An ophthalmic OCT image layer structure automatic segmentation method and system

Abstract

The present application relates to the technical field of ophthalmology, and in particular to an ophthalmic OCT image layer structure automatic segmentation method and system, comprising the following steps: obtaining an ophthalmic OCT image converted from an original physical interference signal collected by an optical coherence tomography device, and calculating the light reflectivity gradient of each pixel point in the OCT image as a prior initial feature representing the optical attenuation physical property; inputting the OCT image into a lesion perception encoder to extract the lesion feature tensor of the image, and outputting a two-dimensional pathological heat map representing the probability of the existence of pathological lesions in the corresponding region of the image through a lesion detection head in the lesion perception encoder. The present application effectively solves the technical problems of deep feature loss, lesion boundary misjudgment and topological layer sequence confusion in the complex pathological OCT image of the traditional deep learning model by deeply fusing the optical attenuation physical law, the lesion negative perception mechanism and the graph neural network anatomy priori.

Description

Technical Field

[0001] This invention relates to the field of ophthalmology, specifically to an automatic layer structure segmentation method and system for ophthalmic OCT images. Background Technology

[0002] Optical coherence tomography (OCT), a non-invasive high-resolution ophthalmic imaging technique, can clearly visualize the physiological structures of the retina, serving as a crucial basis for the clinical diagnosis of fundus diseases such as macular edema and drusen. In recent years, deep learning-based convolutional neural networks have been widely applied to the automatic segmentation of retinal layers in OCT images. However, existing segmentation algorithms still suffer from the following core technical limitations when dealing with clinical OCT images containing severely destructive lesions: First, existing image feature extraction algorithms often deviate from the optical physics foundation of OCT hardware imaging. In actual imaging, the probe beam undergoes severe nonlinear scattering and absorption dissipation as it penetrates deep retinal biological tissue, causing the reflectivity gradient of deep tissues (such as the choroid) to decay exponentially. Current techniques mostly employ uniform, conventional convolutional kernels to extract global gradients, failing to introduce optical physics compensation mechanisms for depth coordinates. This results in weak boundaries in deep tissues being easily masked by multiplicative speckle noise, leading to feature loss.

[0003] Secondly, existing deep learning attention mechanisms suffer from feature-directed errors when dealing with edema or cystic lesions. Conventional attention mechanisms follow the logic of "positive incentive for salient features," but in pathological fundus conditions, the edges of edema lesions often exhibit extremely high local image contrast. Existing models are highly prone to misclassifying these high-intensity "pseudo-boundaries" generated by these pathological areas as normal physiological layer boundaries of the retina, leading to serious misguidance in feature extraction by the network.

[0004] Finally, existing fully convolutional network architectures lack anatomical topological constraints on the strict physiological spatial hierarchy of the retina. The retina possesses a rigorous top-to-bottom layered physical property, but when local image features are extensively damaged due to severe lesions, existing networks often output predictions that violate biological anatomy, such as inverted depth coordinates or spatial layer intersections between adjacent layers. Furthermore, existing conventional loss functions (such as cross-entropy loss) only evaluate network performance based on pixel-level numerical errors, equating "layer intersection" errors with severe spatial violations to ordinary pixel classification biases. This fails to apply targeted topological correction penalties during model backpropagation, resulting in structural breaks and hierarchical disorder in the final output segmentation surface. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention provides an automatic layer structure segmentation method for ophthalmic OCT images, including the following steps: S100, acquiring an ophthalmic OCT image converted from the original physical interference signal acquired by an optical coherence tomography device, and calculating the light reflectance gradient of each pixel in the OCT image as a priori initial feature characterizing the physical properties of optical attenuation. S200. Input the OCT image into the lesion perception encoder to extract the lesion feature tensor of the image, and output a two-dimensional pathological heat map representing the probability of the presence of pathological lesions in the corresponding area of ​​the image through the lesion detection head in the lesion perception encoder. S300: Construct a graph network structure using preset retinal physiological layers as graph network nodes and inter-layer anatomical adjacent position relationships as edges. Aggregate the prior initial features into the corresponding graph network nodes. Use a graph convolutional network to infer the graph network structure and output a topology correction vector to constrain the relative physical spatial positions of each retinal layer. S400. Based on the lesion existence probability of each region in the two-dimensional pathological heat map, a dynamic cross-gating signal is constructed. In a collaborative manner, the topology correction weight is adaptively increased as the lesion existence probability increases. The lesion feature tensor and the topology correction vector are dynamically cross-gated and fused to obtain a high-dimensional fusion feature with lesion robustness and topology consistency. S500: The high-dimensional fusion features are decoded by the decoder to output the spatial boundaries of each layer of the retina, thus completing the image layer structure segmentation. During the model training phase, the lesion-aware encoder, graph convolutional network, and decoder utilize an asymmetric topological penalty loss function for iterative optimization of network parameters. This asymmetric topological penalty loss function is configured such that when a layer crossing that violates the anatomical physical order is detected in the predicted depth coordinates of adjacent layers along the longitudinal scanning depth direction, an exponentially amplified penalty weight is applied based on the difference in crossing depth between the adjacent layers.

[0006] Further, in step S100, the light reflectance gradient of each pixel in the OCT image is calculated, specifically including the following steps: A two-dimensional spatial coordinate system is established with the transverse cross-sectional scan as the X-axis and the longitudinal depth scan (A-scan direction) as the Z-axis; the coordinates of the OCT image are extracted. For each pixel, obtain its corresponding physical light reflection intensity value. ; The initial longitudinal gradient is calculated along the longitudinal depth scanning direction using the center difference algorithm. The calculation formula is: ; Combining the exponential attenuation characteristics of the light beam within biological tissue in optical coherence tomography (OCT), a physical attenuation compensation coefficient is introduced. The optical reflectivity gradient after depth attenuation compensation is calculated. The calculation formula is: ; The light reflectance gradient after depth attenuation compensation The maximum and minimum values ​​are normalized to form the prior initial features that characterize the physical properties of optical attenuation. Before calculating the light reflectance gradient of each pixel in the OCT image, the method further includes: Adaptive nonlocal mean filtering is performed on the OCT image to denoise it. The filtering weights are calculated by utilizing the high similarity of the internal structure of the image, which reduces optical speckle noise while preserving the original interlayer physical reflection edge of the retina.

