A dual-flow stroke lesion segmentation method and system based on symmetric difference learning

Abstract

The present application belongs to the technical field of computer vision and medical image processing, and discloses a double-flow stroke lesion segmentation method and system based on symmetric difference learning. The present application performs intelligent instance enhancement processing on the input multi-modal brain magnetic resonance image; the processed multi-modal image is physically decoupled according to the modality semantics; a symmetric difference module SDM is embedded in each feature layer of the main stream encoder to generate a fine spatial symmetric difference mask, and the main stream feature is adaptively enhanced to output the feature; a cross-modality channel calibration module CMCC is embedded in each level of the jump connection between the main stream encoder and the decoder; the adaptive fusion feature is obtained; the voxel-level segmentation mask is generated by progressive upsampling; and the segmentation loss is calculated simultaneously at multiple scale levels of the decoder in the training stage, and the overall optimization target is the weighted sum of the losses at each level. The present application significantly reduces the risk of clinical missed diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and medical image processing technology, and in particular relates to a method and system for segmenting dual-flow stroke lesions based on symmetry difference learning. Background Technology

[0002] Ischemic stroke, as one of the leading causes of disability and death worldwide, has a long-term disability level closely related to the timeliness of intervention after onset. Guided by the clinical principle that "time is brain," rapidly and accurately defining the infarct core area is crucial for assessing tissue viability and developing individualized reperfusion therapy strategies. With the development of computer vision technology, deep learning-based automated tools have become an important means of assisting radiologists in decision-making. However, in clinical practice, due to the extremely high complexity of stroke lesions on multimodal magnetic resonance imaging (MRI), including diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC), and fluid-attenuated inversion recovery sequence (FLAIR), achieving robust and high-precision lesion segmentation still faces many challenges.

[0003] Stroke lesions exhibit extreme class imbalance and morphological heterogeneity in their spatial distribution, becoming a major bottleneck limiting the performance of existing models. Stroke lesions range in size from large cortical infarcts to extremely small lacunar lesions. For these small lesions, due to their extremely low proportion in the global space, detailed features are easily masked during continuous feature downsampling by conventional deep learning models, leading to significant missed diagnoses in clinical practice. Currently, there are two main mitigation strategies for this problem: one is to introduce hard sample mining or composite loss functions (such as an improved FocalLoss combined with boundary awareness), but these methods only passively adjust during the gradient update stage and do not directly increase the signal-to-noise ratio of small targets at the original feature extraction level; the other is to use generative models (such as diffusion models) to synthesize artifact data with lesions to expand the training set, but such generative augmentation methods are extremely computationally expensive, and the generated lesions often lack explicit constraints on their physical volume and anatomical rationality, making it difficult to stably increase the distribution density of small lesions.

[0004] Secondly, efficient integration and decoupling of multimodal features remains a core challenge in current medical imaging research. Different MRI sequences have significant physical meanings and diagnostic complementarity: DWI and ADC can sensitively capture acute cytotoxic edema, while FLAIR provides crucial structural background, helping to differentiate between acute infarction and old white matter lesions. Most current models employ an early fusion channel stitching strategy. This static mapping strategy assumes equal contributions from each modality, but in complex stroke scenarios, artifacts in the FLAIR modality can easily contaminate DWI features that are highly sensitive to acute cytotoxic edema. Although some cutting-edge research has introduced cross-modal attention modules (such as Transformer-based cross-modal interactions) to attempt to learn modal relevance, the complex calculation of the attention matrix significantly increases the number of model parameters and inference latency, failing to meet the high timeliness requirements of clinical practice. More importantly, most existing methods only calculate feature similarity at a mathematical level, failing to explicitly physically isolate and semantically decouple the lesion-sensitive modality from the structural background modality from a clinical diagnostic perspective.

