Automatic tumor target volume delineation analysis system based on multi-modal radiological image fusion

Abstract

The application discloses a tumor target area automatic delineation analysis system based on multi-modal radiological image fusion, which comprises five functional modules. The image registration module completes CT, MRI rigid registration and B-spline deformation registration, and outputs multi-modal registration signals; the multi-modal feature segmentation module generates an initial tumor probability graph by relying on independent CT and MRI encoders, a shared decoder and a hierarchical cross-modal channel attention unit; the anatomical probability constraint module filters voxels that do not conform to physiological structures by using an anatomical atlas to obtain a constrained probability graph; the conditional random field optimization module combines the probability graph and image edges to construct a fully connected conditional random field iterative optimization, and outputs a tumor target area mask; and the quantitative analysis module calculates the tumor volume, the maximum three-dimensional diameter and the gray level co-occurrence matrix texture parameters based on the mask and outputs analysis results. The application makes up for the problem of missing information of single-mode image, weakens subjective errors of manual delineation, and improves the efficiency and accuracy of tumor target area delineation.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and radiotherapy planning technology, specifically to an automatic tumor target delineation and analysis system based on multimodal radiographic image fusion. Background Technology

[0002] In tumor radiotherapy, accurate target delineation is a crucial prerequisite for radiotherapy planning, directly determining dose distribution and therapeutic efficacy. Clinically, computed tomography (CT) and magnetic resonance imaging (MRI) are widely used to guide target identification. CT clearly displays high-density tissues such as bone, while MRI offers high contrast resolution for soft tissue tumor and edema boundaries. However, single-modal imaging has inherent limitations; CT can blur the boundaries of some soft tissue tumors, and MRI is susceptible to geometric distortion and signal inhomogeneity. Traditional target delineation is performed manually layer by layer by physicians, taking tens of minutes. The results are highly dependent on individual experience, and consistency between different physicians and even within the same physician is difficult to guarantee. This manual approach has become a bottleneck restricting the accuracy and efficiency of radiotherapy. Deep learning-based automatic segmentation methods offer a potential solution, but most existing methods are based on single-modal imaging and cannot capture multimodal complementary information, resulting in insufficient ability to discriminate tumor boundaries in complex cases. Some multimodal fusion schemes simply stitch together two images at the input or feature layer without modeling deep semantic interactions across modalities, making it difficult to highlight the unique characteristics of tumors in multimodal contexts. Furthermore, these methods generally lack integration of anatomical priors, often resulting in segmentation results that violate normal anatomical structures, misclassifying adjacent bone, blood vessels, or hollow organs as target areas. The probability maps output by segmentation networks have blurred boundaries; directly converting them into masks produces jagged edges and isolated noise. Even with post-processing using conditional random fields, they often only utilize single-modal gradient information, failing to leverage the complementary advantages of multimodal edges. Moreover, existing systems often stop at contour output, lacking automatic extraction of quantitative indicators such as tumor volume, maximum 3D diameter, and texture heterogeneity, resulting in insufficient clinical practicality and completeness. Therefore, there is an urgent need for an automatic tumor target delineation scheme that can deeply fuse multimodal image information and embed anatomical constraints and boundary optimization. Summary of the Invention

[0003] In view of the shortcomings of the prior art, the purpose of this invention is to provide an automatic tumor target delineation and analysis system based on multimodal radiographic image fusion, which solves the problems of insufficient information from single-modal images and the high subjectivity and low efficiency of manual delineation. This invention forms aligned image pairs by rigid and deformation registration of computed tomography (CT) and magnetic resonance imaging (MRI) images. An initial tumor probability map is generated by a multi-feature segmentation network containing independent encoders and cross-modal channel attention fusion units. Spatial filtering is performed using a pre-built anatomical probability map, and the boundary is iteratively optimized using a fully connected conditional random field combined with multimodal edge signals. Finally, a target mask is output, and the volume, three-dimensional diameter, and texture parameters are calculated, achieving fully automatic delineation.

[0004] This invention provides an automated tumor target delineation and analysis system based on multimodal radiographic image fusion, comprising: The image registration module performs rigid registration and B-spline-based deformation registration on CT and MRI images to generate multimodal registration signals. The multimodal feature segmentation module receives multimodal registration signals and generates an initial tumor probability map signal through a multi-feature fusion network containing independent CT encoders, MRI encoders, and a shared decoder. The encoder has cross-modal channel attention fusion units between each level to pass the weighted fused feature map to the decoder. The anatomical probability constraint module receives the initial tumor probability map signal, uses a pre-built anatomical probability map atlas to spatially filter the probability map, removes voxels that do not conform to the anatomical constraints, and forms a constraint probability map signal. The Conditional Random Field (CRF) optimization module receives the constraint probability map signal and the gradient edge signal from CT images. It constructs a fully connected CRF with the constraint probability map as a univariate potential and the gradient edge as a binary potential. After iterative evolution, it forms a tumor target mask signal. The quantitative analysis module receives the tumor target area mask signal, calculates and outputs the analysis result signal, which includes volume, maximum three-dimensional diameter, and texture parameters based on the gray-level co-occurrence matrix.

[0005] In one embodiment of the present invention, the image registration module uses CT images as fixed images and MRI images as floating images. First, it calculates the normalized mutual information similarity measure between the two images. Then, it uses gradient descent to search for rigid transformation parameters that maximize the mutual information value. These rigid transformation parameters include three translational components and three rotational components. After rigid registration, a B-spline free-deformation mesh with consistent control point spacing is constructed based on the rigid registration results. Using normalized mutual information as the objective function, the floating image is deformed step-by-step by optimizing the displacement vector of each control point. Iteration stops when the increment of the objective function is lower than a preset convergence threshold. Finally, the deformed MRI and CT images are resampled to the same spatial coordinate system to form a multimodal registration signal. The formula for calculating normalized mutual information is as follows: in, The output value is the normalized mutual information function value, where F is the gray-level distribution variable of the fixed CT image and M is the gray-level distribution variable of the deformed floating MRI image. For rigid transformation of three-dimensional translation and rotation parameter vectors, Let be the displacement vector of the control points of the B-spline deformed mesh. For indexing the grayscale range of CT images, For traversing the voxel space of a 3D image. The gray-voxel probability distribution function of CT images and deformed MRI images. Let be the probability distribution function of grayscale edges in CT single-modality mode. This represents the marginal probability distribution function of a single-modal voxel in deformation MRI.