[0007] Furthermore, the lesion detection encoder includes a multi-scale convolutional feature extraction network and a lightweight lesion auxiliary detection head, and S200 specifically includes the following steps: The OCT image containing the prior initial features is input into the multi-scale convolutional feature extraction network to extract a multi-channel shallow feature tensor containing image edge gradients. ; The multi-channel shallow feature tensor The data is input into the lightweight lesion-assisted detection head, and sequentially undergoes pointwise convolutional dimensionality reduction and nonlinear activation function mapping to output the two-dimensional pathological heatmap. ; Among them, the two-dimensional pathological thermogram The dimension is consistent with the two-dimensional spatial resolution of the OCT image, and the coordinates of each pixel in the image are... numerical value Between 0 and 1 The magnitude of the value positively represents the probability of the presence of destructive pathological lesions such as macular edema or vitreous condyloma in the corresponding spatial coordinate region.

[0008] Furthermore, the lesion-aware encoder also includes a lesion-suppression attention module, and the step of extracting the lesion feature tensor of the image specifically includes: The two-dimensional pathological thermogram The spatial attention mask signal is input into the lesion suppression attention module; Within the lesion suppression attention module, a negative activation function based on pathological probability is constructed, and the lesion suppression attention weight matrix is ​​calculated. The calculation formula is: ; in, To and A matrix of all one dimensions; Using the lesion-inhibiting attention weight matrix For the multi-channel shallow feature tensor Perform pixel-by-pixel weighted Hadamard product calculation, the formula is: ; The weighted tensor After mapping by a deep convolutional layer, the lesion feature tensor is output; through the pixel-by-pixel Hadamard product weighting operation, the lesion probability in the two-dimensional pathological heatmap is calculated. The higher the region, the lower the weight of the corresponding shallow image edge gradient features are forced, thereby achieving adaptive physical suppression of damaged pathological edge features.

[0009] Further, in step S300, the construction of the graph network structure using preset retinal physiological layers as graph network nodes and inter-layer anatomical adjacent position relationships as edges specifically includes the following steps: To the retina Each physiological layer, arranged according to the physical longitudinal depth from shallow to deep in the actual anatomical order, is defined as a set of nodes in a graph network structure. ; Based on the actual anatomical biological sequence, an adjacency matrix representing the anatomical adjacency relationships between layers is constructed. , where matrix elements The mathematical value selection rules are constrained as follows: If and only if the first Physiological Stratification and the First In ophthalmic anatomy, the physiological layers are strictly adjacent to each other, meaning they satisfy the absolute value condition. When, define ; For other inter-layer or non-adjacent inter-layer relationships, define ; Based on the adjacency matrix A self-connection mechanism is introduced to construct a physical location topology graph and calculate the self-loop adjacency matrix. ,in To and Identity matrices of the same dimension.

[0010] Further, in step S300, the reasoning of the graph network structure using a graph convolutional network to output a topological correction vector for constraining the relative physical spatial positions of each layer of the retina specifically includes the following calculation process: Using a local adaptive pooling algorithm, the prior initial features in each physiological layer region of the OCT image are extracted and aggregated to the corresponding nodes. In the middle, the splicing forms the initial node feature matrix. ; Obtain the self-loop adjacency matrix The calculated diagonal degree matrix The diagonal elements of the degree matrix satisfy... ; Forward propagation of physical space correlation features is performed through a graph convolutional neural network. The formula for calculating graph convolution propagation of a layer is: ; in, For the first The learnable weight matrix of a layered graph convolutional network. It is a non-linear activation function; After multi-layer graph convolutional propagation iterations, the output feature matrix of the last layer node is used as the topology correction vector; through the degree matrix Adjacency matrix with self-loop Regularized multiplication operations ensure that graph convolutional networks are strictly restricted to the anatomically adjacent layers during inference. Information is transmitted within the physical topological constraints.

[0011] Furthermore, the collaborative method of constructing a dynamic cross-gating signal based on the lesion presence probability of each region in the two-dimensional pathological heatmap, and adaptively increasing the topology correction weight as the lesion presence probability increases, specifically includes the following calculation steps: Extract the two-dimensional pathological thermogram output by the lesion sensing encoder, denoted as where pixel coordinates The value at the location Characterizes the probability of the presence of local lesions, and ; Constructing an adaptive topology-gated weight matrix Using natural constants The exponential activation function maps the probability of the presence of the local lesion to a topological correction weight, and the calculation formula is as follows: ; in, These are preset positive gain hyperparameters; The exponential activation function generates a dynamic cross-gated signal, such that in healthy retinal areas, the probability of the presence of lesions increases. Approaching The topology correction weights maintain the basic gain state; in the edematous area of ​​the lesion, due to the probability of the lesion existing... Approaching 1 triggers a nonlinear mutation in the exponential activation function, causing the topology correction weight to be exponentially adaptively amplified, thereby forcibly introducing topological dissection priors to dominate the boundary reasoning of the region. The step of dynamically cross-gating and fusing the lesion feature tensor with the topology correction vector to obtain a high-dimensional fusion feature with lesion robustness and topological consistency specifically includes: Obtain the topology correction vector output by the graph convolutional network, perform a spatial dimension broadcast operation on it, and generate a tensor with the lesion feature. Topological space tensors with the same two-dimensional spatial resolution ; Using the adaptive topology gated weight matrix For the topological space tensor Perform pixel-by-pixel enhanced Hadamard product gating to obtain physically weighted topology correction features. The formula is: ; The lesion feature tensor Topological correction features after physical weighting Cascade along the channel dimension and input to Cross-channel feature dimensionality reduction and fusion are performed in the convolutional layer, and the final high-dimensional fused feature is output.