[0005] Existing algorithms treat brain images as ordinary 3D voxel arrays, generally ignoring the natural axial symmetry constraints inherent in human neuroanatomy. When interpreting stroke images, radiologists typically compare the affected hemisphere with the healthy contralateral hemisphere; this bilateral difference is crucial prior knowledge for identifying hidden lesions. To introduce symmetry constraints, current techniques primarily employ two approaches: one is non-rigid registration based on image space preprocessing, which heavily relies on registration accuracy and suffers from high inference latency; the other uses a Siamese Networks bi-branch architecture or a globally symmetric attention architecture. However, these implicit modeling methods within the feature space drastically and disruptively alter the network's fundamental topology, leading to a dramatic increase in the number of parameters and significant difficulties in seamlessly integrating with current highly integrated adaptive baseline frameworks, often requiring the sacrifice of the framework's original deep supervision or dynamic tuning capabilities. Summary of the Invention

[0006] To overcome the problems existing in related technologies, the present invention discloses a method and system for segmenting dual-flow stroke lesions based on symmetry difference learning, specifically involving a method for segmenting ischemic stroke lesions based on multimodal magnetic resonance imaging (MRI). This addresses the three main problems in existing technologies: high missed diagnosis rate of small lesions, multimodal noise interference, and lack of prior knowledge of anatomical symmetry, achieving higher accuracy and stronger robustness in automatic stroke lesion segmentation.

[0007] The technical solution is as follows: a dual-flow stroke lesion segmentation method based on symmetric difference learning, comprising the following steps: S1. Perform intelligent instance enhancement processing on the input multimodal brain magnetic resonance images; S2. Decouple the processed multimodal images physically according to modal semantics; S3. Embed the Symmetric Difference Module (SDM) in each feature layer of the mainstream encoder to generate a refined spatial symmetric difference mask, perform adaptive enhancement on the mainstream features, and output the features. S4. At each jump connection point between the mainstream encoder and decoder, a cross-modal channel calibration module (CMCC) is embedded to obtain adaptive fusion features. S5. The adaptive fusion features at each level are fed into a unified 3D decoder and upsampled level by level to generate voxel-level segmentation masks. During the training phase, the segmentation loss is calculated simultaneously at multiple scale levels of the decoder. The overall optimization objective is the weighted sum of the losses at each level.

[0008] In step S1, performing intelligent instance enhancement processing on the input multimodal brain magnetic resonance images includes: S1.1 Data Acquisition and Preprocessing: Acquire the brain magnetic resonance imaging sequence data to be segmented, store it in three-dimensional voxel format, and denote it as the input tensor. ,in, These correspond to three modes: diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC), and liquid attenuation inversion recovery sequence (FLAIR). These represent the height, width, and depth in three-dimensional space, respectively. Given the set of real numbers, input X is a number that exists in... A four-dimensional real tensor in a continuous space, where every voxel eigenvalue belongs to the real number field.

[0009] S1.2, Intelligent Instance Enhancement Processing (SCP): Lesion Masking of Input Images Performing 3D connected component labeling operations using a 3×3×3 cube structuring element: ; In the formula, For the labeled voxel matrix, different instances are assigned independent integer labels. This represents the total number of independent lesion cases. This is a connected component labeling function, representing the execution of a three-dimensional connected component labeling operation; This is the lesion mask for the input image. It is a 3×3×3 cube structural element; For any instance Physical volume Defined as: ; In the formula, In three-dimensional voxel coordinates, For indicator functions, Independent positive integer labels for different instances; Constructing volume-constrained candidate lesion pools : ; in, voxels, voxel, trigger probability set to ; When the enhancement operation is triggered, the candidate lesion pool is constrained by volume. Randomly selected target instances are used to construct a brain parenchyma safe zone using the physical properties of ADC and FLAIR sequences; for arbitrary modalities voxels The local adaptive cross threshold at a given location, and the dynamic threshold calculation, are as follows: ; In the formula, voxels Local dynamic threshold at that location These represent the signal mean and standard deviation within a local three-dimensional sliding window centered on the voxel, respectively. This is a parameter for adjusting contrast sensitivity. This is the dynamic range constant of the image signal intensity; After obtaining the dynamic threshold matrix, morphological erosion operations are used to shrink the physical boundaries of the safe zone, resulting in the final brain parenchyma safe zone mask. The generation process is defined as follows: ; In the formula, and For the three-dimensional feature tensor of the corresponding mode, and This is the adaptive threshold matrix calculated based on the mean and variance of the local sliding window. Represents voxel-level logical intersection operation. This indicates a morphological erosion operation performed using three-dimensional structuring elements. For local masks Perform cross-modal synchronous cropping and generate fused images through a mask-guided voxel replacement mechanism. With corresponding tags : ; ; In the formula, For the target image, This is a binary mask for cropping candidate microlesion instances and their local regions. The source lesion image contains candidate lesion instances. The original segmentation label corresponding to the candidate lesion instance. ⊙ represents the native segmentation label of the target image, and ⊙ represents the Hadamard product.

