A method for evaluating the prognosis of gastric low-grade intraepithelial neoplasia based on deep learning imageomics features

Abstract

The application provides a gastric low-grade intraepithelial neoplasia outcome evaluation method based on deep learning imageomics features, and relates to the field of medical information technology. The method comprises the following steps: acquiring a gastric medical image, extracting a lesion region of interest and expanding outward to construct an extended background region containing a surrounding mucosa microenvironment; extracting a high-dimensional imageomics feature set from the lesion region of interest; inputting the above two regions into a multi-branch convolutional neural network to extract deep semantic feature maps, and calculating the spatial feature difference to obtain an edge weight prior matrix; then constructing a spatial heterogeneity graph, and obtaining deep imageomics fusion features through collaborative updating of a graph attention network; finally, inputting the deep imageomics fusion features into an outcome prediction classifier to obtain a final outcome evaluation result. The application realizes deep fusion modeling of macro-microenvironment differences and micro-tissue heterogeneity, and significantly improves the accuracy of outcome prediction.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features. Background Technology

[0002] Low-grade gastric intraepithelial neoplasia (LSIN) is a precancerous stage in the progression of gastric mucosal cells towards gastric cancer. Clinical follow-up shows that the progression of this lesion exhibits high individual variability, with outcomes including lesion reversal and regression, maintenance of the current state, or progression to high-grade LSIN and early gastric cancer. Currently, clinical assessment of the prognosis of LSIN relies heavily on endoscopic biopsy and pathological examination. However, biopsy is an invasive procedure, subject to sampling errors, and pathological diagnoses are easily influenced by physician subjective experience, making it difficult to perform a comprehensive quantitative assessment of the prognosis risk of the entire lesion area non-invasively.

[0003] In recent years, advancements in medical image analysis technology have propelled the field towards objectivity and quantification. Radiomics, by extracting quantitative image features such as morphology, first-order statistics, and higher-order textures from routine medical images in high throughput, provides a technical means for non-invasively revealing the microscopic heterogeneity of precancerous lesions. Simultaneously, deep convolutional neural networks, with their adaptive feature extraction capabilities, are widely used to extract macroscopic semantic features from medical images. Integrating radiomics features with deep learning features into predictive models to assist physicians in developing clinical intervention plans for endoscopic resection or long-term follow-up is currently a mainstream research trend in the field of auxiliary diagnostics.

[0004] While the aforementioned combined methods have made some progress in lesion assessment, existing technologies still have significant limitations. On the one hand, existing multi-feature fusion schemes typically employ a method of directly concatenating omics feature arrays with deep semantic feature arrays. This process severs the spatial topological relationship between the lesion's microscopic texture and macroscopic imaging representation, leading to feature dimensional redundancy and limiting the model's generalization ability. On the other hand, the malignant evolution of low-grade gastric intraepithelial neoplasia is directly influenced by the mucosal microenvironment surrounding the lesion. However, existing methods often strictly limit the region of interest for feature extraction to within the lesion boundary, failing to incorporate the morphological evolution features of the lesion edge and surrounding mucosa into the assessment system. This results in insufficient quantification of the spatial evolution trend of the lesion by the model. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features. This invention solves the problem in existing technologies that fail to effectively integrate the spatial topological features of the lesion area and the surrounding microenvironment, resulting in low accuracy in predicting the prognosis of low-grade intraepithelial neoplasia.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features includes: Acquire gastric medical images of the patient to be evaluated, and extract the region of interest for the lesion from the gastric medical images; Based on the boundary of the region of interest of the lesion, an extended background region including the surrounding mucosal microenvironment is constructed by expanding outward. The region of interest of the lesion is transformed to extract a set of radiomics features that reflect the heterogeneity of tissue microtexture; The region of interest of the lesion and the extended background region are respectively input into a pre-trained multi-branch convolutional neural network, which outputs a first deep semantic feature map and a second deep semantic feature map respectively. Calculate the spatial feature difference between the first deep semantic feature map and the second deep semantic feature map to obtain the spatial difference matrix; The correlation between feature elements is calculated based on the spatial difference matrix to obtain the edge weight prior matrix; A spatial heterogeneity map is constructed, and each feature in the image omics feature set is mapped to a graph node. The edge weight prior matrix is ​​used as the initial connection weight, and the attention coordination coefficient between the graph nodes is calculated and the state is updated through a graph attention network to obtain deep image omics fusion features. The deep radiomics fusion features are input into the outcome prediction classifier for probability mapping to obtain the progression risk probability distribution of low-grade gastric intraepithelial neoplasia, and the final outcome assessment result is obtained based on the progression risk probability distribution.

[0007] The present invention discloses the following technical effects: This invention provides a method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features. The invention accurately quantifies the macroscopic evolutionary correlation by calculating the deep semantic space difference matrix between the lesion and its surrounding microenvironment, and uses this as prior information for the edge weights of a graph attention network to guide the collaborative updating of microscopic radiomics node features. This mechanism effectively overcomes the defects of topological association fragmentation and dimensional redundancy caused by simple feature splicing in traditional methods. It achieves deep fusion modeling of macroscopic microenvironmental differences and microscopic tissue heterogeneity, accurately characterizing the evolutionary trend of the lesion, thereby significantly improving the accuracy and robustness of assessing the prognosis of low-grade gastric intraepithelial neoplasia. This provides a reliable quantitative basis for developing non-invasive, personalized intervention strategies in clinical practice. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in 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.

[0009] Figure 1 The flowchart illustrates a method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features, as provided in this embodiment of the invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0012] like Figure 1 As shown, this invention provides a method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features, including: Step 100: Acquire gastric medical images of the patient to be evaluated, and extract the region of interest for the lesion from the gastric medical images; Step 200: Expand outward from the boundary of the region of interest of the lesion to construct an extended background region that includes the surrounding mucosal microenvironment; Step 300: Transform the region of interest of the lesion to extract a set of radiomics features that reflect the heterogeneity of tissue microtexture; Step 400: Input the region of interest of the lesion and the extended background region into a pre-trained multi-branch convolutional neural network, respectively, and output a first deep semantic feature map and a second deep semantic feature map accordingly; Step 500: Calculate the spatial feature difference between the first deep semantic feature map and the second deep semantic feature map to obtain the spatial difference matrix; Step 600: Calculate the correlation between feature elements based on the spatial difference matrix to obtain the edge weight prior matrix; Step 700: Construct a spatial heterogeneity map, map each feature in the image omics feature set to a graph node, use the edge weight prior matrix as the initial connection weight, and calculate the attention synergy coefficient between the graph nodes through a graph attention network and perform state updates to obtain deep image omics fusion features. Step 800: Input the deep radiomics fusion features into the outcome prediction classifier for probability mapping to obtain the progression risk probability distribution of low-grade gastric intraepithelial neoplasia, and obtain the final outcome assessment result based on the progression risk probability distribution.

