the operating space of the laser treatment device, while avoiding accidental injury to surrounding normal tissues through the design of the safety buffer zone. Among them, the voxel-level boundary mapping model is a three-dimensional grid-based model that can accurately represent the tumor boundary and the safety area around it, while the kinematic parameters of the laser device involve the moving speed, acceleration and other characteristics of the laser head. These parameters are used to calculate the width of the safety buffer zone to ensure the safety and effectiveness of the treatment.

[0157] Specifically, when projecting the two-dimensional boundary segmentation results to the three-dimensional space coordinate system, it is necessary to convert the pixel coordinates in the two-dimensional image into three-dimensional coordinates in the operating space of the laser treatment device. This process can be achieved through geometric transformation or interpolation algorithm. The width of the safety buffer is determined by the laser head movement speed, system response delay and safety margin coefficient. The laser head movement speed refers to the speed at which the laser moves along the predetermined path during the treatment process. The system response delay refers to the time difference between the detection of boundary information and the actual execution of the laser action. The safety margin coefficient is a parameter determined by experience or experiment, which is used to increase the additional safety distance to ensure the safety of treatment.

[0158] Preferably, when building a voxel-level boundary mapping model, a high-resolution three-dimensional grid can be used to more accurately represent the tumor boundary and the safety buffer zone. For example, a grid with a voxel size of 1 mm³ can be used to build the model, thereby improving the accuracy of the treatment. The calculation of the safety buffer zone can be optimized according to the specific parameters of the laser device. For example, for high-precision laser equipment, the safety margin coefficient can be appropriately reduced to improve the treatment efficiency; while for equipment with lower precision, the safety margin coefficient needs to be increased to ensure the safety of the treatment. In addition, a real-time feedback mechanism can be introduced to dynamically adjust the width of the safety buffer zone according to the real-time data during the laser treatment process to further improve the safety and adaptability of the treatment.

[0159] In some embodiments, it also includes:

[0160] A multi-parameter safety interlock mechanism is established to trigger treatment suspension when the following conditions are met at the same time:

[0161] (1) Boundary fuzziness index > threshold, where is the boundary gradient change, is the maximum boundary strength value;

[0162] (2) Respiratory movement mutation rate >Preset warning value, A is the breathing amplitude;

[0163] (3) Equipment positioning deviation >Accuracy tolerance, is the Euclidean distance between the actual coordinates and the theoretical coordinates.

[0164] It should be noted that the multi-parameter safety interlock mechanism established in the present invention is used to monitor key parameters in real time and trigger treatment pause during tumor image boundary identification and laser treatment to ensure the safety and reliability of treatment. The boundary fuzziness index is a parameter that measures boundary clarity by calculating the ratio of the boundary gradient change to the maximum boundary intensity value. The respiratory motion mutation rate refers to the rate of change of the respiratory amplitude per unit time, which is used to evaluate the stability of the respiratory motion. The equipment positioning deviation refers to the Euclidean distance between the actual coordinates of the laser equipment and the theoretical coordinates, which is used to monitor the positioning accuracy of the equipment. When these parameters exceed the preset threshold, the treatment pause is triggered to avoid treatment risks caused by inaccurate boundary identification, abnormal respiratory movement or equipment positioning deviation.

[0165] Specifically, the calculation of the boundary fuzziness index is based on the boundary gradient change and the maximum boundary intensity value. The boundary gradient change reflects the degree of change in the pixel intensity of the boundary area, while the maximum boundary intensity value is the maximum value of the pixel intensity of the boundary area. The calculation of the respiratory motion mutation rate requires real-time monitoring of the changes in the respiratory amplitude and calculating its change rate by the differential method. The calculation of the device positioning deviation requires real-time acquisition of the actual coordinates and theoretical coordinates of the laser device, and the deviation value is calculated by the Euclidean distance formula. The preset warning value and precision tolerance are thresholds determined based on experimental data and clinical experience, which are used to determine whether to trigger a treatment pause. For example, the threshold of the boundary fuzziness index can be set to 0.1, the warning value of the respiratory motion mutation rate can be set to 0.5, and the precision tolerance of the device positioning deviation can be set to 1mm.