[0006] In one embodiment of the present invention, the cross-modal channel attention fusion unit performs global average pooling on the feature maps output by the CT encoder and MRI encoder at the same level, compressing the feature map of each channel into a scalar value to obtain a first channel descriptor vector and a second channel descriptor vector. The first channel descriptor vector and the second channel descriptor vector share the same weight generation network, which is composed of a fully connected layer, a linear rectified activation function, a second fully connected layer, and a sigmoid activation function connected in sequence. This weight generation network outputs a CT channel weight vector and an MRI channel weight vector to the first channel descriptor vector and the second channel descriptor vector, respectively. The CT channel weight vector is multiplied with the feature map output by the CT encoder along the channel dimension, and the MRI channel weight vector is multiplied with the feature map output by the MRI encoder along the channel dimension to obtain a weighted first feature map and a weighted second feature map. The weighted first feature map and the weighted second feature map are added element-wise, and then the number of channels is restored to the number of channels before addition through a convolutional layer with a kernel size of 1x1, forming the fused feature map of this level passed to the corresponding level of the shared decoder. The calculation formula of the CT channel weight vector is as follows: Where w is the channel weight vector identifier, CT is the CT encoder-specific distinguishing identifier, and c is the index of the currently calculated feature channel. This is the identifier for the Sigmoid activation function. This is the weight matrix for the second fully connected layer. This is the identifier for the linear rectification activation function. Here, represents the weight matrix of the first fully connected layer, H represents the number of pixels in the height of a single-channel feature map, W represents the number of pixels in the width of a single-channel feature map, i represents the traversal index in the height direction of the feature map, and j represents the traversal index in the width direction of the feature map. This is the original feature map of the c-th channel of the CT encoder, where (i,j) represents the feature vector. Figure 2 3D pixel coordinate index, This is the bias vector for the first fully connected layer. This is the bias vector for the second fully connected layer.

[0007] In one embodiment of the present invention, the anatomical probability constraint module internally deploys a pre-trained multi-organ segmentation network. This multi-organ segmentation network receives multimodal registration signals and performs organ region segmentation on CT images and MRI images respectively, outputting semantic segmentation results containing multiple organ labels. The anatomical probability constraint module extracts organ mask regions that allow tumor occurrence from the semantic segmentation results according to a preset list of organs associated with lesion occurrence, and simultaneously extracts organ mask regions that prohibit tumor occurrence from the semantic segmentation results. The probability values ​​of voxels falling into the organ mask regions that prohibit tumor occurrence in the initial tumor probability map signal are set to zero, and the probability values ​​of voxels falling outside the organ mask regions that allow tumor occurrence are also set to zero. For the retained non-zero probability voxels, the distance transformation value between them and the allowed organ mask regions is further calculated. When the distance transformation value is lower than a preset distance threshold, the probability value of the voxel remains unchanged. When the distance transformation value exceeds the preset distance threshold, the probability value of the voxel is proportionally attenuated. After completing the above operations, a constraint probability map signal is formed.

[0008] In one embodiment of the present invention, the anatomical probability constraint module performs connected component analysis before generating the constraint probability map signal. It performs threshold binarization on the probability map after probability zeroing and attenuation processing, extracts all spatially connected candidate regions, calculates the volume of each connected region, deletes connected regions with volumes smaller than a preset minimum tumor volume, and for connected regions with volumes larger than a preset maximum tumor volume, shrinks them inward along the contour normal direction to the number of voxels corresponding to the preset maximum tumor volume while preserving their outer contour. It then performs three-dimensional shape analysis based on principal component analysis on the preserved connected regions, extracting the major axis, median axis, and minor axis directions and their corresponding length values. It calculates the ratio of the major axis to the median axis and the ratio of the median axis to the minor axis. When the ratio exceeds a preset aspect ratio threshold, the connected region is determined to be a non-nodular morphology and is removed. The constraint probability map signal is formed through the above connected component analysis.

[0009] In one embodiment of the present invention, the anatomical probability constraint module further performs edge normalization on the constraint probability map signal by constructing a shape prior. The shape prior is constructed by taking the voxel with the highest probability value in the initial tumor probability map signal as the seed point, emitting virtual rays from the seed point along the direction vector of uniform sampling in three-dimensional space, recording the voxel coordinates on each virtual ray where the probability value first falls below a preset cutoff threshold, connecting all the recorded voxel coordinates to form a three-dimensional convex hull, setting the voxel probability value inside the three-dimensional convex hull as the original probability value of the voxel in the initial tumor probability map signal, and setting the voxel probability value outside the three-dimensional convex hull as zero, so that the boundaries of each connected region in the constraint probability map signal tend to be smooth.

[0010] In one embodiment of the present invention, the energy function of the fully connected conditional random field constructed by the conditional random field optimization module consists of a univariate potential term and a binary potential term. The univariate potential term is directly taken from the negative logarithm of the probability value of each voxel in the constraint probability map signal, representing the independent cost of marking the voxel as tumor or background. The binary potential term consists of a weighted sum of two Gaussian kernel functions. The first Gaussian kernel function takes the spatial position difference of the voxel and the gradient magnitude difference of the corresponding voxel in the gradient edge signal from the CT image as input. The spatial position difference is calculated by dividing the square of the Euclidean distance between the two voxels by the square of the first spatial bandwidth parameter. The gradient magnitude difference is calculated by dividing the square of the difference between the gradient magnitudes of two voxels by the square of the first gradient bandwidth parameter. The second Gaussian kernel function takes only the spatial position difference of voxels as input and is calculated by dividing the square of the Euclidean distance between two voxels by the square of the second spatial bandwidth parameter. The first Gaussian kernel function tends to assign the same label to adjacent voxels with similar gradient magnitudes. The second Gaussian kernel function is used to eliminate isolated small false detection regions. The univariate potential term and the binary potential term are summed to form the global energy function. The average field approximation algorithm is used to iteratively update the label probability distribution of each voxel until the energy function converges and the tumor target area mask signal is output.

[0011] In one embodiment of the present invention, the gradient edge signal derived from CT images is fused with the phase-consistency edge signal derived from MRI images before being input into the binary potential term. The phase-consistency edge signal of the MRI images is obtained by calculating the Fourier domain phase consistency value of the MRI images at multiple scales and in multiple directions. For each voxel, the maximum value of the phase consistency value in each direction is taken as the phase-consistency edge response of that voxel. The local signal-to-noise ratio of the CT images at each voxel and the local signal-to-noise ratio of the MRI images at each voxel are calculated. The ratio of the local signal-to-noise ratio of CT to the local signal-to-noise ratio of MRI is used as a weighting coefficient. The CT gradient amplitude and the MRI phase-consistency edge response are weighted and summed to generate the fused edge signal. The fused edge signal replaces the gradient edge signal derived from the CT images as the gradient amplitude difference input of the first Gaussian kernel function in the binary potential term.