[0010] In step S2, the physical decoupling of the processed multimodal image according to modal semantics includes: The main stream containing DWI and ADC sequences and the auxiliary stream containing FLAIR sequences are fed into their respective independent parallel 3D encoder paths to extract multi-scale feature pyramids step by step. and The three-dimensional encoders are all composed of four feature extraction stages and a bottom bottleneck layer stacked together; in any given... Each feature extraction stage Representing the network depth-level index, both the mainstream and auxiliary streams employ two consecutive 3D convolutional modules for feature encoding; among them, For mainstream 3D encoders in the first Mainstream feature maps output at each scale level For the auxiliary flow three-dimensional encoder in the first Auxiliary flow feature maps output at each scale level; The data flow of a single convolutional module passes through the following layers in sequence: a 3×3×3 three-dimensional convolutional layer for extracting local spatial and pathological / anatomical features; a three-dimensional instance normalization layer to accelerate network convergence and reduce data bias caused by differences in different MRI equipment and individual patients; and a LeakyReLU activation function layer to introduce non-linear expressive power and maintain the stability of feature propagation; each encoder layer increases the number of feature channels step by step while performing feature downsampling. After the above-mentioned step-by-step independent encoding, the main stream and auxiliary stream respectively output multi-scale feature pyramids corresponding to different spatial resolutions and abstract semantic levels. and ,in, Representing the first The channel number, height, width, and depth of the level features provide physically isolated feature bases for cross-modal channel calibration fusion.

[0011] In step S3, a refined spatial symmetric difference mask is generated, and adaptive enhancement is performed on the mainstream features. The output features include: embedding a symmetric difference module (SDM) in each feature layer of the mainstream encoder; performing a mirror flip along the sagittal plane on the mainstream features; generating an original difference map by calculating the tensor absolute difference between the original feature map and the mirror reference feature map; and generating a refined spatial symmetric difference mask through a lightweight convolutional bottleneck layer. Adaptive enhancement is applied to the mainstream features, and the output features are then optimized. Defined as: ; In the formula, For activation function, In the first The final adaptive fusion feature output by the cross-modal channel calibration module CMCC at each scale level.

[0012] Generate refined spatially symmetric difference mask include: The Symmetric Difference Module (SDM) first moves along the sagittal plane to the main feature map. Perform mirror flip operation Construct mirror reference features: ; In the formula, For mirror reference feature map; Calculate the absolute difference between the original feature map and its mirror image to generate the original difference map. ; In the formula, This is the original symmetrical difference diagram; Will The input consists of a lightweight convolutional bottleneck layer composed of two 1×1×1 convolutional layers and a ReLU activation function, which is then mapped by a Sigmoid function to generate a refined spatially symmetric difference mask. : ; In the formula, and These are the weight matrices for dimensionality reduction and dimensionality increase convolutions, respectively. It is the ReLU activation function. This is the Sigmoid activation function.

[0013] In step S4, the cross-modal channel calibration module CMCC is embedded to obtain adaptive fusion features, including: Will and After concatenation along the channel dimension, cross-modal context information is extracted using global average pooling, and auxiliary flow channel attention weights are generated through a linear bottleneck structure. After applying weights to the auxiliary flow features, the residuals are superimposed with the mainstream features to obtain adaptive fusion features: ; In the formula, For the first The adaptive fusion feature map is generated after each network layer is calibrated and fused by CMCC. To add element by element, For the first Auxiliary stream encoder in each network layer Extracted feature maps, The channel attention weights are dynamically generated by the CMCC module based on the global context. These are the channel-level gating weights learned adaptively by the network.