[0166] Preferably, a more sophisticated monitoring method can be introduced in the multi-parameter safety interlock mechanism. For example, for the boundary fuzziness index, a dynamic evaluation can be performed in combination with the local area of ​​the image, rather than a single global calculation, so as to more accurately reflect the changes in boundary clarity. For the respiratory motion mutation rate, the sliding window method can be used to calculate its short-term change trend to improve the monitoring sensitivity. For equipment positioning deviation, a predictive deviation assessment can be performed in combination with the kinematic model of the laser equipment to detect potential positioning problems in advance. In addition, a redundant monitoring mechanism can be introduced, such as using multiple sensors to simultaneously monitor respiratory motion or equipment positioning, to improve the reliability and accuracy of the monitoring results.

[0167] The above-mentioned embodiments of the present invention have the following beneficial effects: the present invention can improve the accuracy and stability of tumor image boundary recognition by fusing fluorescence boundaries and tumor tissue features and combining technical means such as multi-scale noise suppression, dynamic feature extraction, respiratory motion compensation, boundary optimization, and safety buffer design. In the image preprocessing stage, non-local mean filtering, anisotropic diffusion and wavelet transform are used to eliminate high-frequency noise, suppress fluorescence leakage artifacts and retain low-frequency fluorescence features, thereby improving image quality and providing clearer input for subsequent boundary extraction. In the feature extraction and fusion process, a dynamic feature extraction network is constructed and a fluorescence attenuation compensation factor is introduced to extract fluorescence intensity distribution characteristics and compensate for fluorescence attenuation differences. At the same time, combined with the texture structure characteristics of tumor tissue, the initial tumor boundary contour containing multi-dimensional feature weights can be generated by dynamically calculating the fusion weights, further improving the accuracy of boundary recognition. Dynamic interpolation correction of the intensity distribution matrix through the respiratory motion compensation model can solve the boundary displacement problem caused by respiratory motion and ensure the stability of boundary recognition. In the boundary optimization stage, dynamic threshold segmentation and multi-scale morphological optimization can be used to eliminate boundary discontinuity and pseudo-boundary residues and generate a more accurate tumor boundary contour. The optimized boundary is mapped to the spatial coordinate system of the laser treatment device and a boundary mapping model is generated, which can provide an accurate positioning basis for laser treatment. At the same time, combined with the dynamic safety buffer design, the buffer width can be adjusted in real time according to the kinematic parameters of the laser device, further improving the safety of treatment.

[0168] In addition, the present invention also establishes a multi-parameter safety interlock mechanism, which can timely trigger a treatment pause when the boundary fuzziness exceeds the threshold, the respiratory movement mutation rate exceeds the warning value, or the equipment positioning deviation exceeds the accuracy tolerance, thereby effectively avoiding treatment risks caused by inaccurate boundary identification or equipment operation errors, and improving the safety and reliability of laser treatment; through preoperative calibration and real-time correction of tissue displacement reference, the influence of respiratory movement on the boundary can be more accurately compensated, and the dynamic adaptability of boundary identification can be further improved; overall, the present invention can provide more efficient and safer technical support for the precise diagnosis and laser treatment of tumors.

[0169] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or terminal devices such as a computer, a server, a mobile phone, and a tablet.