[0012] In one embodiment of the present invention, the quantitative analysis module performs connected component labeling on the tumor target area mask signal, identifies each independent tumor region, counts the total number of voxels contained in each tumor region and multiplies it by the physical volume of a single voxel to obtain the tumor volume, traverses all voxel pairs within the tumor region to calculate the Euclidean distance between each voxel pair, takes the maximum value among all Euclidean distances as the maximum three-dimensional diameter, constructs a gray-level co-occurrence matrix within the tumor region, calculates four texture feature values ​​of the gray-level co-occurrence matrix: energy, contrast, correlation, and homogeneity, and organizes the tumor volume, maximum three-dimensional diameter, energy, contrast, correlation, and homogeneity into a structured analysis result signal and outputs it.

[0013] In one embodiment of the present invention, the multi-feature fusion network in the multimodal feature segmentation module uses a joint loss function for parameter optimization during the training phase. The joint loss function is composed of a weighted sum of a first loss term, a second loss term, and a third loss term. The first loss term is the sum of the Dessian loss and the focal loss between the initial tumor probability map signal and the manually labeled tumor mask. The second loss term is the weighted cross-entropy loss between the constraint probability map signal and the manually labeled tumor mask. In the manually labeled tumor mask, voxels with a voxel value of one are assigned a first weight, and voxels with a voxel value of zero are assigned a second weight, and the first weight is greater than the second weight. The third loss term is the regularization loss obtained by taking the sum of the squares of the second norm of all trainable parameters of the network. The first loss term, the second loss term, and the third loss term are multiplied by preset first coefficients, second coefficients, and third coefficients, respectively, and then summed to serve as the objective function for backpropagation.

[0014] The present invention provides an automatic tumor target delineation and analysis system based on multimodal radiographic image fusion. The present invention forms aligned image pairs by performing rigid and deformation registration on computed tomography and magnetic resonance images. An initial tumor probability map is generated by a multi-feature segmentation network containing independent encoders and cross-modal channel attention fusion units. Spatial filtering is performed using a pre-built anatomical probability map. The boundary is then iteratively optimized by combining a fully connected conditional random field with multimodal edge signals. Finally, a target mask is output and the volume, three-dimensional diameter, and texture parameters are calculated to achieve fully automatic delineation. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1This is a system architecture diagram of an automated tumor target delineation and analysis system based on multimodal radiographic image fusion. Figure 2 A flowchart illustrating the internal breakdown of the image registration module; Figure 3 This is a flowchart for multimodal segmentation, anatomical constraints, and CRF refinement. Detailed Implementation

[0017] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0018] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0019] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0020] Please see Figure 1-3The image shows an automatic tumor target delineation and analysis system based on multimodal radiographic image fusion according to the present invention. It includes an image registration module that performs rigid registration and B-spline-based deformation registration on CT and MRI images to form a multimodal registration signal; a multimodal feature segmentation module that receives the multimodal registration signal and generates an initial tumor probability map signal through a multi-feature fusion network comprising an independent CT encoder, an MRI encoder, and a shared decoder. The encoder layers are equipped with cross-modal channel attention fusion units to transmit the weighted fused feature map to the decoder; and an anatomical probability constraint module. The system receives the initial tumor probability map signal, uses a pre-built anatomical probability map to spatially filter the probability map, removes voxels that do not conform to anatomical constraints, and forms a constrained probability map signal; the conditional random field optimization module receives the constrained probability map signal and gradient edge signals from CT images, constructs a fully connected conditional random field with the constrained probability map as a univariate potential and the gradient edges as a binary potential, and forms a tumor target mask signal after iterative evolution; the quantization analysis module receives the tumor target mask signal, calculates and outputs the analysis result signal containing volume, maximum three-dimensional diameter, and texture parameters based on the gray-level co-occurrence matrix.

[0021] Figure 1 As shown in the flowchart, this is the overall vertical workflow of the complete system for automatic delineation of tumor target areas using multimodal radiographic imaging. It includes nine functional nodes, and the process involves branch selection logic and iterative loop logic, rather than a single linear serial link. The starting node is the CT / MRI image input. This node is responsible for receiving the raw data from two sets of clinically acquired multimodal radiographic images, which serve as the raw input source for all operations of the entire system. Without any spatial alignment or feature extraction processing, the data is directly transmitted down to the image registration module. After receiving the two types of raw images, the image registration module sequentially performs rigid spatial alignment and B-spline free deformation fine registration. Relying on normalized mutual information measures, it continuously optimizes spatial transformation parameters, unifying the spatially misaligned CT and MRI images to the same three-dimensional spatial coordinate system, outputting a standardized multimodal registration signal, and completing the pre-processing of cross-modal image spatial matching. After processing, the data flows to the branch decision node for whether to enable anatomical constraints. This branch node provides two independent data transmission paths. The first path enables constraint processing, with data flowing directly into the anatomical probability constraint module. The second path disables constraint processing, with data directly jumping to the conditional random field optimization module. After receiving the initial tumor probability map signal, the anatomical probability constraint module performs organ semantic segmentation using a pre-trained multi-organ segmentation network. Through multiple operations such as organ mask filtering, distance transform probability attenuation, connected component filtering, and 3D convex hull shape prior regularization, it removes pseudo-tumor voxels that do not conform to human anatomical and physiological structures, outputting a constraint probability map signal with significantly reduced noise. This processed signal is then uniformly sent to the conditional random field optimization module.

[0022] The Conditional Random Field (CRF) optimization module integrates constrained probability map signals with multimodal fusion edge signals to construct a fully connected CRF energy function containing univariate and dual-Gaussian kernel binary potentials. It continuously updates the probability distribution of all element labels using a mean-field approximation algorithm. After each iteration, data flows into a convergence check node. This node forms the core loop of the system. If the global energy value has not reached a stable convergence state, the data will flow back to the CRF optimization module for iterative evolution, continuously correcting the tumor region boundary. If the energy value converges stably, the smoothed and complete tumor target mask signal is passed down to the quantification analysis module. The quantification analysis module performs three-dimensional quantitative calculations on each independent tumor connected domain, statistically analyzing tumor volume and maximum three-dimensional diameter. Simultaneously, it extracts four types of tumor texture features based on the gray-level co-occurrence matrix, integrating all quantitative indicators into a standardized structured analysis result signal. This signal is then transmitted to the structured analysis result output node, completing the entire system's computational process and outputting quantitative report data that can be directly used for clinical radiotherapy evaluation. The entire process relies on branch selection to enable controllable switching of anatomical constraint functions, and on iterative loops to achieve refined optimization of target area boundaries. Each node has a clear division of labor, and data flows in one direction with local backflow loops, fully covering the entire technical logic of multimodal imaging from input to quantitative indicator output.