[0013] Furthermore, the specific implementation process of step 100 is as follows: Acquire initial gastric medical images of the patient to be evaluated. In clinical practice, the raw digital images acquired by relevant medical equipment (i.e., the initial gastric medical images described in the claims) are often affected by factors such as hardware noise, uneven lighting, or mucus reflection. Therefore, the initial gastric medical images must undergo image denoising and pixel normalization processing. Specifically, a Gaussian filtering algorithm can be used to eliminate high-frequency noise, and a linear mapping formula can be used to uniformly scale the pixel grayscale value range of the image to a standard value range, thereby eliminating baseline drift caused by different imaging devices and obtaining gastric medical images of uniform quality that conform to the neural network input specifications.

[0014] The gastric medical image is input into a pre-trained lesion segmentation network to obtain an initial lesion mask. The lesion segmentation network is a fully convolutional neural network whose parameters have been iteratively optimized using a large sample set with contours annotated by authoritative doctors. This network can adaptively extract multi-scale semantic features from the image, automatically identify abnormal gastric mucosal lesion areas, and output the initial lesion mask composed of binary pixels (e.g., setting the value of pixels predicted as lesions to 1 and background pixels to 0).

[0015] Considering that the segmentation results directly output by the network may contain edge spikes, isolated noise points, or internal micro-holes, this embodiment further performs morphological boundary smoothing on the initial lesion mask. In specific implementation, by using structuring elements of a preset size, opening and closing operations in mathematical morphology are sequentially performed on the initial lesion mask. The opening operation is used to eliminate small mis-segmented patches that are detached from the main body, and the closing operation is used to fill the gaps inside the mask, thereby smoothing the rough boundaries and finally obtaining a target lesion mask with good connectivity and continuous contours.

[0016] Based on the target lesion mask, the gastric medical image is localized and cropped. Specifically, the non-zero pixel coordinate boundaries of the target lesion mask are extracted, and its minimum bounding rectangle or true contour mapping in the global coordinate system is calculated. This mapping is then used to crop the image at the corresponding spatial location on the gastric medical image, removing a large amount of redundant normal gastric tissue background to obtain the region of interest for the lesion. This step not only significantly reduces the amount of data processed by subsequent algorithms but also provides a precise spatial boundary basis for subsequent extraction of tissue microtexture heterogeneity features.

[0017] Furthermore, the specific implementation process of step 200 is as follows: This embodiment extracts the boundary contour coordinates of the region of interest (ROI) of the lesion and generates a corresponding binary mask of the lesion in the global coordinate system of the gastric medical image. Specifically, this embodiment traverses the closed contour of the ROI of the lesion, assigning a value of 1 to all pixels inside the contour to represent the main lesion region with pathological abnormalities, and assigning a value of 0 to all pixels outside the contour to represent the background region without lesions. Through the above assignment rules, this embodiment constructs a binary mask of the lesion containing only morphological and topological boundary information in a global two-dimensional matrix, for example, with a size of 512 by 512 pixels, thereby providing an accurate mathematical mapping benchmark for subsequent spatial scale expansion operations.

[0018] This embodiment obtains a preset physical reference distance for the microenvironment and, in conjunction with the pixel spatial resolution of the gastric medical image, calculates the size of the structural element for morphological operations. The physical reference distance for the microenvironment refers to the actual physical extent of the microenvironment surrounding the lesion, which may contain an inflammatory-cancer transformation microenvironment, determined based on clinical pathology statistical priors. This embodiment extracts the pixel spatial resolution in the horizontal and vertical directions of the image by parsing the underlying metadata of the medical image, and calculates the equivalent pixel spatial resolution based on these two values. Subsequently, this embodiment divides the physical reference distance for the microenvironment by the equivalent pixel spatial resolution and rounds the calculated quotient up, thereby determining the number of pixels corresponding to the radius of the isotropic structural element used for morphological expansion. For example, when the clinically set physical reference distance for the microenvironment is 10 mm and the equivalent pixel spatial resolution of the image is 0.5 mm per pixel, after division and rounding up, this embodiment determines the radius of the structural element to be precisely 20 pixels.

[0019] This embodiment constructs an isotropic structural element based on the previously calculated structural element size, and performs a morphological dilation operation on the lesion binarization mask using the isotropic structural element to obtain an expanded binarization mask. The isotropic structural element is a two-dimensional digital detection matrix with centrosymmetry; for example, in this embodiment, a solid circular disk matrix with a radius of 20 pixels is constructed in memory. During the morphological dilation operation, this embodiment performs a pixel-by-pixel global sliding scan of the center of the solid circular matrix on the lesion binarization mask; when at least one pixel with a value of 1 exists within the local receptive field covered by the matrix, the pixel value at the output position corresponding to the center point of the matrix is ​​set to 1. Through this morphological operation, the original topological boundary of the lesion will expand outward uniformly by a specified pixel distance, thereby generating an expanded binarization mask with a larger overall area that completely covers the surrounding microenvironment.

[0020] This embodiment calculates the pixel logical difference between the extended binary mask and the lesion binary mask to obtain a peripheral mucosal ring mask. Since the simple extended binary mask still contains the original lesion region, this embodiment needs to perform region removal logic calculations to extract the pure peripheral mucosal microenvironment. Specifically, this embodiment first performs an element-wise inversion operation on the original lesion binary mask, that is, inverting the values ​​of the original lesion region from 1 to 0, and the background from 0 to 1. Next, this embodiment performs an element-wise multiplication operation between the inverted mask and the previously generated extended binary mask, i.e., a binary logical AND operation at spatially corresponding positions. Through this seamless data processing, the pixel values ​​of the originally overlapping lesion center region are all zeroed out, while only the pure ring region with a constant width of 20 pixels generated by the extended expansion is retained, and the pixel value of this region remains 1. This generates a peripheral mucosal ring mask with a clear spatial topology and that does not overlap with the lesion.

[0021] This embodiment utilizes the peripheral mucosal ring mask to perform pixel mapping and cropping of the gastric medical image to obtain the extended background region. Specifically, this embodiment uses the generated peripheral mucosal ring mask as a spatial perspective filter, precisely superimposed onto the original high-resolution gastric medical image with three color channels. For spatial coordinates in the image corresponding to a mask pixel value of 1, this embodiment fully preserves the original mucosal color and tissue texture pixel data; for positions corresponding to a mask pixel value of 0, their pixel values ​​are forcibly set to background zero. Through this point-to-point pixel mapping and cropping filtering, this embodiment successfully and non-destructively extracts the ring-shaped mucosal tissue image closely adjacent to the lesion but not containing the lesion itself from the original medical image, i.e., the extended background region, thus providing pure and high-quality data support for subsequent deep learning networks to accurately quantify and extract the inflammatory cell infiltration features at the lesion edge.

[0022] Specifically, the calculation expression for the size of the structuring element in the morphological operation is as follows: ; in, This is a preset physical reference distance for the microenvironment; The pixel spatial resolution of gastric medical images in the horizontal direction; The pixel spatial resolution of gastric medical images in the vertical direction; For the reason and The calculated equivalent pixel space resolution; is the radius (in pixels) of the isotropic structuring element used for morphological dilation; "••" is the floor operator; The calculation expression for the peripheral mucosal annular mask is as follows: ; in, The binary mask for the lesion (elements are either 0 or 1); For radius isotropic structural elements; As structural elements Morphological dilation operator performed on the input binary mask; This is the expanded binarized mask obtained after dilation; For the peripheral mucosa ring mask; For element-wise multiplication (corresponding to binary logical AND); This is to invert each element of the binary mask for the lesion.