[0170] The above descriptions are only some preferred embodiments of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A tumor image boundary recognition method integrating fluorescence boundary, characterized in that: The following steps are involved: The dynamic fluorescence image sequence of the tumor area is continuously collected by the fluorescence imaging device, and the real-time respiratory waveform data output by the respiratory motion monitoring sensor is simultaneously obtained; Performing multi-scale noise suppression processing on the dynamic fluorescence image sequence, adaptively adjusting the filtering parameters based on the image resolution, and generating a pre-processed fluorescence image with artifacts removed; The spatial fluorescence intensity distribution features are extracted from the preprocessed fluorescence image through a multi-layer feature weighted fusion network to generate a fluorescence intensity distribution matrix that integrates multi-scale features. A motion compensation model is constructed based on real-time respiratory waveform data, and the intensity distribution matrix is ​​corrected for boundary displacement through phase angle tracking and motion vector prediction to obtain the corrected fluorescence intensity distribution characteristics. The corrected fluorescence intensity distribution features are fused with the texture structure features of the tumor tissue to generate an initial tumor boundary contour containing multi-dimensional feature weights; Iterative topological optimization is performed on the initial tumor boundary contour, and boundary discontinuities are eliminated through dynamic threshold segmentation and morphological operations to obtain the optimized tumor boundary. The optimized tumor boundary is mapped to the spatial coordinate system of the laser treatment device, and a boundary mapping model including a safety buffer zone is generated and updated in real time.

2. The method according to claim 1, characterized in that The spatial fluorescence intensity distribution features extracted from the preprocessed fluorescence image through a multi-layer feature weighted fusion network include: Three layers of convolution kernels with different scales are used to extract local features, regional features and global features of the fluorescence image respectively; The features of each layer are corrected by the feature attenuation compensation formula, which is expressed as: ； in, represents pixel coordinates, represents the fluorescence intensity feature extracted by the i-th convolution kernel, represents the fluorescence diffusion attenuation coefficient, represents the fluorescence attenuation compensation factor, represents the fluorescence intensity feature extracted by the i-th convolution kernel after correction; Generating a fluorescence intensity distribution matrix that fuses multi-scale features includes: performing weighted fusion on the corrected features to generate a fluorescence intensity distribution matrix, which is expressed as: ； in, is the fluorescence intensity distribution matrix, is the weight coefficient of the i-th feature layer.

3. The method according to claim 2, characterized in that The corrected fluorescence intensity distribution characteristics are obtained by: Perform fast Fourier transform on the respiratory waveform data, extract the dominant frequency component and calculate the respiratory phase angle; Based on the respiratory phase angle, a respiratory motion compensation coefficient model is constructed to obtain a respiratory motion compensation coefficient. The respiratory motion compensation coefficient model is expressed as: ； in, represents the respiratory phase angle, represents the spatial displacement weight, represents the speed impact factor, represents the respiratory phase change rate, C represents the respiratory motion compensation coefficient; The fluorescence intensity distribution matrix is ​​dynamically interpolated and corrected according to the respiratory motion compensation coefficient to obtain the corrected fluorescence intensity distribution characteristics, wherein the corrected matrix satisfies the following conditions: ； in, , is the tissue displacement reference value calibrated before surgery, and F is the fluorescence intensity distribution matrix function.

4. The method according to claim 3, characterized in that The method of fusing the corrected fluorescence intensity distribution feature with the texture structure feature of the tumor tissue to generate an initial tumor boundary contour containing multi-dimensional feature weights includes: The fusion weights of fluorescence features and tissue features are calculated as follows: ； in, represents the standard deviation of the fluorescence intensity distribution matrix F, Represents the standard deviation of the texture feature matrix T; Based on the fusion weight and the corrected fluorescence intensity distribution characteristics, a fusion boundary feature map is generated, which is expressed as follows: ； in, represents the fused boundary feature map, represents the fusion weight, Represents the texture feature matrix; represents the corrected fluorescence intensity distribution matrix; The fused boundary feature map is binarized to generate the initial tumor boundary contour.

5. The method according to claim 4, characterized in that The binarization process of the fused boundary feature map to generate the initial tumor boundary contour comprises: The dynamic segmentation threshold is calculated based on the fused boundary feature map, expressed as: ； Among them, τ represents the dynamic segmentation threshold, N represents the total number of pixels, represents the mean of the feature map, is the sensitivity adjustment coefficient, Represents the fused boundary feature map; Compare the fused boundary feature map with the dynamic segmentation threshold. If the pixel of the fused boundary feature map is greater than the segmentation threshold, the comparison result is 1; if the pixel of the fused boundary feature map is less than the segmentation threshold, the comparison result is 0. generating an initial tumor boundary contour based on the comparison results; Multi-scale morphological optimization is performed, and the size of the structure element decreases proportionally to the number of iterations to eliminate pseudo-boundary residues.