[0023] The image registration module is responsible for aligning computed tomography (CT) images and magnetic resonance imaging (MRI) images to the same spatial coordinates. CT images reflect the electron density distribution of tissues, while MRI images reflect the proton relaxation characteristics of tissues. These two images differ fundamentally in spatial resolution, geometric accuracy, and tissue contrast mechanisms, and direct pixel-level fusion would result in positional bias. Therefore, this module uses the CT image as a fixed reference image and the MRI image as a floating image to be transformed, eliminating spatial offset in two stages. The first stage performs rigid registration based on normalized mutual information. Normalized mutual information, as a similarity measure, can measure the statistical correlation of grayscale distributions at corresponding voxel positions in two images from different imaging modalities, without requiring linear assumptions about image grayscale values. The optimizer searches for rigid transformation parameters containing three translations and three rotations in a six-dimensional parameter space. When the mutual information value reaches its maximum, the two images achieve initial alignment in global position and orientation. The second stage superimposes a non-rigid registration based on a B-spline free deformation model onto the rigid registration result. This is achieved by constructing an equally spaced grid of control points within the image space, assigning each control point an independent 3D displacement vector. The displacements between control points are interpolated using B-spline basis functions to obtain a dense deformation field. The free deformation grid covers the entire image domain, iteratively updating the displacement values ​​of each control point to further increase the normalized mutual information until the increment of the objective function falls below a preset convergence threshold. Finally, the floating image is resampled voxel-by-voxel through this deformation field, achieving voxel-level spatial correspondence with the fixed image, forming a multimodal registration signal that is then transmitted to subsequent modules.

[0024] After receiving the multimodal registration signal, the multimodal feature segmentation module feeds the spatially aligned computed tomography (CT) images and magnetic resonance imaging (MRI) images into two independent encoder branches. Both encoders employ a combination of convolutional stacking and downsampling to extract multi-scale feature maps from shallow texture to deep semantics layer by layer. The outputs of the CT encoder and MRI encoder at each level are not directly concatenated but adaptively integrated through a cross-modal channel attention fusion unit. This fusion unit performs global average pooling on both feature maps, compressing the spatial information of each channel into a channel descriptor scalar, allowing subsequent operations to focus on feature selection at the channel dimension. The two sets of channel descriptors share a weight generation network consisting of a fully connected layer, a linear rectified activation function, a second fully connected layer, and a sigmoid activation function in sequence. This network has a non-linear bottleneck structure, first compressing the number of channels and then restoring them, learning the dependencies between channels, and finally outputting the CT channel weight vector and the MRI channel weight vector. The two weight vectors are multiplied channel-by-channel along the channel dimension by their respective source feature maps to achieve input-driven adaptive channel recalibration, enhancing high-response channels and suppressing low-response channels. The weighted feature maps are then element-wise summed, fusing the representations of the same spatial location from both modalities into a single entity. This is then passed through a 1x1 convolutional layer to restore the number of channels to the original number of channels in a single feature branch before fusion. This structure allows the network to dynamically adjust the fusion strategy at each scale based on the information contribution of the two modalities, rather than forcibly adding equal weights. The fused multi-scale feature map is passed to the shared decoder via skip connections. The shared decoder gradually restores the spatial resolution through upsampling and convolution operations, ultimately outputting an initial tumor probability map signal. This signal is a three-dimensional volumetric data point, where each voxel has a value between zero and one, representing the probability that the voxel belongs to a tumor region. The formula for calculating the CT channel weight vector is as follows: Where w is the channel weight vector identifier, CT is the CT encoder-specific distinguishing identifier, and c is the index of the currently calculated feature channel. This is the identifier for the Sigmoid activation function. This is the weight matrix for the second fully connected layer. This is the identifier for the linear rectification activation function. Here, represents the weight matrix of the first fully connected layer, H represents the number of pixels in the height of a single-channel feature map, W represents the number of pixels in the width of a single-channel feature map, i represents the traversal index in the height direction of the feature map, and j represents the traversal index in the width direction of the feature map. This is the original feature map of the c-th channel of the CT encoder, where (i,j) represents the feature vector. Figure 2 3D pixel coordinate index, This is the bias vector for the first fully connected layer. This is the bias vector for the second fully connected layer. In actual computation, the cross-modal channel attention fusion unit simultaneously performs the same structure weight generation calculation on the MRI encoder features to obtain the MRI channel weight vector. The two sets of weights are then multiplied along the channel dimension by the corresponding original feature maps to complete the weighting. The weighted CT and MRI features are then added element-wise, and a fusion feature map is generated after unifying the channel dimension through a 1×1 convolution and fed into the shared decoder. This formula automatically assigns weights based on the representation ability of single-channel features for tumor lesions. Channels with strong tumor feature responses receive high weights, while channels dominated by background noise have weights close to zero. This addresses the modal information imbalance problem where CT images show prominent grayscale of bones and soft tissues, while MRI images show rich details of soft tissue lesions. Each level of the encoder independently runs this weight calculation formula, completing cross-modal feature complementary fusion layer by layer. Finally, multiple layers of fused features are stacked and input into the shared decoder to generate the initial tumor probability map, forming a complete differentiable attention calculation link. This supports end-to-end backpropagation training of the multi-feature fusion network. If this adaptive weight calculation is canceled, the CT and MRI... Features will be fused by indiscriminate direct addition, which is prone to the defect that single modality information overwhelms lesion features, significantly reducing the sensitivity of tumor region segmentation.