[0023] More specifically, the "microenvironment physical reference distance" is derived from multi-center retrospective clinical pathology statistical data. Medical experts have found that abnormal changes in the submucosal microvessels and inflammatory microenvironment of low-grade gastric intraepithelial neoplasia often extend beyond the lesion boundary visible to the naked eye under endoscopy, typically concentrated within a specific physical distance (e.g., 5 to 15 mm) of the lesion's extension. Therefore, this parameter is a core threshold connecting macroscopic clinical experience with microscopic pixel calculations, and its function is to define the physical range of the microenvironment with the highest predictive value for outcome. The "equivalent pixel spatial resolution" is a low-level calibration parameter introduced to eliminate the hardware sampling rate differences of the endoscopic probe in different scanning axes. It is obtained by calculating the arithmetic mean of the pixel spatial resolution in the horizontal and vertical directions. The function of this parameter is to ensure that in anisotropically sampled medical images, the generated morphologically dilated structural elements can be mapped to a true physical circle, thereby ensuring that the outward expansion distance of the extended background region remains absolutely consistent in all physical directions, completely avoiding the technical defects of microscopic feature extraction distortion caused by spatial deformation.

[0024] Furthermore, the specific implementation process of step 300 is as follows: To capture the subtle heterogeneous structures hidden in pathological sections at different observation scales, this embodiment utilizes wavelet transform filters and Laplace Gaussian filters to perform multi-scale filtering on the region of interest of the lesion, obtaining multi-scale derived images. Specifically, this embodiment performs smoothing and edge detection convolution operations on the pixel matrix of the lesion region using Laplace Gaussian filters with different standard deviations to extract tissue textures and nodule edge responses of varying coarseness. Simultaneously, this embodiment employs a discrete wavelet transform algorithm to decompose the image spatial frequency into high and low frequencies along the horizontal and vertical directions, generating multiple frequency sub-band feature maps reflecting the local details and global contours of the lesion. Through the above-mentioned joint filtering processing in the frequency and spatial domains at multiple scales, early microscopic changes in the gastric mucosa that are difficult to discern with the naked eye under endoscopy are amplified and explicitly transformed into multi-dimensional digital matrix entities in memory, thus laying a rich data foundation for subsequent extraction of deep quantitative indicators.

[0025] Furthermore, to comprehensively quantify the pathological features of low-grade gastric intraepithelial neoplasia, this embodiment extracts first-order statistical features, two-dimensional shape features, and higher-order texture features from the region of interest of the lesion and the multi-scale derived images to obtain an initial radiomics feature pool. Specifically, this embodiment calculates statistical measures such as skewness and kurtosis based on the original pixel intensity distribution to describe the non-uniformity of lesion density, and defines the complex physical boundary morphology of the lesion region by extracting geometric measures such as the boundary perimeter and compactness of the contour. Based on this, this embodiment further constructs higher-order spatial distribution models such as the gray-level co-occurrence matrix and the gray-level run-length matrix, and deeply mines the internal texture disorder caused by abnormal cell proliferation in the tumor microenvironment by statistically analyzing the joint probability distribution of adjacent pixel pairs under specific directions and distances. The above-mentioned multi-dimensional feature extraction process accurately resolves the abstract medical image patch into a numerical feature vector containing thousands of independent dimensions, physically and concretely characterizing the highly complex spatial heterogeneous microenvironment within the lesion.

[0026] Finally, addressing the strong collinearity and information overlap defects in the high-dimensional feature space, this embodiment uses principal component analysis to perform feature dimensionality reduction and redundancy removal on the initial radiomics feature pool, obtaining the radiomics feature set. In the actual underlying data processing flow, this embodiment first performs zero-mean standardization on all feature vectors in the initial radiomics feature pool, then calculates the covariance matrix between each feature dimension and performs eigenvalue decomposition. This embodiment sorts the eigenvalues ​​in descending order and extracts the top-ranking orthogonal principal component vectors based on a preset cumulative variance contribution rate threshold to construct a projection transformation matrix. Using this projection transformation matrix, this embodiment maps the original large and sparse initial radiomics feature pool to a new low-dimensional orthogonal feature space, thoroughly filtering out invalid noise dimensions and redundant data while preserving the key pathological variation information of the original lesions to the greatest extent. This significantly reduces the computer memory consumption and floating-point operation load during subsequent machine learning model training in engineering implementation, and physically outputs a compact and highly condensed radiomics feature set matrix entity, ensuring the efficiency and robustness of the regression prediction algorithm.

[0027] Specifically, to ensure that the radiomics features are spatially consistent with the subsequent graph structure reasoning process, this embodiment organizes the radiomics feature set into a set of local features corresponding to a spatial grid. Specifically, this embodiment uses the spatial resolution of the deep semantic feature map as an alignment reference, dividing the region of interest of the lesion into multiple local sub-regions according to the corresponding grid. For each local sub-region, first-order statistical features, two-dimensional shape features, and higher-order texture features are extracted from the original image and multi-scale derived images, forming an initial set of radiomics feature vectors corresponding to that local sub-region. Subsequently, this embodiment uniformly performs principal component analysis for dimensionality reduction and redundancy removal on the above local feature sets, ensuring that each local sub-region ultimately outputs a local radiomics feature representation with consistent dimensions. Through this method, the radiomics feature set not only retains the quantitative information of texture heterogeneity within the lesion but also corresponds one-to-one with the nodes in the subsequent construction of the spatial heterogeneity map in terms of spatial index, thereby ensuring that the overall process can be directly implemented.

[0028] Furthermore, the specific implementation process of step 400 is as follows: To thoroughly decouple the visual features of the lesion from its surrounding microenvironment in the high-dimensional feature space of deep learning, this embodiment constructs a multi-branch convolutional neural network comprising a first feature extraction branch and a second feature extraction branch. In practical deployment, this embodiment uses a parameter-independent dual-stream residual network architecture (i.e., the multi-branch convolutional neural network in the claims) as the physical entity for feature extraction. The first and second feature extraction branches each comprise a deep network structure consisting of alternating stacks of two-dimensional convolutional kernels with 3x3 receptive fields and max-pooling layers. By setting mutually independent weight parameter matrices in the two independent branches, this embodiment can construct two parallel forward propagation data stream channels in the memory of the computing device. This mechanism effectively avoids premature information aliasing between abnormal features inside the lesion and normal or inflammatory microenvironment features at the edge in the shallow network, providing a solid structural foundation for the accurate quantification of subsequent spatial differences.