[0012] Further, in step S500, the asymmetric topology penalty loss function is configured as follows: when a layer intersection violating the anatomical physical order is detected in the predicted depth coordinates of adjacent layers along the longitudinal scanning depth direction, an exponentially amplified penalty weight is applied based on the intersection depth difference of the adjacent layers. The specific calculation rules include: Set along the A-scan physical scan direction along the longitudinal scan depth, the first retinal layer in lateral position The network predicted depth coordinates are , No. The network predicts the depth coordinates of the layer as follows: ,in The total number of layers in the retina; Define the physical depth difference between adjacent layers as ; Construct the asymmetric topological penalty loss function Its mathematical formula is: ; in, The preset penalty scaling hyperparameter; In the backpropagation iteration of network parameters, when the predicted depth conforms to biological anatomy, that is... When the asymmetric topological penalty loss function is 0, the output value of the loss function is 0; when a physical space violation occurs... At that time, the network is affected by the base data The exponential asymmetric penalty forces the model to learn the correct anatomical boundaries in severely affected areas.

[0013] Furthermore, the step of decoding the high-dimensional fused features using a decoder to output the spatial boundaries of each layer of the retina and complete image layer structure segmentation specifically includes: During the model training phase, the basic similarity loss between the predicted segmentation probability map and the real doctor annotation labels is calculated. ; The basic similarity loss With the aforementioned asymmetric topological penalty loss function Linear joint operation is performed according to preset weights, and a global joint loss function is constructed to optimize the joint parameters of the lesion-sensing encoder, graph convolutional network and decoder end-to-end. During the model inference phase, the trained network outputs two-dimensional physical boundary coordinates, and through a three-dimensional surface reconstruction algorithm, the two-dimensional segmentation boundaries of multiple consecutive B-scan frames are assembled and rendered into a three-dimensional layered digital model of the retina without hierarchical intersections or pathological breaks.

[0014] An automatic layer structure segmentation system for ophthalmic OCT images, comprising: The physical prior feature extraction module is used to acquire ophthalmic OCT images converted from the original physical interference signals collected by the optical coherence tomography (OCT) equipment, and to calculate the light reflectance gradient of each pixel in the OCT image as a prior initial feature characterizing the physical properties of optical attenuation. The lesion perception and extraction module is used to input the OCT image into the lesion perception encoder to extract the lesion feature tensor of the image, and output a two-dimensional pathological heat map representing the probability of the presence of pathological lesions in the corresponding area of ​​the image through the lesion detection head in the lesion perception encoder. The anatomical topology reasoning module is used to construct a graph network structure with preset retinal physiological layers as graph network nodes and the anatomical adjacent position relationship between layers as edges. The prior initial features are aggregated into the corresponding graph network nodes, and the graph network structure is reasoned using a graph convolutional network to output a topology correction vector used to constrain the relative physical spatial position of each retinal layer. The dynamic cross-gated fusion module is used to construct a dynamic cross-gated signal based on the lesion existence probability of each region in the two-dimensional pathological heat map. In a collaborative manner, the topology correction weight is adaptively increased as the lesion existence probability increases. The lesion feature tensor and the topology correction vector are dynamically cross-gated and fused to obtain a high-dimensional fusion feature with lesion robustness and topological consistency. The topology consistency decoding and optimization module is used to decode the high-dimensional fusion features through the decoder, output the spatial boundaries of each layer of the retina, and complete the image layer structure segmentation. In the model training phase, the lesion perception encoder, graph convolutional network and decoder in the system use an asymmetric topology penalty loss function to iteratively optimize the network parameters. The asymmetric topology penalty loss function is configured as follows: when a layer crossing that violates the anatomical physical order is detected in the predicted depth coordinates of adjacent layers along the longitudinal scanning depth direction, an exponentially amplified penalty weight is applied based on the difference in the crossing depth of the adjacent layers.

[0015] This invention effectively solves the technical problems of traditional deep learning models in complex pathological OCT images, such as loss of deep features, misjudgment of lesion boundaries, and disordered topological order, by deeply integrating the physical laws of optical attenuation, the negative perception mechanism of lesions, and the anatomical priors of graph neural networks. This invention introduces an exponential optical attenuation compensation formula based on vertical depth to restore the true underlying physical boundary features; by constructing an inverse mask matrix based on pathological heatmaps, it utilizes Hadamard product operations to achieve efficient suppression and adaptive physical isolation of pathological "pseudo-boundaries" such as edema in the underlying feature space; and further constructs a graph network branch constrained by an absolute value diagonal matrix, which, in conjunction with an asymmetric topological penalty loss function, applies an exponential calculus penalty to correct "layer crossover" predictions that violate biological anatomical order during model training and inference. Through the synergistic operation of the above multi-dimensional algorithmic structure and physical and mathematical constraints, the system can still output a three-dimensional continuous layered structure model without breaks, without layer crossovers, and strictly conforming to medical anatomical norms, even in areas where retinal tissue is severely damaged by lesions. This provides a highly robust and accurate digital structural basis for clinical ophthalmological diagnosis. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a schematic diagram of the lesion sensing encoder architecture of the present invention; Figure 3 This is a schematic diagram of the anatomical topology reasoning graph network construction of the present invention; Figure 4 This is a schematic diagram of the dynamic cross-gating fusion mechanism of the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but includes other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] The present invention will now be described in further detail with reference to the accompanying drawings: Example: Figures 1-4 As shown, an automatic layer structure segmentation method for ophthalmic OCT images includes the following steps: S100. Acquire ophthalmic OCT images converted from the original physical interference signals collected by the optical coherence tomography (OCT) device, and calculate the light reflectance gradient of each pixel in the OCT image as a priori initial feature characterizing the physical properties of optical attenuation. S200. Input the OCT image into the lesion perception encoder to extract the lesion feature tensor of the image, and output a two-dimensional pathological heat map representing the probability of the presence of pathological lesions in the corresponding area of ​​the image through the lesion detection head in the lesion perception encoder. S300: Construct a graph network structure using the preset physiological layers of the retina as graph network nodes and the anatomical adjacent position relationship between layers as edges. Aggregate the prior initial features into the corresponding graph network nodes, use the graph convolutional network to reason about the graph network structure, and output the topology correction vector used to constrain the relative physical spatial position of each layer of the retina. S400. Based on the probability of lesion presence in each region of the two-dimensional pathological heat map, a dynamic cross-gated signal is constructed. In a collaborative manner, the topology correction weight is adaptively increased as the probability of lesion presence increases. The lesion feature tensor and the topology correction vector are dynamically cross-gated and fused to obtain a high-dimensional fusion feature with lesion robustness and topological consistency. S500 decodes the high-dimensional fusion features through the decoder and outputs the spatial boundaries of each layer of the retina to complete the image layer structure segmentation. Among them, the lesion perception encoder, graph convolutional network and decoder use the asymmetric topological penalty loss function to iteratively optimize the network parameters during the model training stage. The asymmetric topological penalty loss function is configured as follows: when a layer crossing that violates the anatomical physical order is detected in the predicted depth coordinates of adjacent layers along the longitudinal scanning depth direction, a penalty weight that is exponentially amplified is applied based on the difference in the crossing depth of adjacent layers.