[0025] The anatomical probability constraint module receives the initial tumor probability map signal and uses prior anatomical knowledge to filter out false positive regions that do not conform to anatomical patterns. This module pre-loads an anatomical probability atlas constructed from statistics of a large number of labeled cases, which records the probability distribution of each organ's location in a standard spatial coordinate system. Simultaneously, the module deploys a pre-trained multi-organ segmentation network that directly processes multimodal registration signals, performs semantic segmentation on normal organs in computed tomography and magnetic resonance imaging, and outputs organ labels for each voxel. Based on the anatomical patterns of tumor development, the module predefines a list of organs that allow tumors and a list of organs that prohibit tumors, extracting corresponding binary masks from the semantic segmentation results. For each non-zero probability voxel in the initial tumor probability map signal, it checks which organ mask its spatial coordinates fall into. If the voxel is within the mask of an organ in the prohibited list, its probability value is set to zero; if the voxel is outside the mask of any organ in the allowed list, it is also set to zero. After this filtering step, the remaining voxels are all within the allowable organ range. The module further calculates the distance transformation value from these voxels to the boundary of the organ mask. If the distance value exceeds a preset threshold, it means that the voxel is too close to or embedded outside the organ wall, and its probability value is reduced proportionally, thus weakening its contribution in subsequent processing. After zeroing and attenuating the probability, the module performs connected component analysis on the entire probability map. After binarization with a preset probability threshold, all spatially connected candidate regions are marked, and scattered fragments with too small a volume are removed. For regions with too large a volume, they are shrunk inward along the contour normal direction to a reasonable upper limit, and the principal component directions of the connected regions are extracted. The length ratios of the long, medium, and short axes are calculated, and flat or linear regions with aspect ratios that deviate significantly from the morphological characteristics of nodular lesions are removed. The module can also selectively perform shape prior constraints based on the convex hull of the seed point. Virtual rays are emitted in all directions of space using the voxel with the highest probability as the seed, and boundary voxels whose probabilities fall to the cutoff threshold are recorded. These boundary voxels are connected to form a convex hull. The original probability is retained inside the convex hull, and the outside is cleared to zero, making the boundary of the constrained probability map tend to be smooth. After the above processing is completed, a constrained probability map signal is generated.

[0026] like Figure 2As shown in the flowchart, this flowchart illustrates the detailed vertical calculation process within the image registration module. It employs a two-layer independent iterative loop structure, distinguishing between rigid registration and B-spline deformation registration optimization processes. Multiple loop backflow logics are present, and a single linear execution method is not used. The starting node is a fixed CT image + a floating MRI image. This node distinguishes between the two types of image attributes, fixing the CT image as the spatial reference image and defining the MRI image as a floating image to be spatially aligned. Simultaneously, it reads all grayscale pixel information from both sets of images and transmits the original image data to the rigid transformation parameter gradient descent optimization node. The rigid transformation parameter gradient descent optimization node uses normalized mutual information as the optimization objective, synchronously iteratively updating the rigid transformation parameters, which include six-dimensional components (three-axis translation and three-axis rotation). After each round of parameter updates, it calculates the mutual information matching value between the two sets of images. The calculation result is transmitted to the rigid mutual information increment threshold judgment node, forming the first layer of the loop. If the increase in mutual information value before and after the current iteration is not lower than the preset convergence threshold, it indicates that the rigid spatial alignment accuracy is insufficient. The data flows back to the gradient descent optimization node for rigid transformation parameters to continue updating the transformation parameters. If the increase in mutual information meets the convergence criterion, the rigid registration stage ends, and the data flows downward to the B-spline deformation mesh construction node. Based on the rigidly registered and aligned image, the B-spline deformation mesh construction node generates a three-dimensional free deformation mesh with uniform control point spacing, establishing a deformation model with finely adjustable local spatial positions. After the mesh is built, the mesh parameters and the rigidly aligned image are fed into the optimized control point displacement vector node. The formula for calculating normalized mutual information is as follows: in, The output value is the normalized mutual information function value, where F is the gray-level distribution variable of the fixed CT image and M is the gray-level distribution variable of the deformed floating MRI image. For rigid transformation of three-dimensional translation and rotation parameter vectors, Let be the displacement vector of the control points of the B-spline deformed mesh. For indexing the grayscale range of CT images, For traversing the voxel space of a 3D image. The gray-voxel probability distribution function of CT images and deformed MRI images. Let be the probability distribution function of grayscale edges in CT single-modality mode. This represents the marginal probability distribution function of a single-modal voxel in deformation MRI. In the registration process, the module first iteratively updates the rigid transformation vector T using gradient descent. In each iteration, the normalized mutual information formula is substituted into the matrix to calculate the matching degree until the rigid transformation converges. Then, an equally spaced B-spline deformation mesh is constructed, and the control point displacement vector δ is successively fine-tuned while repeatedly solving this formula. The incremental change of the objective function is monitored in real time. When the increment falls below a preset convergence threshold, the deformation iteration terminates. Finally, the optimal T and... The formula completes MRI image resampling and outputs multimodal registration signals in a unified spatial coordinate system. It solves the problem of large differences in grayscale values ​​between CT and MRI images and the failure of direct pixel matching. It achieves accurate cross-modal alignment by relying on information entropy measurement, providing spatially aligned and offset-free input data for subsequent multimodal feature segmentation modules. Without this normalized mutual information calculation logic, the misalignment of lesion positions between CT and MRI will directly lead to large-area positioning deviations in the subsequent tumor probability map generation, significantly reducing the overall accuracy of target area delineation. The formula ensures the robustness of cross-modal similarity measurement calculation and is adaptable to radiological image input scenarios with different locations and different scanning slice thicknesses.

[0027] The optimized control point displacement vector node continuously adjusts the 3D displacement vectors of all control points within the grid. Each displacement adjustment recalculates the normalized mutual information value between the floating MRI image and the baseline CT image. The calculation result is transmitted to the deformation mutual information increment criterion judgment node to form a second-layer loop. If the mutual information increment after deformation iteration does not reach the preset convergence threshold, the data flows back to the optimized control point displacement vector node for continuous fine-tuning of the control point displacement. If the increment meets the convergence criterion, the deformation registration stage is completed, and the data flows into the multimodal image resampling and alignment node. The multimodal image resampling and alignment node performs 3D resampling on the deformed MRI image based on the optimal rigidity transformation parameters and the optimal B-spline deformation control point displacement, ensuring complete uniformity of voxel size and spatial coordinates between the CT and MRI images, eliminating intermodal spatial offset errors. The finally processed data is transmitted to the output multimodal registration signal node, outputting a standardized multimodal registration signal that is spatially perfectly aligned and can be directly used for subsequent feature segmentation calculations. The entire subdivision process completes global rigid coarse alignment and local deformation fine alignment through a two-layer loop. It relies on the change of normalized mutual information value to control the start and stop of the loop, fully restores all iterative optimization operation logic inside the image registration module, and executes each node layer by layer in vertical order. The loop backflow does not disrupt the overall vertical arrangement structure, and it is a complete technical implementation process that adapts to the cross-modal spatial matching of multimodal radiographic images.