[0029] Furthermore, in this embodiment, the region of interest of the lesion is input into the first feature extraction branch for multi-layer convolution and pooling operations to extract feature representations of the macroscopic pathological appearance inside the lesion, and the first deep semantic feature map is output accordingly. In the specific computation, this embodiment uses a matrix of gastric mucosal lesion images cropped from the original image as source data input to this branch, and performs local feature extraction layer by layer using the bottom-level convolutional kernel array. Through up to 50 layers of nonlinear feature mapping and downsampling operations with a stride of 2, the network adaptively transitions from focusing on low-level visual responses such as edge gradients to extracting highly abstract macroscopic pathological appearances inside the lesion, such as nodular protrusions, local congestion, or erosion on the mucosal surface. Finally, this forward propagation process physically and concretely outputs a high-level tensor matrix entity in the GPU memory, corresponding to the first deep semantic feature map. For example, the spatial size of the output first deep semantic feature map is constrained to 14 by 14 and contains 256 channels, each of which maps the pathological morphological information of a specific dimension of the main body of the low-grade gastric intraepithelial neoplasia lesion with high fidelity.

[0030] Simultaneously, to ensure feature space alignment and accurately capture evolutionary clues of surrounding tissues, this embodiment simultaneously inputs the extended background region into the second feature extraction branch for multi-layer convolution and pooling operations to extract feature representations characterizing the evolutionary state of the microenvironment surrounding the lesion, and outputs the corresponding second deep semantic feature map. During this operation, this embodiment uses data from a pure annular mucosal region stripped of the lesion body as input, and performs feature encoding through computational units with network hierarchical structures equivalent to but with independent parameters to those of the first feature extraction branch. This process focuses on capturing and quantifying key clues characterizing the evolutionary state of the microenvironment surrounding the lesion, such as the degree of inflammatory cell infiltration and abnormal microvascular proliferation patterns. After rigorous alignment and multi-level dimensionality reduction processing, this embodiment physically outputs a feature matrix completely consistent with the aforementioned branch in both spatial and channel dimensions (i.e., also possessing a high-dimensional tensor of 14 x 14 x 256 dimensions), correspondingly outputting the second deep semantic feature map. This dual-stream synchronous independent extraction design precisely produces two sets of digital features with completely equal scales but distinct semantic perceptual domains in the physical memory, thus completely clearing the underlying computational obstacles of matrix dimension mismatch for the next step of cross-spatial dimension feature difference analysis.

[0031] Furthermore, the specific implementation process of steps 500-600 is as follows: To accurately quantify the evolutionary differences in macroscopic morphology between the lesion and its surrounding microenvironment, this embodiment spatially aligns the first and second deep semantic feature maps and performs element-wise subtraction to obtain an initial feature difference matrix. In the underlying computation, this embodiment first verifies the resolution of the two high-dimensional tensors output by the preceding dual-branch network, ensuring they have completely consistent spatial receptive domains (e.g., both are strictly aligned to a 14x14 spatial resolution and 256 feature channels). Subsequently, this embodiment allocates a blank tensor of the same size as the original feature map in the memory pool and uses the processor instruction set to traverse the corresponding spatial coordinates and channel indices, subtracting the pixel scalar values ​​of the second deep semantic feature map representing the surrounding microenvironment from the pixel scalar values ​​of the first deep semantic feature map representing the lesion. Through this point-to-point element-wise subtraction operation, this embodiment physically realizes the morphological evolution gradient of the internal aberrant tissue relative to the external mucosal background, thereby outputting a complete initial feature difference matrix that records the differences in all channels in the device.

[0032] If two deep semantic feature maps have slight differences in spatial dimensions, this embodiment first performs forced alignment to eliminate differential errors caused by pixel-level offsets. Alignment can be achieved through resampling: one feature map is spatially resampled according to the target size to make it completely consistent with the other feature map in terms of length and width; if insufficient or out-of-bounds boundaries occur after resampling, missing boundaries are filled with zeros; if redundant boundaries appear, the redundant parts are cropped so that the two final output feature maps have the same spatial coordinate system and the same effective receptive field coverage. After completing the above alignment, element-wise subtraction can be performed to avoid computational instability caused by dimension mismatch and misaligned comparisons.

[0033] Furthermore, to filter out invalid negative feedback noise generated in the difference calculation and highlight key lesion response information, this embodiment inputs the initial feature difference matrix into a nonlinear activation function for feature response enhancement to obtain the spatial difference matrix. The mechanism of the nonlinear activation function is to perform a mathematical operation logic on the input scalar or tensor elements, taking the larger value. Specifically, the calculation process of this function is as follows: the input single-point data value is strictly compared with the baseline value 0, and the larger value between the two is always extracted as the final output result of a single operation. For example, when the input scalar value of a specific channel at a certain spatial location in the initial feature difference matrix is ​​negative (e.g., -2.5), it indicates that there is no positive evolutionary difference between the lesion and background features in that area. The activation function outputs the larger value 0 by comparison, thereby directly suppressing the negative feedback channel to zero; conversely, when the input value is positive (e.g., 3.2), the activation function directly transmits and outputs the original positive value 3.2 to fully preserve the significant pathological difference response. By comparing and filtering each element and performing nonlinear mapping operations, this embodiment completely eliminates the suppressive redundant information in the feature space, generating the spatial difference matrix with high-value representation capabilities.

[0034] Subsequently, to construct a node relationship pre-structure suitable for graph neural network topology reasoning, this embodiment performs tensor reshaping on the spatial difference matrix in terms of channel dimensions, extracting feature vectors corresponding to different spatial locations. In this data processing stage, this embodiment flattens the spatial difference matrix, originally in a three-dimensional geometric structure (e.g., the aforementioned 14x14 spatial resolution and 256 deep channels), into a two-dimensional matrix form through low-level addressing mapping rules. Specifically, this embodiment retains the numerical depth of the 256 feature channels unchanged, stretching and merging the 14x14 two-dimensional spatial coordinate axes into a one-dimensional sequence containing 196 independent spatial locations. After the above-mentioned physical reshaping operation of the tensor dimension, this embodiment successfully extracts 196 independent one-dimensional arrays, each of which constitutes a feature vector of a specific dimension containing 256 element values. These feature vectors are arranged linearly in physical storage, and each feature vector condenses the lesion and microenvironment evolution differences of a specific local spatial block (such as a receptive field with a side length of 32 pixels) in the original gastric medical image, thus successfully discretizing the continuous feature map into basic feature physical entities suitable for graph node interaction.

[0035] Finally, this embodiment calculates the cosine similarity between each pair of feature vectors to construct a global similarity matrix, and normalizes the global similarity matrix to obtain the edge weight prior matrix. Specifically, in this embodiment, a double-loop nested calculation method is used in the set of 196 feature vectors. A dot product operation is performed on any two feature vectors, and the result is divided by the product of the magnitudes of the two feature vectors, thereby accurately quantifying the consistency of their directional angles in the multidimensional feature space. The closer the cosine similarity value is to positive 1, the more homogeneous the pathological evolution patterns of the two spatial locations are. Through traversal calculation, this embodiment generates a symmetric two-dimensional tensor of size 196 by 196 in the GPU memory, which is the global similarity matrix. To eliminate the risk of gradient explosion during graph network training caused by absolute numerical differences, this embodiment further uses an exponential normalization function to perform probabilistic scaling on each row of the global similarity matrix, ensuring that the sum of the similarity weights between each spatial location node and all other nodes is strictly equal to 1. Finally, this embodiment outputs the edge weight prior matrix with standard probability distribution characteristics, perfectly realizing the mathematical-physical mapping from image convolution semantic differences to graph structure edge connection strength.