[0020] Furthermore, the specific implementation process of step S100 is as follows: In step S100, the system performs the operation of acquiring an ophthalmic OCT image converted from the original physical interference signal collected by the optical coherence tomography (OCT) device, and calculating the light reflectance gradient of each pixel in the OCT image as a priori initial feature characterizing the physical properties of optical attenuation.

[0021] In the specific implementation process, the original two-dimensional or three-dimensional structural data scanned and output by an optical coherence tomography (OCT) device is first acquired. Due to the inherent physical coherence of the optical interference imaging process, OCT images are usually accompanied by multiplicative speckle noise. To strictly preserve the physical reflection edges between the physiological layers of the retina while filtering out speckle noise, the system performs adaptive nonlocal mean filtering on the OCT image before performing gradient calculation. The specific filtering logic is as follows: utilizing the high similarity of the internal structure of the image, the Euclidean distance between the target pixel block and the candidate pixel block is calculated globally or within a set search window, and the corresponding Gaussian weight is generated by mapping this Euclidean distance as the independent variable; then, the pixel gray values ​​within the search window are weighted and summed using this Gaussian weight, and the denoised pixel values ​​are output. Through this spatial domain weighting operation, the high-frequency pixel abrupt changes caused by speckle noise are smoothed.

[0022] After filtering, the system loads the image data into the processor's memory and establishes a two-dimensional coordinate system with the horizontal cross-sectional scanning direction as the X-axis and the vertical depth scanning direction (A-scan) as the Z-axis. The system then iterates through and extracts the coordinates from the OCT image. For each pixel, read its corresponding physical light reflection intensity value. Considering that the physical morphology of the retinal physiological layer boundary is mainly manifested as abrupt changes in reflectivity along the longitudinal depth direction, this embodiment excludes the calculation of the lateral gradient and calculates the initial longitudinal gradient separately along the longitudinal depth scanning direction using the central difference algorithm. The specific formula for calculating the difference is defined as follows: Through the above calculations, the system constructs the initial directional derivative using the difference in light reflection intensity between adjacent depth pixels.

[0023] During optical coherence tomography (OCT), the probe beam inevitably undergoes physical scattering and absorption dissipation based on depth variations as it penetrates deep retinal tissue, resulting in a nonlinear attenuation of the reflected signal in the deep tissue region. To eliminate the bias caused by this optical attenuation leading to underestimation of the deep tissue gradient calculation, this system incorporates a physical attenuation compensation mechanism within the processor.

[0024] Specifically, the system presets a physical attenuation compensation coefficient. Combining the exponential decay law of light beams within biological tissues, the initial longitudinal gradient... Perform depth attenuation compensation calculation to obtain the light reflectivity gradient. The specific compensation calculation formula is as follows: In the process of this calculation, natural constants are used. Exponential term with base As a coordinate with vertical depth An increasing and monotonically increasing compensation factor. After performing the above multiplication operation, the processor can apply an adaptive numerical amplification mapping to the weak gradient features in deep tissues that are weakened by optical attenuation.

[0025] The system extracts the depth-attenuation-compensated optical reflectivity gradient calculated from the entire image within the tensor processing unit. The global maximum and minimum values ​​are determined, and maximum-minimum normalization is performed. A linear transformation is then used to constrain the attenuation compensation gradient values ​​at all spatial coordinate points to a specific floating-point range, eliminating differences in absolute dimensions. This yields the final prior initial feature matrix characterizing the physical properties of optical attenuation. This prior initial feature matrix is ​​then stored as a floating-point data tensor, serving as the underlying input data for the network model.

[0026] Furthermore, the specific implementation process of step S200 is as follows: The system inputs prior initial features into the lesion perception encoder to extract the lesion feature tensor of the image, and outputs a two-dimensional pathological heat map representing the probability of the presence of pathological lesions in the corresponding area of ​​the image through the lesion detection head in the lesion perception encoder.

[0027] In the specific hardware and software collaborative execution process, the system's graphics processing unit (GPU) first receives the prior initial feature matrix representing the physical properties of optical attenuation output from step S100, and encapsulates it into a low-level data tensor with spatial and channel dimensions. This tensor is then input into the multi-scale convolutional feature extraction network inside the lesion perception encoder. This multi-scale convolutional feature extraction network is configured with parallel convolutional kernels of different receptive field sizes. By performing two-dimensional discrete convolution operations, it extracts a multi-channel shallow feature tensor containing the gradient information of the physical reflection edges of each layer of the retina, while preserving the overall spatial topology. This tensor is denoted as... .

[0028] The system opens up parallel computational branches in the tensor data stream.

[0029] On the one hand, multi-channel shallow feature tensors Import the data into a lightweight lesion detection head. Inside this head, the tensor is first subjected to pointwise convolution. The layer performs a linear weighted summation across channel dimensions to achieve feature dimensionality reduction; the dimensionality-reduced tensor is immediately input into a sigmoid nonlinear activation function for numerical mapping. After this mapping operation, the system outputs a two-dimensional pathological heatmap. In terms of data structure definition, a two-dimensional pathological heatmap... This is a two-dimensional floating-point matrix whose two-dimensional spatial resolution is strictly consistent with the original input OCT image. The spatial coordinates of each pixel in this matrix... The value at the location Strictly constrained within a closed interval of 0.0 to 1.0, the magnitude of this floating-point value positively represents the mathematical probability of the presence of destructive pathological lesions such as macular edema and vitreous warts in the corresponding physical space coordinate region.