[0028] The Conditional Random Field (CRF) optimization module receives constraint probability map signals, gradient edge signals from computed tomography (CT) images, and phase consistency edge signals from magnetic resonance imaging (MRI) images to construct a fully connected CRF for fine-grained boundary evolution. The energy function of this CRF consists of the sum of a univariate potential term and a binary potential term. The univariate potential term is directly taken from the negative logarithm of the probability values ​​of each voxel in the constraint probability map signal, representing the cost of independently labeling the voxel as either tumor or background; this signal already encodes gross anatomical constraints and morphological screening information. The binary potential term is composed of the weighted sum of two Gaussian kernel functions, establishing the correlation between voxels. The first Gaussian kernel function considers both spatial location differences and edge feature differences. The spatial location difference is the square of the Euclidean distance between the two voxels divided by the first spatial bandwidth parameter; the edge feature difference is the square of the difference in the fused edge amplitude of the two voxels divided by the first gradient bandwidth parameter. The fusion edge magnitude is obtained by weighted summation of the computed tomography gradient magnitude and the magnetic resonance phase-consistency edge response. The weighting coefficients are dynamically determined by the local signal-to-noise ratio ratio of the two modes at each voxel; a higher signal-to-noise ratio assigns a greater weight. The phase-consistency edge response is obtained through multi-scale, multi-directional Fourier domain analysis of the magnetic resonance image. This response is independent of the absolute grayscale value variation of the image and can capture tissue boundaries in the magnetic resonance image that are visible to the naked eye but difficult to detect by gradient operators. The second Gaussian kernel function depends only on spatial location differences and is calculated by dividing the square of the Euclidean distance between the two voxels by the second spatial bandwidth parameter. It is used to penalize isolated noise points that are spatially adjacent but have inconsistent labels. The superposition of the two kernel functions forms a binary potential term, which encourages adjacent voxel labels to be consistent at the boundary, while allowing voxels that cross significant edges to maintain label differences. The mean field approximation algorithm is used to iteratively update the probability values ​​of each voxel belonging to the tumor and the background. In each iteration, the expected contribution of all other voxels to the binary potential of the current voxel is calculated and superimposed on the unary potential. After normalization, a new probability distribution is obtained. This process continues until the energy change of two adjacent iterations is lower than the termination condition. Finally, a binarized tumor target mask signal is output.

[0029] The quantitative analysis module receives the tumor target area mask signal, performs connected component labeling on the mask to distinguish independent tumor lesions, and extracts clinically commonly used quantitative indicators for each lesion. Tumor volume is calculated by multiplying the total number of voxels contained in the lesion by the actual volume of a single voxel in physical space. The maximum 3D diameter is calculated by traversing all voxel coordinate pairs within the lesion, calculating the Euclidean distance for each pair, and taking the maximum value among all distances. This value reflects the maximum extension range of the lesion in any direction in 3D space. Texture heterogeneity parameters are calculated based on the gray-level co-occurrence matrix. A gray-level co-occurrence matrix is ​​constructed within the lesion region according to a preset voxel pair offset direction and distance. The frequency of occurrence of each pair of gray values ​​is counted, and energy is calculated from this matrix to reflect the uniformity of the texture. Contrast is calculated to reflect the severity of local gray-level changes, correlation is calculated to reflect the linear dependence of gray values ​​along a specific direction, and homogeneity is calculated to reflect the smoothness of the gray-level distribution. The above volume, diameter, and texture parameters are summarized into a structured analysis result signal and output to the radiotherapy planning system or clinical record database.

[0030] like Figure 3 As shown, this flowchart represents a vertical sub-process combining multimodal feature segmentation, anatomical constraints, and conditional random fields for refined processing. It incorporates network operation mode branch links and conditional random field self-iterative loop links. The process structure exhibits branching and local backflow, and is not a purely linear serial process. The starting node is the multimodal registration signal input. This node receives fused CT and MRI image data with unified spatial coordinates, processed by the image registration module, and transmits the standardized 3D image signal downwards to the network operation mode branch node. This node divides into two independent parallel computation paths, corresponding to the model training scenario and the clinical inference scenario, respectively. The first path is the inference mode, where data flows directly into the cross-modal attention feature fusion inference node. This node relies on independent CT encoders and MRI encoders to extract dual-modal image features layer by layer. Cross-modal channel attention fusion units are deployed at each level of the encoder. Through global average pooling and weight generation network, adaptive weights for each channel are calculated. After weighted fusion of dual-modal features, the channel dimensions are unified by 1×1 convolution, and the fused features are output and sent to the shared decoder to generate the initial tumor probability map signal. The second path is the training mode, where data flows into the joint loss backpropagation training node. This node relies on a composite joint loss function including Dessell loss, focus loss, weighted cross-entropy loss, and regularization loss to perform gradient backpropagation updates on all trainable parameters of the multi-feature fusion network, continuously optimizing the encoder, decoder, and attention weight network parameters. After parameter iteration, the initial tumor probability map signal is also output. The data from both branch paths eventually converge to the node that generates the initial tumor probability map.

[0031] The node that generates the initial tumor probability map outputs a coarse segmentation probability map that has not been filtered by anatomical structures. The data is then transmitted down to the anatomical probability constraint processing node. This node relies on the multi-organ segmentation network to complete the semantic segmentation of human organs. It performs a full set of processing, including zeroing the organ mask probability, attenuating the distance transformation probability, filtering the volume and shape of connected components, and regularizing the edges of the three-dimensional convex hull. It removes false positive voxels that are anatomically impossible to have tumors and outputs the constraint probability map signal, which is then sent to the fully connected CRF energy iteration optimization node. The fully connected CRF energy iterative optimization node integrates the constraint probability map with CT-MRI fusion edge signals to construct a global energy function containing a univariate potential term and a dual-Gaussian kernel binary potential term. A mean-field approximation algorithm is used to iteratively update the probability distribution of all tumor and background labels. After each iteration, the global energy value is recalculated. If the energy does not reach a convergent and stable state, the data is directly fed back to the node for further iterative optimization, continuously smoothing the tumor target boundary and eliminating isolated small false detection areas. When the global energy value converges, the node outputs a tumor target mask signal with continuous contours and no artifact interference, which is finally transmitted to the output tumor target mask node, completing all calculations of the entire segmentation constraint refinement sub-process. The entire sub-process relies on branches to distinguish between model training and clinical inference scenarios, and uses self-looping to continuously optimize the target contour. It connects three core algorithm modules: cross-modal feature fusion, anatomical and physiological constraints, and global probability field refinement correction, completely covering all intermediate calculation steps from tumor probability map generation to final standardized mask output.