[0036] Specifically, the "spatial dimension alignment" originates from the fundamental prerequisite of multi-source feature cross-stream fusion in deep learning. Because the first and second feature extraction branches may have slight physical device addressing rounding errors in the edge padding strategy of the underlying convolutional kernel or the truncation of the pooling stride, the output feature maps can easily experience pixel-level offsets. Therefore, "spatial dimension alignment," as a forced verification and tensor resampling error correction mechanism, has the core function of establishing a unified coordinate system reference standard. This ensures that the length, width, and depth values ​​of the two feature maps have absolute consistency when performing matrix subtraction, completely avoiding computational crashes caused by out-of-bounds data dimensions. The "edge weight prior matrix" originates from the real-world need in graph theory to measure the initial topological connectivity between nodes in complex systems. Traditional graph networks typically use a Gaussian distribution to randomly initialize the connection weights of nodes, while this invention directly incorporates macroscopic environmental association information (i.e., this matrix) extracted from real medical images as a preset experience into the network. Its application function is to inject the similarity of morphological evolution of specific regions observed in macroscopic images into the collaborative update mechanism of microomics graph nodes in advance, so that lesion regions with similar inflammatory or carcinogenic microenvironments are forced to be given a very large communication weight threshold when transmitting graph information, thereby providing expert prior guidance with solid physical meaning for graph attention networks and greatly narrowing the feature search space of the model.

[0037] To obtain numerically stable prior weights that are convenient for use as graph connection strength, this embodiment employs row-wise probabilistic scaling of the global similarity matrix. Specifically, for the similarity values ​​in a row corresponding to any spatial location in the matrix, this embodiment first performs exponential mapping on each value in that row to eliminate negative values ​​and enhance differences. Then, it sums all the exponential mapping results in that row as a normalization benchmark. Finally, it divides the exponential result at each position by the sum of the row's total, ensuring that the sum of all weights in that row is strictly one. Through this row-wise probabilistic scaling process, the global similarity matrix can be transformed into a prior matrix of edge weights with probabilistic assignment meaning, thereby reducing the risk of gradient instability caused by inconsistent weight scales during subsequent graph network training.

[0038] Specifically, the expression for the nonlinear activation function is: ; in, It is a non-linear activation function; These are the input scalar or tensor elements of the activation function; To obtain the larger value operator.

[0039] Furthermore, the specific implementation process of step 700 is as follows: To overcome the limitation of traditional independent feature vectors lacking spatial topological interaction, this embodiment constructs a non-Euclidean space data structure. First, this embodiment maps the imagemic feature set into a linear transformation layer as node feature representations, and defines each node feature representation as a graph node. In specific implementation, this embodiment extracts, for example, a 128-dimensional imagemic feature set retained after initial dimensionality reduction, and inputs it into a single-layer perceptron (i.e., the linear transformation layer) composed of learnable weight matrices. Through matrix multiplication, this embodiment maps the original feature vectors into feature tensors with a uniform hidden layer dimension (e.g., the dimension is uniformly reduced to 64), thus obtaining the node feature representations. Subsequently, this embodiment defines, for example, 196 spatially partitioned independent feature tensors as graph nodes in a graph theory model, serving as the physical entity basis for information transmission. Next, this embodiment uses the element values ​​of the edge weight prior matrix as the initial connection weights between the graph nodes to construct the spatial heterogeneity graph containing the graph nodes and their edge connection relationships. Specifically, in this embodiment, a 196x196 two-dimensional similarity matrix generated in the preceding steps is extracted, and its probabilistic values ​​are directly assigned to the directed connections between corresponding graph nodes. If the value corresponding to two points in the matrix is ​​greater than a set sparsity truncation threshold (e.g., 0.05), an information transmission channel is physically established in the memory space of this embodiment, and this value is used as the initial connection weight, thereby successfully constructing the spatial heterogeneity map in the computer memory that highly restores the microscopic and macroscopic anatomical associations of the gastric mucosa.

[0040] Furthermore, after completing the basic graph construction, this embodiment performs linear projection on the node feature representations of source and target graph nodes with edge connections in the graph attention network, and then concatenates and fuses the projected features with the initial connection weights to obtain initial attention coefficients. During the forward propagation of the graph neural network, to calculate the joint effect of micro-organizational differences and macro-environmental responses between nodes, this embodiment allocates a shared linear transformation parameter matrix and simultaneously performs dot product projection operations on the node feature representations of the source graph node (as the information sender) and the target graph node (as the information receiver), aligning them to the same feature subspace used for attention calculation. After projection, this embodiment creatively performs a cross-dimensional tensor concatenation operation: the projected features of the source graph node, the projected features of the target graph node, and the pre-extracted initial connection weights representing differences in macro-morphological evolution are concatenated along the feature channel dimension. For example, this embodiment concatenates two feature vectors of length 64 with a scalar weight of length 1 into a joint feature vector entity of length 129. Subsequently, in this embodiment, the concatenated joint feature vector is input into the shared attention evaluation module for inner product operation, and a scalar value representing the absolute information interaction strength between nodes is physically calculated, namely the initial attention coefficient.

[0041] In calculating the attention coefficient, this embodiment constrains the interaction strength between nodes by two types of information: one is the matching degree between the source node and the target node in the feature space, and the other is the macroscopic morphological evolution similarity prior provided by the edge weight prior matrix. Specifically, this embodiment first maps the features of the source node and the target node to the same feature subspace to ensure their comparability; then, the mapped features and the corresponding prior connection weights are fused along the channel dimension and input into the attention evaluation module to obtain an intermediate scalar representing the interaction strength between the two nodes. The larger this intermediate scalar is, the more consistent the two nodes are in terms of lesion evolution differences and macroscopic association priors, and the higher the contribution ratio of their information transmission channels is allocated in the subsequent weighted aggregation. Through this mechanism of "feature matching information and prior connection information participating together," the neighborhood aggregation of the graph attention network can have a clearer physical meaning in medical images.

[0042] Subsequently, to ensure the numerical stability of the attention mechanism during backpropagation and to conform to the probability distribution characteristics, this embodiment applies a nonlinear activation function to the initial attention coefficients and performs normalization processing to obtain the attention coordination coefficients between the graph nodes. Since the scalar values ​​calculated previously may have extreme negative ranges, this embodiment first uses a linear rectified activation function with leakage correction (i.e., the nonlinear activation function) to perform feature response enhancement on the initial attention coefficients. For example, when the input scalar is negative, this embodiment assigns it a fixed slope of 0 to 2 for linear decay to ensure that the network gradient does not completely die, enhancing the ability to capture subtle evolutionary differences. Next, this embodiment performs normalization operations on the activated coefficients along the dimensions of all adjacent nodes of the same target graph node based on a normalized exponential function. Through power operations and summation / division operations on the mathematical base, this embodiment strictly maps the chaotic real values ​​to a probability range of 0 to 1, and ensures that the sum of all incoming edge weights received by a single target node is always equal to 1, thereby outputting smooth attention coordination coefficients that conform to the communication allocation rules in the physical device.