[0030] In generating two-dimensional pathological heatmaps Next, the processor proceeds to the calculation phase for lesion suppression feature extraction. The system then generates the two-dimensional pathological heatmap. As a spatial attention mask signal, it is callbacked to the Lesion Suppression Attention Module (LSAM) in the main data stream. Within the computational unit of this module, the processor first performs matrix operations on the negative activation function based on the pathological probability, specifically constructing a matrix... Using a single matrix 1 with the same two-dimensional dimensions, and controlling the arithmetic logic unit to execute pixel-by-pixel matrix subtraction instructions, the lesion suppression attention weight matrix is ​​calculated. Its strict operational expression is .

[0031] After obtaining the weight matrix, the system calls hardware-level tensor multiplication instructions to utilize the lesion suppression attention weight matrix. For the aforementioned multi-channel shallow feature tensor Spatial alignment is performed, and pixel-by-pixel Hadamard product weighted operation is executed. Its algebraic expression is: Through this low-level matrix element-wise multiplication operation, the system achieves physical selection of features at the numerical computation level: Probability of lesions In higher coordinate regions, the corresponding mask weight values ​​approach 0, which, after multiplication, forcibly suppress the gradient magnitude of the damaged pathological edge features in that region; while in healthy regions, the mask weight values ​​approach 1, and after multiplication, the original physical reflection edge gradient is completely preserved. Finally, the tensor after the Hadamard product weighted operation... The data is continuously fed into subsequent deep convolutional layers to perform high-dimensional nonlinear feature mapping, and finally outputs and caches the lesion feature tensor for subsequent cross-modal fusion.

[0032] Furthermore, the specific implementation process of step S300 is as follows: In step S300, the system initiates the anatomical topology reasoning branch in parallel, constructs a graph network structure with the preset physiological layers of the retina as graph network nodes and the anatomical adjacent position relationship between layers as edges, aggregates the prior initial features into the corresponding graph network nodes, uses the graph convolutional network to reason about the graph network structure, and outputs the topology correction vector used to constrain the relative physical spatial position of each layer of the retina.

[0033] In the specific underlying operational logic, the system's main control chip first initializes the graph data structure in memory. Based on prior knowledge in ophthalmology, the retina has a strict layered arrangement, and the system pre-sets the total number of physiological retinal layers to be [number missing]. Layers. The processor allocates memory contiguously in the heap at physical addresses. A data structure, which will... Each physiological layer is instantiated as a set of nodes in a graph network structure, arranged in the order of vertical depth from shallow to deep in the real biological space. Subsequently, the system instantiates a dimension as... A two-dimensional square matrix, serving as the adjacency matrix representing the anatomical adjacent positional relationships between layers. To translate biophysical barrier laws into computer-executable mathematical constraints, the processor's internal arithmetic logic unit (ALU) processes this adjacency matrix. Strict conditional addressing and assignment instructions are executed: the system traverses all row and column indices of the matrix. Execute absolute value condition judgment logic If and only if this condition is true (i.e., the first...) Layer and First When layers are in a strict vertical adjacency relationship in three-dimensional physical space, an assignment instruction is triggered. For all For cross-level or non-adjacent relationships, force assignment. Using this absolute value judgment logic, the system generates an extremely sparse tridiagonal matrix with rigorous anatomical properties. The processor generates the matrix. identity matrix of the same dimension And perform matrix addition to construct a self-loop adjacency matrix. This introduces a mechanism for preserving the physical characteristics of nodes in the graph topology.

[0034] After completing the low-level initialization of the topology graph structure, the system performs high-dimensional aggregation of the feature data. The system retrieves the prior initial feature matrix representing the physical properties of optical attenuation, which was output in step S100. For each preset set of spatial coordinates corresponding to a physiological layer, the system calls the graphics processor's local adaptive pooling operator to reduce the dimensionality of the discretely distributed light reflectivity gradient values ​​in the two-dimensional pixel space and aggregate them into a one-dimensional feature vector, which is then written into the corresponding graph node. In the data domain. The processor will handle all of them. The one-dimensional feature vectors of each graph node are concatenated along the feature dimension to generate an initial node feature matrix, denoted as . .

[0035] The system invokes the Graph Convolutional Neural Network (GCN) engine to perform forward physical spatial relational feature propagation on the graph network structure. Before entering each convolutional layer operation, the processor first performs the propagation based on the aforementioned self-loop adjacency matrix. Calculate the diagonal degree matrix The assignment rule for its main diagonal elements is to sum the row vectors: In the first Within the graph convolution propagation period of the layer, the tensor core computational unit of the system strictly executes the following spectral graph convolution propagation formula: ; In the aforementioned low-level instruction flow, the processor first calculates... Perform a symmetric normalization operation to generate a normalized Laplace operator. This operator is then used with the... Layer node feature matrix Perform large-scale matrix multiplication, and finally multiply by a learnable weight matrix optimized by backpropagation of the network. And via a nonlinear activation function After mapping, the data is stored in the next level register. From a physical, microscopic computational perspective, due to the matrix... The internal force was forcibly applied The zero-value truncation of the matrix multiplication operation completely severs the feature connectivity paths between non-adjacent layers at the mathematical algebraic level. This means that during the forward inference process of the network, any abnormal high-frequency gradient features generated by any local pathological lesion (such as cystic edema growing across layers) can only be attenuated and transmitted to physically adjacent layers according to the strict biological layering of the retina, and cannot cross layers to contaminate the feature expressions of other healthy layers.

[0036] After iterative execution of multi-layer graph convolution operations, the system outputs a final-layer feature tensor. This tensor contains topological information that integrates global anatomical arrangement patterns. The system extracts and caches this information from GPU memory as a topology correction vector for subsequent cross-modal calls by the fusion computation module.

[0037] Furthermore, the specific implementation process of step S400 is as follows: The system executes a dynamic cross-gating signal based on the probability of lesion presence in each region of the two-dimensional pathological heatmap. In a collaborative manner, the topology correction weight is adaptively increased as the probability of lesion presence increases. The lesion feature tensor and the topology correction vector are dynamically cross-gated and fused to obtain a high-dimensional fusion feature with lesion robustness and topological consistency.