[0032] During the training phase, the multi-feature fusion network of the multimodal feature segmentation module performs end-to-end parameter optimization through a joint loss function. This joint loss function consists of a weighted sum of three loss terms. The first loss term calculates the sum of the Descein loss and the focal loss between the initial tumor probability map signal and the manually labeled tumor mask. The Descein loss directly optimizes the overlap between the segmented region and the labeled region, while the focal loss imposes a higher penalty on boundary voxels that are difficult to classify. The second loss term calculates the weighted cross-entropy loss between the constrained probability map signal and the manually labeled tumor mask, assigning higher weights to positive sample voxels belonging to the tumor and lower weights to negative sample voxels belonging to the background, thereby guiding the network to focus more on the tumor region itself rather than the large background in the output phase. The third loss term is the L2 regularization loss of all trainable parameters, suppressing parameter overfitting. The first, second, and third coefficients control the contribution ratio of the three losses, respectively, and the sum serves as the objective function for backpropagation, updating the parameters of each layer through the gradient descent algorithm until convergence. After training, the network parameters are fixed, and the multimodal feature segmentation module only performs forward inference and no longer calculates loss terms.

[0033] The present invention provides an automatic tumor target delineation and analysis system based on multimodal radiographic image fusion. This system forms aligned image pairs by performing rigid and deformation registration on computed tomography (CT) and magnetic resonance imaging (MRI) images. An initial tumor probability map is generated by a multi-feature segmentation network containing independent encoders and cross-modal channel attention fusion units. Spatial filtering is performed using a pre-built anatomical probability map. The boundary is then iteratively optimized using a fully connected conditional random field combined with multimodal edge signals. Finally, a target mask is output and its volume, three-dimensional diameter, and texture parameters are calculated, achieving fully automatic delineation.

[0034] Therefore, the automatic tumor target delineation and analysis system based on multimodal radiographic image fusion of the present invention can solve the problems of insufficient information from single-modal images and the high subjectivity and low efficiency of manual delineation.

[0035] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. An automatic tumor target delineation and analysis system based on multimodal radiographic image fusion, characterized in that, include: The image registration module performs rigid registration and B-spline-based deformation registration on CT and MRI images to generate multimodal registration signals. The multimodal feature segmentation module receives the multimodal registration signal and generates an initial tumor probability map signal through a multi-feature fusion network containing an independent CT encoder, an MRI encoder, and a shared decoder. The encoder has a cross-modal channel attention fusion unit between each level, which transmits the weighted fused feature map to the decoder. The anatomical probability constraint module receives the initial tumor probability map signal, uses a pre-built anatomical probability map atlas to spatially filter the probability map, removes voxels that do not conform to the anatomical constraints, and forms a constraint probability map signal. The Conditional Random Field (CRF) optimization module receives the constraint probability map signal and the gradient edge signal from the CT image. It constructs a fully connected CRF with the constraint probability map as a univariate potential and the gradient edge as a binary potential. After iterative evolution, it forms a tumor target mask signal. The quantitative analysis module receives the tumor target area mask signal, calculates and outputs the analysis result signal, which includes volume, maximum three-dimensional diameter, and texture parameters based on the gray-level co-occurrence matrix.

2. The automatic tumor target delineation and analysis system based on multimodal radiographic image fusion according to claim 1, characterized in that, The image registration module uses CT images as fixed images and MRI images as floating images. First, it calculates the normalized mutual information similarity measure between the two images. Then, it uses gradient descent to search for rigid transformation parameters that maximize the mutual information value. These rigid transformation parameters include three translational components and three rotational components. After rigid registration, a B-spline free-deformation mesh with consistent control point spacing is constructed based on the rigid registration results. Using normalized mutual information as the objective function, the floating image is deformed step-by-step by optimizing the displacement vector of each control point. Iteration stops when the increment of the objective function falls below a preset convergence threshold. Finally, the deformed MRI and CT images are resampled to the same spatial coordinate system to form the multimodal registration signal. The formula for calculating the normalized mutual information is as follows: in, The output value is the normalized mutual information function value, where F is the gray-level distribution variable of the fixed CT image and M is the gray-level distribution variable of the deformed floating MRI image. For rigid transformation of three-dimensional translation and rotation parameter vectors, Let be the displacement vector of the control points of the B-spline deformed mesh. For indexing the grayscale range of CT images, For traversing the voxel space of a 3D image. The gray-voxel probability distribution function of CT images and deformed MRI images. Let be the probability distribution function of grayscale edges in CT single-modality mode. This represents the marginal probability distribution function of a single-modal voxel in deformation MRI.

3. The automatic tumor target delineation and analysis system based on multimodal radiographic image fusion according to claim 1, characterized in that, The cross-modal channel attention fusion unit performs global average pooling on the feature maps output from the same level of the CT encoder and MRI encoder, compressing the feature map of each channel into a scalar value to obtain a first channel descriptor vector and a second channel descriptor vector. The first and second channel descriptor vectors share the same weight generation network, which is composed of a fully connected layer, a linear rectified activation function, a second fully connected layer, and a sigmoid activation function connected in sequence. This weight generation network outputs CT channel weight vectors and MRI channel weight vectors for the first and second channel descriptor vectors, respectively. The CT channel weight vector is multiplied along the channel dimension by the feature map output from the CT encoder, and the MRI channel weight vector is multiplied along the channel dimension by the feature map output from the MRI encoder, to obtain a weighted first feature map and a weighted second feature map. The weighted first feature map and the weighted second feature map are then added element-wise, and then the number of channels is restored to the number of channels before addition by a convolutional layer with a kernel size of 1x1, forming the fused feature map of this level passed to the corresponding level of the shared decoder. The calculation formula for the CT channel weight vector is as follows: Where w is the channel weight vector identifier, CT is the CT encoder-specific distinguishing identifier, and c is the index of the currently calculated feature channel. This is the identifier for the Sigmoid activation function. This is the weight matrix for the second fully connected layer. This is the identifier for the linear rectification activation function. Here, represents the weight matrix of the first fully connected layer, H represents the number of pixels in the height of a single-channel feature map, W represents the number of pixels in the width of a single-channel feature map, i represents the traversal index in the height direction of the feature map, and j represents the traversal index in the width direction of the feature map. Here is the original feature map of the c-th channel of the CT encoder, where (i,j) is the two-dimensional pixel coordinate index of the feature map. This is the bias vector for the first fully connected layer. This is the bias vector for the second fully connected layer.