[0043] Furthermore, the normalization process in this embodiment uses "all adjacent incoming edges of the same target node" as the normalization range. This means that the interaction strengths of all adjacent nodes received by the same target node are uniformly scaled probabilistically, ensuring that the weights of each incoming edge are on the same comparable scale and that the sum of these incoming edge weights is a fixed constant. Through this constraint, the graph network can achieve stable information allocation when performing neighborhood feature aggregation: strongly correlated neighbors receive a higher contribution ratio, while the contributions of weakly correlated neighbors are compressed, thereby improving the numerical stability of state updates and the interpretability of aggregation results.

[0044] Finally, this embodiment uses the attention synergy coefficient to perform a weighted summation of the node feature representations of adjacent graph nodes to obtain the updated node features, and then performs a global pooling operation on the updated node features to obtain the deep image omics fusion features. In the information aggregation stage of the graph structure, this embodiment assigns a corresponding information transfer ratio to the node feature representation of each source graph node according to the normalized output attention synergy coefficient. Subsequently, this embodiment performs a tensor matrix summation operation on the weighted features of all directly adjacent nodes of the target graph node (e.g., setting the maximum number of effective neighboring nodes of the local receptive field to 8), thereby fully absorbing the information of the evolutionary changes in the surrounding inflammation-cancer transformation microenvironment. This summation result will cover the original memory address region of the target graph node, completing a single iteration of node state update and generating the updated node features. After undergoing multiple layers (e.g., stacked 3 layers) of graph attention aggregation operations, in order to adapt to the input specifications of the terminal classifier, this embodiment performs a global average pooling operation on the set matrix composed of all 196 updated node features. By taking the arithmetic mean along the node count axis, this embodiment completely reduces the dimensionality of the two-dimensional graph structure feature matrix into a high-order one-dimensional global vector of length 64. This one-dimensional vector perfectly condenses the microscopic texture of local radiomics and the macroscopic topological information of the surrounding mucosal environment, i.e., the final output deep radiomics fusion feature, providing solid data entity support for the automated assessment of pathological outcomes.

[0045] Furthermore, the specific implementation process of step 800 is as follows: The deep radiomics fusion features are input into a probabilistic mapping classifier to obtain the progression risk probability distribution of low-grade gastric intraepithelial neoplasia, and the final outcome assessment result is obtained based on the progression risk probability distribution. Specifically, to achieve a physical leap from high-dimensional abstract features to specific clinical diagnostic indicators, this embodiment inputs the deep radiomics fusion features into the progression prediction classifier containing multiple fully connected layers for nonlinear dimensionality reduction to obtain a progression category feature vector. At the underlying physical storage and execution level of the computer, this embodiment extracts, for example, a one-dimensional global high-order feature vector of length 64 produced in the preceding steps, and feeds it sequentially into a classification network entity containing multiple hidden layer nodes. Each fully connected layer performs intensive tensor multiplication and addition and nonlinear activation truncation operations on the input features through its internally allocated learnable synaptic weight matrix and bias tensor. After multi-level data transmission, the micro-omics and macro-topology fusion representation originally distributed in 64-dimensional space is stripped of redundancy and densely compressed at each level, finally converging to output a low-dimensional floating-point array structure containing only 3 node values. The array structure is physically instantiated in video memory as the regression category feature vector, successfully achieving engineering-level shrinkage of feature dimensions and high purification of core discriminative information.

[0046] Furthermore, in order to transform the floating-point classification features, which do not yet possess clear statistical significance, into a risk assessment system that conforms to medical evidence-based logic, this embodiment performs probability mapping on the regression category feature vector to obtain the progression risk probability distributions corresponding to the three categories of reversal, regression, maintenance, and deterioration. In the specific underlying data processing pipeline, this embodiment calls the normalized exponential function processing module for the aforementioned generated regression category feature vector containing three node values. This processing module first calculates the exponent with the natural constant as the base for each of the three discrete real values, thereby forcibly mapping the feature response of arbitrary polarity to a positive excitation value; then the processor calculates the sum of these three positive excitation values ​​as a common denominator, and then determines the strict proportion of each feature channel in the sum. Through this underlying hardware instruction scheduling, this embodiment generates a normalized array with a constant length of 3 and an absolute sum of values ​​of its internal elements within the memory allocation area. The three elements of this array are strictly bound in sequence to the three targeted evolutionary directions of the natural course of low-grade gastric intraepithelial neoplasia on the physical output interface, accurately and materially outputting the probability distribution of progression risk that reflects the future dynamic evolution tendency of the lesion.

[0047] Finally, after obtaining the support of a standard statistical probability model, this embodiment compares the probability values ​​of the three categories in the progression risk probability distribution, extracts the category with the highest probability value, and uses the category corresponding to the highest probability value as the final outcome assessment result for low-grade gastric intraepithelial neoplasia. At the final node of the engineered execution, this embodiment schedules the logic comparator instruction set of the central processing unit to perform rapid numerical sorting and threshold determination on the three floating-point probability values ​​in the distribution array. For example, when the probability value representing reversal regression is 0.15, maintaining the status quo is 0.10, and progression deterioration is 0.75, this embodiment accurately locates the response channel index containing the highest value, 0.75, through an addressing algorithm, and extracts the high-risk category label anchored to that channel. Subsequently, this embodiment encapsulates the category label mapped by the highest probability value into a structured data message and outputs it to a visible terminal interface or downstream medical decision support system, formally establishing it as the final outcome assessment result. This modular output step completely breaks through the subjective limitations of traditional biopsy pathology in terms of space and time span, physically delivering a highly robust and consistent outcome prediction report, providing precise quantitative basis for clinicians to formulate intervention strategies for invasive surgical resection or non-invasive conservative follow-up.

[0048] Specifically, to improve the model's stability in representing the highly heterogeneous microenvironment of low-grade gastric intraepithelial neoplasia, the graph attention network employs a multi-head attention mechanism. Regarding the process of using the attention coordination coefficient to weightedly sum the node feature representations of adjacent graph nodes to obtain the updated node features, this embodiment, based on the multi-head attention mechanism, performs multiple parallel calculations of the attention coordination coefficient in multiple independent feature subspaces. Specifically, at the physical hardware thread scheduling level, this embodiment splits the original single graph attention calculation task into, for example, eight completely independent parallel computation branches. For each branch, this embodiment uses its own dedicated linear projection matrix to independently map the graph node feature vectors in the original high-dimensional space (e.g., a 64-dimensional tensor space) to a lower-dimensional, non-interfering feature subspace (e.g., a low-dimensional tensor space of length 8). Within this isolated physical memory space, this embodiment drives the underlying tensor operation unit to independently perform inner product matching and normalization processing for each group of projected features, thereby solving in parallel the attention coordination coefficients between 8 groups of graph nodes with completely different focuses, laying a rich computing power foundation for capturing the evolution law of the microenvironment from a multi-dimensional perspective.