[0038] In the specific implementation process, the system processor first extracts the two-dimensional pathological thermogram output by step S200. and define it in memory as having A probability distribution matrix of dimension 1. The processor iterates through this matrix to extract the coordinates of each pixel. The value at the location To achieve nonlinear adaptive adjustment of topological constraint strength, the system constructs a structure based on the natural constant in the arithmetic logic unit. The exponential activation function. Specifically, the system reads the preset positive gain hyperparameter. An adaptive topology-gated weight matrix is ​​generated by calculating the exponential mapping instruction. Its specific scalar calculation expression is as follows: Under this operational logic, when a local area is in a healthy state ( When the value approaches 0, the exponential term outputs the base gain; when a lesion is detected in a local area. At that time, due to the non-linear growth characteristics of the exponential function, the topological weight value at that coordinate will undergo a step amplification.

[0039] Simultaneously, the system obtains the topology correction vector output by the graph convolutional network in step S300. Since this vector initially belongs to a one-dimensional node feature set, its dimension is represented as... (in The number of physiological strata, The number of feature channels does not match the two-dimensional spatial dimension of the lesion feature tensor. Therefore, the system performs a spatial dimension broadcasting operation: first, based on the width of the input image... and height The vector is then tiled and copied along the axis to generate a topological space tensor with the exact same spatial resolution as the lesion feature tensor. .

[0040] After dimension alignment is completed, the system calls the hardware multiplier and uses the adaptive topology-gated weight matrix obtained from the previous calculation. For topological space tensors Perform pixel-by-pixel enhanced Hadamard product operations to calculate physically weighted topology correction features. Its algebraic expression is This step uses pixel-level gating switches to forcibly enhance the system's topological anatomical constraint signals in areas with severe lesion damage, thereby using the global structural prior of the graph network to compensate for the lack of local image features.

[0041] Finally, the system uses the aforementioned lesion feature tensor Topological correction features after physical weighting A cascaded operation is performed along the tensor channel dimension to construct a heterogeneous composite tensor that integrates local pathological perception and global topological constraints. Subsequently, the system inputs this composite tensor into a variable configured as follows: In the feature fusion layer of the convolutional kernel, deep compression and redundancy elimination of feature information are achieved by performing linear weighted combination and nonlinear activation mapping on cross-channel features, and finally outputting high-dimensional fused features, which are then passed to the decoder.

[0042] Furthermore, the specific implementation process of step S500 is as follows: The system executes the decoding of high-dimensional fused features through a decoder, outputting the spatial boundaries of each layer of the retina, and completing the image layer structure segmentation operation. This step is divided into the parameter optimization execution logic in the model training phase and the 3D morphology construction execution logic in the model inference phase.

[0043] In the specific implementation process, the system's graphics processor first inputs the high-dimensional fused features output from step S400 into the decoder network. The decoder network is configured with multiple transposed convolutional layers and upsampling operators in a cascaded configuration. The processor gradually recovers the two-dimensional spatial resolution of the feature tensor by performing two-dimensional discrete deconvolution operations, and finally outputs the physiological layers of the retina at various lateral positions through a pointwise classifier layer. The predicted depth coordinates.

[0044] During the model training phase, the processor's loss function calculation module is activated to perform iterative parameter optimization. To enforce that the network model adheres to realistic biological anatomy priors, the system is configured with an asymmetric topological penalty loss function for backpropagation optimization of network parameters. The specific underlying computer execution logic is as follows: The system defines the loss function calculation module along the A-scan physical scan direction along the longitudinal scan depth. retinal layer in lateral position The network predicted depth coordinates are , No. The network predicts the depth coordinates of the layer as follows: ,in The total number of layers in the retina. The Arithmetic Logic Unit (ALU) iterates through the predicted depth values ​​of adjacent physiological layers on the same horizontal coordinate, performs scalar subtraction, and calculates the physical depth difference between adjacent layers. .

[0045] The system calls the tensor computation unit to construct the asymmetric topological penalty loss function. Its calculus and algebraic expression is defined as follows: When the processor executes this formula, a non-linear activation mechanism with threshold truncation is introduced. Specifically, when the depth predicted by the network conforms to biological anatomy, that is, the upper layer coordinate values ​​are less than or equal to the lower layer coordinate values ​​(… )hour, The function instruction outputs a constant 0, causing the topology penalty calculation result at that spatial coordinate point to be cleared; once the processor detects a layer intersection phenomenon that violates the anatomical physical order in the predicted depth coordinates of adjacent layers, i.e. At that time, the floating-point coprocessor immediately intervenes and reads the preset penalty scaling hyperparameters. and execute with natural constant The system uses an exponential function mapping with a base of 0. Therefore, the processor generates a penalty gradient signal that explodes exponentially based on the cross-depth difference. The system calculates the basic cross-entropy similarity loss between the predicted segmentation probability map and the actual doctor's labeled data. Then, it is compared with the asymmetric topological penalty loss function. A linear weighted fusion is performed to construct the global joint loss. Finally, the partial derivatives of the global joint loss with respect to the network parameters are calculated using the chain rule, triggering the backpropagation mechanism. The internal weight parameters of the lesion-sensing encoder, graph convolutional network, and decoder are then updated in physical memory using the Adam algorithm for adaptive moment estimation.

[0046] During the model inference phase, when the model training converges and switches to deployment on actual medical devices, the system removes the loss function calculation module. The network directly outputs two-dimensional physical boundary coordinates during forward propagation. Subsequently, the system imports the two-dimensional segmentation boundary coordinate data of multiple consecutive frames of B-scan sections into the three-dimensional surface reconstruction algorithm module. The processor assembles discrete two-dimensional profile lines into a three-dimensional layered digital model of the retina by executing spatial interpolation and mesh rendering instructions. Because the network has been subject to strict anatomical and physical constraints by the asymmetric topological penalty loss function, the three-dimensional model rendered by the system is directly presented as a smooth, multi-layered surface entity without hierarchical spatial intersections and pathological breaks even in areas of severe edema. Finally, it is output to an external medical imaging display device for clinicians to perform three-dimensional morphological diagnosis.