6. The method according to claim 2, characterized in that The fluorescence attenuation compensation factor is determined by the following method: Use fluorescence calibration films at different depths to conduct benchmark attenuation experiments to obtain the initial value of the fluorescence attenuation compensation factor; The fluorescence attenuation compensation factor is dynamically corrected according to the real-time detected tissue optical parameters, as shown below: ； in, represents the nominal absorbance, is the tissue variability coefficient, represents the fluorescence attenuation compensation factor, represents the tissue optical parameters, Indicates the initial value.

7. The method according to claim 3, characterized in that The tissue displacement reference amount is determined as follows: Based on the preoperative four-dimensional CT images, a respiratory phase-tissue displacement mapping relationship table was constructed to record the tissue displacement corresponding to different respiratory phases; Querying a reference displacement from a mapping relationship table based on the real-time respiratory phase; The reference displacement is adjusted according to the real-time breathing amplitude A, and the adjustment formula is: ； ； in, , is the reference displacement, is the breathing amplitude threshold.

8. The method according to claim 1, characterized in that The performing multi-scale noise suppression processing on the dynamic fluorescence image sequence comprises: Non-local mean filtering is used to eliminate high-frequency noise in dynamic fluorescence image sequences, and the filter window size is positively correlated with the image resolution. Fluorescence leakage artifacts are suppressed by anisotropic diffusion algorithm, and the diffusion coefficient is negatively correlated with the local gradient amplitude; The periodic noise is soft-thresholded in the wavelet transform domain to retain the low-frequency fluorescence features.

9. The method according to claim 1, characterized in that: Mapping the optimized tumor boundary to the spatial coordinate system of the laser treatment device comprises the following steps: The optimized two-dimensional boundary segmentation result of the tumor boundary is projected into a three-dimensional space coordinate system to establish a voxel-level boundary mapping model; A dynamic safety buffer is generated based on the kinematic parameters of the laser device. The buffer width is expressed as follows: ； Where, v represents the moving speed of the laser head, Indicates system response delay, is the safety margin factor, Indicates the buffer width; A buffer layer with a thickness of W is generated on the outer surface of the boundary mapping model.

10. A tumor image boundary recognition system integrating fluorescence boundaries, characterized in that: It includes a data acquisition module, a preprocessing module, a matrix generation module, a feature correction module, a contour generation module, a contour optimization module and a data mapping module; A data acquisition module continuously acquires a dynamic fluorescence image sequence of the tumor area through a fluorescence imaging device, and synchronously acquires real-time respiratory waveform data output by a respiratory motion monitoring sensor; A preprocessing module performs multi-scale noise suppression processing on the dynamic fluorescence image sequence, adaptively adjusts the filtering parameters based on the image resolution, and generates a preprocessed fluorescence image with artifacts removed; The matrix generation module extracts the spatial fluorescence intensity distribution features from the preprocessed fluorescence image through a multi-layer feature weighted fusion network to generate an intensity distribution matrix containing the fluorescence diffusion attenuation characteristics; The feature correction module builds a motion compensation model based on real-time respiratory waveform data, performs boundary displacement correction on the intensity distribution matrix through phase angle tracking and motion vector prediction, and obtains the corrected fluorescence intensity distribution feature; The contour generation module fuses the corrected fluorescence intensity distribution features with the texture structure features of the tumor tissue to generate an initial tumor boundary contour containing multi-dimensional feature weights; The contour optimization module performs iterative topological optimization on the initial tumor boundary contour, eliminates boundary discontinuities through dynamic threshold segmentation and morphological operations, and obtains the optimized tumor boundary; The data mapping module maps the optimized tumor boundary to the spatial coordinate system of the laser treatment device, generates a boundary mapping model including a safety buffer zone, and updates it in real time.

Images