[0049] Furthermore, in this embodiment, within the aforementioned multiple independent feature subspaces, a weighted summation operation with the node feature representation is simultaneously performed to obtain multiple subspace node features. During the specific matrix multiplication and addition operation execution phase, this embodiment strictly constrains each set of attention collaboration coefficients calculated previously to be effective only within the corresponding specific feature subspace. This embodiment controls the central processing unit to call multi-threaded concurrent instructions to simultaneously extract node feature representations of the target graph node surrounding a node, with a maximum set of 10 effective neighboring nodes, within 8 subspaces. Subsequently, this embodiment utilizes the dedicated attention weight values ​​within each subspace to perform parallel scalar multiplication and matrix accumulation operations on the local features of these neighboring nodes. Through this high-concurrency underlying hardware data flow, this embodiment effectively avoids gradient collapse or feature assimilation defects that may result from a single attention mechanism, simultaneously producing 8 independent tensor entities, each capturing a specific inflammatory cancer pathological association pattern, i.e., the multiple subspace node features, in physical memory.

[0050] Finally, to reintegrate the multi-dimensional local representations into a unified global high-order semantic entity, this embodiment performs dimensionality reduction processing on the features of each subspace node to obtain the updated node features. In the engineering implementation of data structure reorganization, this embodiment uses tensor splicing instructions to sequentially concatenate the eight parallel-generated subspace node features of length 8 along the feature channel axis, physically reconstructing a joint feature vector with a length restored to 64. To eliminate potential information redundancy at the multi-head splicing interface and promote deep fusion of features from different subspaces, this embodiment then feeds the joint feature vector into a fully connected linear layer composed of internal weight parameter matrices for dimensionality reduction and mapping. After this round of intensive data matrix inner product transformation, the local topological features originally scattered across multiple computation heads are completely converged and fused, ultimately physically instantiating and outputting a scale-standard and highly information-condensed comprehensive tensor array, i.e., the updated node features. This multi-head parallel and end-to-end dimensionality reduction and reorganization physical computation architecture significantly enhances the accuracy and generalization robustness of the network in characterizing the evolution of the complex microenvironment of the gastric mucosa.

[0051] More specifically, in one optional implementation, the lesion segmentation network, multi-branch convolutional neural network, spatial heterogeneity map fusion network, and outcome prediction classifier are all pre-trained using supervised learning. Specifically, the lesion segmentation network takes gastric medical images as input and lesion contour masks annotated by clinical experts as supervision signals. Through iterative optimization, the consistency between the lesion mask output by the network and the annotation results is gradually improved. The multi-branch convolutional neural network takes the region of interest of the lesion and the extended background region as dual inputs. Through an end-to-end training strategy, it learns deep semantic feature representations of the lesion body and the surrounding microenvironment, respectively, and solidifies its parameters for subsequent difference calculations. The spatial heterogeneity map fusion network and the outcome prediction classifier are jointly trained using the outcome category determined by follow-up as supervision signals, so that the fused features after graph structure aggregation have clear category discrimination capabilities. After training, the above model parameters are solidified and deployed to the inference end. The inference stage is strictly executed in the order of steps 100 to 800, which outputs the progression risk probability distribution and the final outcome assessment result, thereby ensuring that the method of the present invention can be repeatedly implemented in clinical auxiliary decision-making scenarios.

[0052] In one engineered deployment approach, this embodiment can deploy the lesion segmentation network, multi-branch convolutional neural network, and spatial heterogeneity graph fusion network on a computing platform with parallel computing capabilities to improve inference efficiency and stability in clinical scenarios. Specifically, the operations such as convolution, pooling, element-wise differencing, similarity matrix calculation, and neighborhood weighted aggregation of the graph attention network can all be implemented using parallel tensor computation and accelerated by a GPU with large-scale parallel instruction scheduling capabilities. This shortens the overall time from input to risk probability distribution output for a single patient image while ensuring output consistency, meeting the real-time and batch processing requirements of clinical decision support.

[0053] This embodiment also provides a system for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features, including: The image acquisition and localization module is used to acquire gastric medical images of the patient to be evaluated and extract the region of interest for the lesion in the gastric medical images; An extended background region construction module is used to expand outward from the boundary of the region of interest of the lesion to construct an extended background region that includes the surrounding mucosal microenvironment. The radiomics feature extraction module is used to transform the region of interest of the lesion and extract a set of radiomics features that reflect the heterogeneity of tissue microtexture. The deep semantic feature extraction module is used to input the region of interest of the lesion and the extended background region into a pre-trained multi-branch convolutional neural network, and output a first deep semantic feature map and a second deep semantic feature map respectively. The spatial difference calculation module is used to calculate the spatial feature difference between the first deep semantic feature map and the second deep semantic feature map to obtain a spatial difference matrix; The edge weight prior acquisition module is used to calculate the correlation between feature elements based on the spatial difference matrix to obtain the edge weight prior matrix; The spatial heterogeneity map fusion module is used to construct a spatial heterogeneity map, map each feature in the image omics feature set to a graph node, use the edge weight prior matrix as the initial connection weight, and calculate the attention coordination coefficient between the graph nodes through a graph attention network and perform state updates to obtain deep image omics fusion features. The outcome assessment and prediction module is used to input the deep radiomics fusion features into the outcome prediction classifier for probability mapping, obtain the progression risk probability distribution of low-grade gastric intraepithelial neoplasia, and obtain the final outcome assessment result based on the progression risk probability distribution.

[0054] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0055] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features, characterized in that, include: Acquire gastric medical images of the patient to be evaluated, and extract the region of interest for the lesion from the gastric medical images; Based on the boundary of the region of interest of the lesion, an extended background region including the surrounding mucosal microenvironment is constructed by expanding outward. The region of interest of the lesion is transformed to extract a set of radiomics features that reflect the heterogeneity of tissue microtexture; The region of interest of the lesion and the extended background region are respectively input into a pre-trained multi-branch convolutional neural network, which outputs a first deep semantic feature map and a second deep semantic feature map respectively. Calculate the spatial feature difference between the first deep semantic feature map and the second deep semantic feature map to obtain the spatial difference matrix; The correlation between feature elements is calculated based on the spatial difference matrix to obtain the edge weight prior matrix; A spatial heterogeneity map is constructed, and each feature in the image omics feature set is mapped to a graph node. The edge weight prior matrix is ​​used as the initial connection weight, and the attention coordination coefficient between the graph nodes is calculated and the state is updated through a graph attention network to obtain deep image omics fusion features. The deep radiomics fusion features are input into the outcome prediction classifier for probability mapping to obtain the progression risk probability distribution of low-grade gastric intraepithelial neoplasia, and the final outcome assessment result is obtained based on the progression risk probability distribution.

2. The method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features according to claim 1, characterized in that, The process of acquiring gastric medical images of the patient to be evaluated and extracting regions of interest for lesions from the gastric medical images includes: The initial gastric medical images of the patient to be evaluated are acquired, and the initial gastric medical images are subjected to image denoising and pixel normalization processing to obtain the gastric medical images. The gastric medical images are input into a pre-trained lesion segmentation network to obtain an initial lesion mask; The initial lesion mask is subjected to morphological boundary smoothing to obtain the target lesion mask; Based on the target lesion mask, the gastric medical image is localized and cropped to obtain the region of interest for the lesion.