[0047] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic layer structure segmentation method for ophthalmic OCT images, characterized in that, Includes the following steps: S100: Acquire an ophthalmic OCT image converted from the original physical interference signal collected by the optical coherence tomography (OCT) device, and calculate the light reflectance gradient of each pixel in the OCT image as a priori initial feature characterizing the physical properties of optical attenuation. S200. Input the OCT image into the lesion perception encoder to extract the lesion feature tensor of the image, and output a two-dimensional pathological heat map representing the probability of the presence of pathological lesions in the corresponding area of ​​the image through the lesion detection head in the lesion perception encoder. S300: Construct a graph network structure using preset retinal physiological layers as graph network nodes and inter-layer anatomical adjacent position relationships as edges. Aggregate the prior initial features into the corresponding graph network nodes. Use a graph convolutional network to infer the graph network structure and output a topology correction vector to constrain the relative physical spatial positions of each retinal layer. S400. Based on the lesion existence probability of each region in the two-dimensional pathological heat map, a dynamic cross-gating signal is constructed. In a collaborative manner, the topology correction weight is adaptively increased as the lesion existence probability increases. The lesion feature tensor and the topology correction vector are dynamically cross-gated and fused to obtain a high-dimensional fusion feature with lesion robustness and topology consistency. S500: The high-dimensional fusion features are decoded by the decoder to output the spatial boundaries of each layer of the retina, thus completing the image layer structure segmentation; wherein, during the model training phase, the lesion perception encoder, graph convolutional network and decoder use an asymmetric topological penalty loss function to iteratively optimize the network parameters. The asymmetric topological penalty loss function is configured as follows: when a layer crossing that violates the anatomical physical order is detected in the predicted depth coordinates of adjacent layers along the longitudinal scanning depth direction, an exponentially amplified penalty weight is applied based on the difference in the crossing depth of the adjacent layers; In step S100, the light reflectance gradient of each pixel in the OCT image is calculated, specifically including the following steps: A two-dimensional spatial coordinate system is established with the horizontal cross-sectional scan as the X-axis and the vertical depth scan (A-scan direction) as the Z-axis; the pixels with these coordinates are extracted from the OCT image, and their corresponding physical light reflection intensity values ​​are obtained. ; The initial longitudinal gradient is calculated along the longitudinal depth scanning direction using the center difference algorithm. The calculation formula is: Combining the exponential attenuation characteristics of the light beam within biological tissue in optical coherence tomography (OCT), a physical attenuation compensation coefficient is introduced. The optical reflectivity gradient after depth attenuation compensation is calculated. The calculation formula is: The light reflectance gradient after depth attenuation compensation The maximum and minimum values ​​are normalized to form the prior initial features that characterize the physical properties of optical attenuation. Before calculating the light reflectance gradient of each pixel in the OCT image, the method further includes: Adaptive nonlocal mean filtering is performed on the OCT image to denoise it. The filtering weights are calculated by utilizing the high similarity of the internal structure of the image, which reduces optical speckle noise while preserving the original interlayer physical reflection edge of the retina. The lesion detection encoder includes a multi-scale convolutional feature extraction network and a lightweight lesion auxiliary detection head. S200 specifically includes the following steps: The OCT image containing the prior initial features is input into the multi-scale convolutional feature extraction network to extract a multi-channel shallow feature tensor containing image edge gradients. ; The multi-channel shallow feature tensor The data is input into the lightweight lesion-assisted detection head, and sequentially undergoes pointwise convolution dimensionality reduction and nonlinear activation function mapping to output the two-dimensional pathological heatmap. ; Among them, the two-dimensional pathological thermogram The dimension is consistent with the two-dimensional spatial resolution of the OCT image, and the coordinates of each pixel in the image are... numerical value Between 0 and 1 The magnitude of the value positively represents the probability of the presence of destructive pathological lesions such as macular edema or vitreous warts in the corresponding spatial coordinate region; The lesion-aware encoder further includes a lesion-suppressing attention module, and the step of extracting the lesion feature tensor of the image specifically includes: The two-dimensional pathological thermogram The spatial attention mask signal is input into the lesion suppression attention module; Within the lesion suppression attention module, a negative activation function based on pathological probability is constructed, and the lesion suppression attention weight matrix is ​​calculated. The calculation formula is: in, To and A matrix of all one dimensions; Using the lesion-inhibiting attention weight matrix For the multi-channel shallow feature tensor Perform pixel-by-pixel weighted Hadamard product calculation, the formula is: The weighted tensor After mapping by a deep convolutional layer, the lesion feature tensor is output; through the pixel-by-pixel Hadamard product weighting operation, the lesion probability in the two-dimensional pathological heatmap is calculated. The higher the region, the lower the weight of the corresponding shallow image edge gradient features are forced, thereby achieving adaptive physical suppression of damaged pathological edge features.

2. The method for automatic layer structure segmentation of ophthalmic OCT images according to claim 1, characterized in that, In step S300, the construction of the graph network structure using preset retinal physiological layers as graph network nodes and interlayer anatomical adjacent position relationships as edges specifically includes the following steps: To the retina Each physiological layer, arranged according to the physical longitudinal depth from shallow to deep in the actual anatomical order, is defined as a set of nodes in a graph network structure. ; Based on the actual anatomical biological sequence, an adjacency matrix representing the anatomical adjacency relationships between layers is constructed. , where matrix elements The mathematical value selection rules are constrained as follows: If and only if the first Physiological Stratification and the First In ophthalmic anatomy, the physiological layers are strictly adjacent to each other, meaning they satisfy the absolute value condition. When, define ; For other inter-layer or non-adjacent inter-layer relationships, define ; Based on the adjacency matrix A self-connection mechanism is introduced to construct a physical location topology graph and calculate the self-loop adjacency matrix. ,in To and Identity matrices of the same dimension.

3. The method for automatic layer structure segmentation of ophthalmic OCT images according to claim 2, characterized in that, In step S300, the inference process performed on the graph network structure using a graph convolutional network to output a topological correction vector for constraining the relative physical spatial positions of each layer of the retina specifically includes the following calculation process: Using a local adaptive pooling algorithm, the prior initial features in each physiological layer region of the OCT image are extracted and aggregated to the corresponding nodes. In the middle, the splicing forms the initial node feature matrix. ; Obtain the self-loop adjacency matrix The calculated diagonal degree matrix The diagonal elements of the degree matrix satisfy... ; Forward propagation of physical space correlation features is performed through a graph convolutional neural network. The formula for calculating graph convolution propagation of a layer is: in, For the first The learnable weight matrix of a layered graph convolutional network. It is a non-linear activation function; After multi-layer graph convolutional propagation iterations, the output feature matrix of the last layer node is used as the topology correction vector; through the degree matrix Adjacency matrix with self-loop Regularized multiplication operations ensure that graph convolutional networks are strictly restricted to the anatomically adjacent layers during inference. Information is transmitted within the physical topological constraints.