4. The automatic tumor target delineation and analysis system based on multimodal radiographic image fusion according to claim 1, characterized in that, The anatomical probability constraint module internally deploys a pre-trained multi-organ segmentation network. This network receives the multimodal registration signal and performs organ region segmentation on CT and MRI images respectively, outputting semantic segmentation results containing multiple organ labels. The anatomical probability constraint module extracts organ mask regions that allow tumor occurrence from the semantic segmentation results based on a preset list of organs associated with lesion occurrence, and simultaneously extracts organ mask regions that prohibit tumor occurrence from the semantic segmentation results. The probability values ​​of voxels falling into the organ mask regions that prohibit tumor occurrence in the initial tumor probability map signal are set to zero, and the probability values ​​of voxels falling outside the organ mask regions that allow tumor occurrence are also set to zero. For the retained non-zero probability voxels, the distance transformation value between them and the allowable organ mask regions is further calculated. When the distance transformation value is lower than a preset distance threshold, the probability value of the voxel remains unchanged. When the distance transformation value exceeds the preset distance threshold, the probability value of the voxel is proportionally attenuated. After completing the above operations, the constraint probability map signal is formed.

5. The automatic tumor target delineation and analysis system based on multimodal radiographic image fusion according to claim 4, characterized in that, Before generating the constraint probability map signal, the anatomical probability constraint module performs connected component analysis. It performs threshold binarization on the probability map after probability zeroing and attenuation processing, extracts all spatially connected candidate regions, calculates the volume of each connected region, and deletes connected regions with volumes smaller than a preset minimum tumor volume. For connected regions with volumes larger than a preset maximum tumor volume, it shrinks inward along the contour normal direction to the number of voxels corresponding to the preset maximum tumor volume while preserving its outer contour. It then performs 3D shape analysis based on principal component analysis on the preserved connected regions, extracting the major, middle, and minor axes and their corresponding lengths. It calculates the ratio of the major axis to the middle axis and the ratio of the middle axis to the minor axis. When these ratios exceed a preset aspect ratio threshold, the connected region is determined to be a non-nodular morphology and is removed. The constraint probability map signal is formed through the above connected component analysis.

6. The automatic tumor target delineation and analysis system based on multimodal radiographic image fusion according to claim 4, characterized in that, The anatomical probability constraint module also performs edge normalization on the constraint probability map signal by constructing a shape prior. The shape prior is constructed by taking the voxel with the highest probability value in the initial tumor probability map signal as the seed point, emitting virtual rays from the seed point along the direction vector of uniform sampling in three-dimensional space, recording the voxel coordinates on each virtual ray where the probability value first falls below a preset cutoff threshold, connecting all the recorded voxel coordinates to form a three-dimensional convex hull, setting the voxel probability value inside the three-dimensional convex hull as the original probability value of that voxel in the initial tumor probability map signal, and setting the voxel probability value outside the three-dimensional convex hull as zero, so that the boundaries of each connected region in the constraint probability map signal tend to be smooth.

7. The automatic tumor target delineation and analysis system based on multimodal radiographic image fusion according to claim 1, characterized in that, The energy function of the fully connected conditional random field constructed by the conditional random field optimization module consists of a univariate potential term and a binary potential term. The univariate potential term is directly taken from the negative logarithm of the probability value of each voxel in the constraint probability map signal, representing the independent cost of labeling the voxel as tumor or background. The binary potential term is composed of a weighted sum of two Gaussian kernel functions. The first Gaussian kernel function takes the spatial position difference of the voxel and the gradient magnitude difference of the corresponding voxel in the gradient edge signal from the CT image as input. The spatial position difference is calculated by dividing the square of the Euclidean distance between the two voxels by the square of the first spatial bandwidth parameter, and the gradient magnitude... The difference is calculated by dividing the square of the difference between the gradient magnitudes of two voxels by the square of the first gradient bandwidth parameter. The second Gaussian kernel function takes only the spatial position difference of the voxels as input and is calculated by dividing the square of the Euclidean distance between the two voxels by the square of the second spatial bandwidth parameter. The first Gaussian kernel function tends to assign the same label to adjacent voxels with similar gradient magnitudes. The second Gaussian kernel function is used to eliminate isolated small false detection regions. The univariate potential term and the binary potential term are summed to form the global energy function. The label probability distribution of each voxel is iteratively updated using the mean field approximation algorithm until the energy function converges, and the tumor target area mask signal is output.

8. The automatic tumor target delineation and analysis system based on multimodal radiographic image fusion according to claim 7, characterized in that, The gradient edge signal derived from CT images is fused with the phase-consistency edge signal derived from MRI images before being input into the binary potential term. The phase-consistency edge signal of the MRI images is obtained by calculating the Fourier domain phase consistency value of the MRI images at multiple scales and in multiple directions. For each voxel, the maximum value of the phase consistency value in each direction is taken as the phase-consistency edge response of that voxel. The local signal-to-noise ratio of the CT images and the local signal-to-noise ratio of the MRI images at each voxel are calculated. The ratio of the CT local signal-to-noise ratio to the MRI local signal-to-noise ratio is used as a weighting coefficient. The CT gradient amplitude and the MRI phase-consistency edge response are weighted and summed to generate the fused edge signal. The fused edge signal replaces the gradient edge signal derived from the CT images as the gradient amplitude difference input of the first Gaussian kernel function in the binary potential term.

9. The automatic tumor target delineation and analysis system based on multimodal radiographic image fusion according to claim 1, characterized in that, The quantitative analysis module performs connected component labeling on the tumor target area mask signal, identifies each independent tumor region, counts the total number of voxels contained in each tumor region and multiplies it by the physical volume of a single voxel to obtain the tumor volume, traverses all voxel pairs within the tumor region to calculate the Euclidean distance between each voxel pair, takes the maximum value among all Euclidean distances as the maximum three-dimensional diameter, constructs a gray-level co-occurrence matrix within the tumor region, calculates four texture feature values ​​of the gray-level co-occurrence matrix: energy, contrast, correlation, and homogeneity, and organizes the tumor volume, maximum three-dimensional diameter, energy, contrast, correlation, and homogeneity into a structured analysis result signal and outputs it.

10. The automatic tumor target delineation and analysis system based on multimodal radiographic image fusion according to claim 1, characterized in that, The multi-feature fusion network in the multimodal feature segmentation module uses a joint loss function for parameter optimization during the training phase. The joint loss function consists of a weighted sum of a first loss term, a second loss term, and a third loss term. The first loss term is the sum of the Dessian loss and the focal loss between the initial tumor probability map signal and the manually labeled tumor mask. The second loss term is the weighted cross-entropy loss between the constraint probability map signal and the manually labeled tumor mask. In the manually labeled tumor mask, voxels with a voxel value of one are assigned a first weight, and voxels with a voxel value of zero are assigned a second weight, with the first weight being greater than the second weight. The third loss term is the regularization loss obtained by taking the sum of the squares of the L2 norm of all trainable parameters of the network. The first loss term, the second loss term, and the third loss term are multiplied by preset first coefficients, second coefficients, and third coefficients, respectively, and then summed to serve as the objective function for backpropagation.

Images