3. The method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features according to claim 1, characterized in that, The expansion outward from the boundary of the region of interest of the lesion to construct an extended background region including the surrounding mucosal microenvironment includes: Extract the boundary contour coordinates of the region of interest of the lesion, and generate the corresponding binary mask of the lesion in the global coordinate system of the gastric medical image; The preset physical reference distance of the microenvironment is obtained, and the size of the structural element for morphological operations is calculated by combining the pixel spatial resolution of the gastric medical image. Based on the size of the structural element, an isotropic structural element is constructed, and the isotropic structural element is used to perform a morphological dilation operation on the binarized mask of the lesion to obtain an expanded binarized mask. Calculate the pixel logical difference set between the extended binarized mask and the lesion binarized mask to obtain the peripheral mucosal ring mask; The extended background region is obtained by using the peripheral mucosal ring mask to perform pixel mapping and cropping of the gastric medical image in terms of spatial location.

4. The method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features according to claim 1, characterized in that, The transformation processing of the region of interest in the lesion to extract a set of radiomics features reflecting the heterogeneity of tissue microtexture includes: The region of interest of the lesion was subjected to multi-scale filtering using wavelet transform filter and Laplace Gaussian filter respectively to obtain multi-scale derived image; First-order statistical features, two-dimensional shape features, and high-order texture features are extracted from the region of interest of the lesion and the multi-scale derived image, respectively, to obtain an initial radiomics feature pool; Principal component analysis is used to perform feature dimensionality reduction and redundancy removal on the initial radiomics feature pool to obtain the radiomics feature set.

5. The method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features according to claim 1, characterized in that, The step of inputting the region of interest of the lesion and the extended background region into a pre-trained multi-branch convolutional neural network, respectively, and outputting a first deep semantic feature map and a second deep semantic feature map, includes: Construct the multi-branch convolutional neural network that includes a first feature extraction branch and a second feature extraction branch; The region of interest of the lesion is input into the first feature extraction branch to perform multi-layer convolution and pooling operations to extract feature representations that characterize the macroscopic pathological appearance inside the lesion, and the first deep semantic feature map is output accordingly. The extended background region is simultaneously input into the second feature extraction branch for multi-layer convolution and pooling operations to extract feature representations of the evolution state of the microenvironment surrounding the lesion, and the corresponding second deep semantic feature map is output.

6. The method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features according to claim 1, characterized in that, The spatial feature difference between the first deep semantic feature map and the second deep semantic feature map is calculated to obtain a spatial difference matrix; Based on the spatial difference matrix, the correlation between feature elements is calculated to obtain the edge weight prior matrix, including: The first deep semantic feature map and the second deep semantic feature map are aligned in spatial dimensions, and element-wise subtraction is performed to obtain an initial feature difference matrix. The initial feature difference matrix is ​​input into a nonlinear activation function to enhance the feature response, thereby obtaining the spatial difference matrix; The spatial difference matrix is ​​reshaped using a channel-dimensional tensor to extract feature vectors corresponding to different spatial locations; Calculate the cosine similarity between each pair of feature vectors to construct a global similarity matrix, and normalize the global similarity matrix to obtain the edge weight prior matrix.

7. The method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features according to claim 1, characterized in that, The construction of the spatial heterogeneity map maps each feature in the imageomics feature set to a graph node, uses the prior edge weight matrix as the initial connection weight, and calculates the attention coordination coefficient between the graph nodes through a graph attention network and performs state updates to obtain deep imageomics fusion features, including: The image omics feature set is input into a linear transform layer and mapped to node feature representations, and each node feature representation is defined as a graph node. The element values ​​of the edge weight prior matrix are used as the initial connection weights between the graph nodes to construct the spatial heterogeneity graph containing the graph nodes and their edge connection relationships. In the graph attention network, the node feature representations of source graph nodes and target graph nodes with edge connections are linearly projected, and the projected features are concatenated and fused with the initial connection weights to obtain the initial attention coefficients. The initial attention coefficients are applied with a nonlinear activation function and then normalized to obtain the attention coordination coefficients between the graph nodes. The attention collaboration coefficient is used to perform a weighted summation of the node feature representations of adjacent graph nodes to obtain the updated node features, and then a global pooling operation is performed on the updated node features to obtain the deep image omics fusion features.

8. The method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features according to claim 1, characterized in that, The process involves inputting the deep radiomics fusion features into a probabilistic mapping classifier to obtain the progression risk probability distribution of low-grade gastric intraepithelial neoplasia, and then obtaining the final outcome assessment result based on the progression risk probability distribution, including: The deep image omics fusion features are input into the regress prediction classifier containing multiple fully connected layers for nonlinear dimensionality reduction to obtain the regress category feature vector. By performing probability mapping on the feature vectors of the regression categories, the probability distributions of the progress risk corresponding to the three categories of reversal and regression, maintaining the status quo, and progress deterioration are obtained respectively; By comparing the probability values ​​of the three categories in the probability distribution of progression risk, the category with the highest probability value is extracted, and the category corresponding to the highest probability value is taken as the final outcome assessment result of low-grade gastric intraepithelial neoplasia.

9. A method for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features according to claim 7, characterized in that, The graph attention network employs a multi-head attention mechanism. The node feature representations of adjacent graph nodes are weighted and summed using the attention coordination coefficients to obtain the updated node features, including: Based on the multi-head attention mechanism, the calculation of the attention coordination coefficient and the weighted summation of the node feature representation are performed in parallel multiple times in multiple independent feature subspaces to obtain node features in multiple subspaces. The features of each subspace node are concatenated and dimensionality reduced to obtain the updated node features.

10. A system for assessing the prognosis of low-grade gastric intraepithelial neoplasia based on deep learning radiomics features, characterized in that, include: The image acquisition and localization module is used to acquire gastric medical images of the patient to be evaluated and extract the region of interest for the lesion in the gastric medical images; An extended background region construction module is used to expand outward from the boundary of the region of interest of the lesion to construct an extended background region that includes the surrounding mucosal microenvironment. The radiomics feature extraction module is used to transform the region of interest of the lesion and extract a set of radiomics features that reflect the heterogeneity of tissue microtexture. The deep semantic feature extraction module is used to input the region of interest of the lesion and the extended background region into a pre-trained multi-branch convolutional neural network, and output a first deep semantic feature map and a second deep semantic feature map respectively. The spatial difference calculation module is used to calculate the spatial feature difference between the first deep semantic feature map and the second deep semantic feature map to obtain a spatial difference matrix; The edge weight prior acquisition module is used to calculate the correlation between feature elements based on the spatial difference matrix to obtain the edge weight prior matrix; The spatial heterogeneity map fusion module is used to construct a spatial heterogeneity map, map each feature in the image omics feature set to a graph node, use the edge weight prior matrix as the initial connection weight, and calculate the attention coordination coefficient between the graph nodes through a graph attention network and perform state updates to obtain deep image omics fusion features. The outcome assessment and prediction module is used to input the deep radiomics fusion features into the outcome prediction classifier for probability mapping, obtain the progression risk probability distribution of low-grade gastric intraepithelial neoplasia, and obtain the final outcome assessment result based on the progression risk probability distribution.

Images