4. The method for automatic layer structure segmentation of ophthalmic OCT images according to claim 3, characterized in that, In step S400, the method of constructing a dynamic cross-gating signal based on the lesion presence probability of each region in the two-dimensional pathological heatmap, and adaptively increasing the topology correction weight as the lesion presence probability increases, specifically includes the following calculation steps: Extract the two-dimensional pathological thermogram output by the lesion sensing encoder, denoted as where pixel coordinates The value at the location Characterizes the probability of the presence of local lesions, and ; Constructing an adaptive topology-gated weight matrix Using natural constants The exponential activation function maps the probability of the presence of the local lesion to a topological correction weight, and the calculation formula is as follows: in, These are preset positive gain hyperparameters; The exponential activation function generates a dynamic cross-gated signal, such that in healthy retinal areas, the probability of the presence of lesions increases. Approaching 0, the topology correction weights maintain the basic gain state; in the edematous region of the lesion, due to the probability of the lesion... Approaching 1 triggers a nonlinear mutation in the exponential activation function, causing the topology correction weight to be exponentially adaptively amplified, thereby forcibly introducing topological dissection priors to dominate the boundary reasoning of the region. The step of dynamically cross-gating and fusing the lesion feature tensor with the topology correction vector to obtain a high-dimensional fusion feature with lesion robustness and topological consistency specifically includes: Obtain the topology correction vector output by the graph convolutional network, perform a spatial dimension broadcast operation on it, and generate a tensor with the lesion feature. Topological space tensors with the same two-dimensional spatial resolution ; Using the adaptive topology gated weight matrix For the topological space tensor Perform pixel-by-pixel enhanced Hadamard product gating to obtain physically weighted topology correction features. The formula is: The lesion feature tensor Topological correction features after physical weighting Cascade along the channel dimension and input to Cross-channel feature dimensionality reduction and fusion are performed in the convolutional layer, and the final high-dimensional fused feature is output.

5. The method for automatic layer structure segmentation of ophthalmic OCT images according to claim 4, characterized in that, In step S500, the asymmetric topology penalty loss function is configured as follows: when a layer intersection violating the anatomical physical order is detected in the predicted depth coordinates of adjacent layers along the longitudinal scanning depth direction, an exponentially amplified penalty weight is applied based on the intersection depth difference of the adjacent layers. The specific calculation rules include: Set along the A-scan physical scan direction along the longitudinal scan depth, the first In the lateral position of the retina The network predicted depth coordinates are , No. The network predicts the depth coordinates of the layer as follows: ,in The total number of layers in the retina; Define the physical depth difference between adjacent layers as ; Construct the asymmetric topological penalty loss function Its mathematical formula is: in, The preset penalty scaling hyperparameter; In the backpropagation iteration of network parameters, when the predicted depth conforms to biological anatomy, that is... When the asymmetric topological penalty loss function is 0, the output value of the loss function is 0; when a physical space violation occurs... At that time, the network is affected by the base data The exponential asymmetric penalty forces the model to learn the correct anatomical boundaries in severely affected areas.

6. The method for automatic layer structure segmentation of ophthalmic OCT images according to claim 5, characterized in that, The process of decoding the high-dimensional fusion features using a decoder to output the spatial boundaries of each retinal layer and complete image layer structure segmentation specifically includes: During the model training phase, the basic similarity loss between the predicted segmentation probability map and the real doctor annotation labels is calculated. ; The basic similarity loss With the aforementioned asymmetric topological penalty loss function Linear joint operation is performed according to preset weights, and a global joint loss function is constructed to optimize the joint parameters of the lesion-sensing encoder, graph convolutional network and decoder end-to-end. During the model inference phase, the trained network outputs two-dimensional physical boundary coordinates, and through a three-dimensional surface reconstruction algorithm, the two-dimensional segmentation boundaries of multiple consecutive B-scan frames are assembled and rendered into a three-dimensional layered digital model of the retina without hierarchical intersections or pathological breaks.

7. An automatic layer structure segmentation system for ophthalmic OCT images, comprising the automatic layer structure segmentation method for ophthalmic OCT images according to any one of claims 1-6, characterized in that the system... include: The physical prior feature extraction module is used to acquire ophthalmic OCT images converted from the original physical interference signals collected by the optical coherence tomography (OCT) equipment, and to calculate the light reflectance gradient of each pixel in the OCT image as a prior initial feature characterizing the physical properties of optical attenuation. The lesion perception and extraction module is used to input the OCT image into the lesion perception encoder to extract the lesion feature tensor of the image, and output a two-dimensional pathological heat map representing the probability of the presence of pathological lesions in the corresponding area of ​​the image through the lesion detection head in the lesion perception encoder. The anatomical topology reasoning module is used to construct a graph network structure with preset retinal physiological layers as graph network nodes and the anatomical adjacent position relationship between layers as edges. The prior initial features are aggregated into the corresponding graph network nodes, and the graph network structure is reasoned using a graph convolutional network to output a topology correction vector used to constrain the relative physical spatial position of each retinal layer. The dynamic cross-gated fusion module is used to construct a dynamic cross-gated signal based on the lesion existence probability of each region in the two-dimensional pathological heat map. In a collaborative manner, the topology correction weight is adaptively increased as the lesion existence probability increases. The lesion feature tensor and the topology correction vector are dynamically cross-gated and fused to obtain a high-dimensional fusion feature with lesion robustness and topological consistency. The topology consistency decoding and optimization module is used to decode the high-dimensional fusion features through the decoder, output the spatial boundaries of each layer of the retina, and complete the image layer structure segmentation. In the model training phase, the lesion perception encoder, graph convolutional network and decoder in the system use an asymmetric topology penalty loss function to iteratively optimize the network parameters. The asymmetric topology penalty loss function is configured as follows: when a layer crossing that violates the anatomical physical order is detected in the predicted depth coordinates of adjacent layers along the longitudinal scanning depth direction, an exponentially amplified penalty weight is applied based on the difference in the crossing depth of the adjacent layers.

Images