[0014] Furthermore, attention weights for the auxiliary flow channels are generated. Including: for the first Layer feature map, and First, the data is spliced ​​together along the channel dimension. The global context vector is extracted through the global average pooling (GAP) operation: ; In the formula, For global context vectors, For the average pooling operation across the entire series, Let it be a 2C-dimensional real vector space. For the first Feature maps spliced ​​from each network layer; channel compression ratio after hidden layers Linear bottleneck structure generates auxiliary flow channel attention weights : ; In the formula, and These are the learnable weight parameter matrices for the dimensionality-reduced and dimensionality-increasing convolutional layers, respectively. For the generated channel attention mask.

[0015] In step S5, the overall optimization objective is... The weighted sum of losses at each level: ; ; In the formula, To optimize overall losses, For cross-entropy loss, For Dice similarity coefficient loss, This represents the total number of output layers for the decoder. The weights used for deep supervision decay exponentially with decreasing resolution. It is the gold standard label for true voxel-level data. For scale-level indexing, Let be the segmentation loss calculated at the l-th resolution level of the decoder. This represents the predicted probability of the network.

[0016] Another objective of this invention is to provide a dual-flow stroke lesion segmentation system based on symmetry difference learning, the system comprising: The intelligent instance enhancement processing module is used to perform intelligent instance enhancement processing on the input multimodal brain magnetic resonance images; The physical decoupling module is used to physically decouple the processed multimodal images according to modal semantics; The symmetric difference module is used to embed in the feature layers of the mainstream encoder at each level to generate a fine-grained spatial symmetric difference mask, perform adaptive enhancement on the mainstream features, and output features; A cross-modal channel calibration module is used to embed adaptive fusion features at various jump connections between the main encoder and decoder. The weighted summation module for the loss at each level is used to feed the adaptive fusion features of each level into a unified 3D decoder and upsample them level by level to generate voxel-level segmentation masks. During the training phase, the segmentation loss is calculated simultaneously at multiple scale levels of the decoder, and the overall optimization objective is the weighted summation of the loss at each level.

[0017] Furthermore, the dual-flow stroke lesion segmentation system based on symmetry difference learning is mounted on a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can realize the functions of the dual-flow stroke lesion segmentation system based on symmetry difference learning.

[0018] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention effectively solves the problem of high missed diagnosis rates for minute lesions. Existing technologies passively compensate for minute lesions at the loss function or generative model level, resulting in high computational costs and a lack of explicit constraints on lesion volume and anatomical rationality. This invention actively identifies easily missed minute infarct instances through the SCP strategy, dynamically increasing their spatial distribution density in training batches from the data source with extremely low preprocessing computational costs, forcing the encoder to establish perceptual sensitivity to small-scale targets at a shallow level. On the ISLES2022 dataset, the overall recall rate reached 86.45%, and the recall rate of minute lesions with a volume of less than 5 mL was significantly improved from the baseline of 52.30% to 76.85%, significantly reducing the risk of clinical missed diagnosis.

[0019] This invention effectively solves the false positive problem caused by multimodal noise interference. Existing technologies often employ early channel splicing strategies, statically assuming equal contributions from each modality, leading to non-specific high-signal artifacts in FLAIR sequences that contaminate DWI features highly sensitive to acute infarction. In step S2, this invention physically isolates the lesion-sensitive main flow from the structural background auxiliary flow. In step S4, the CMCC module adaptively adjusts the proportion of auxiliary flow information incorporated, significantly reducing the absolute volumetric error (AVD) from 21.52% of the baseline to 11.45% with only 32.6M parameters.

[0020] This invention effectively solves the problem of false positives at the boundary caused by the lack of anatomical symmetry priors. Existing technologies based on Siamese networks or global symmetry attention require significant changes to the basic network topology, resulting in a dramatic increase in the number of parameters and incompatibility with adaptive frameworks. In step S3, this invention explicitly utilizes the anatomical symmetry prior of the human brain in a non-invasive manner through the SDM module, endowing the model with the diagnostic logic of "bilateral comparison," effectively filtering out false positive artifacts caused by symmetrical physiological structures such as the ventricular margins. On the ISLES2022 dataset, HD95 is reduced to 4.33 mm, and on the ATLASv2.0 unimodal dataset, HD95 reaches 22.75 mm, both outperforming all existing comparison methods, demonstrating the cross-dataset generalization robustness of this mechanism.

[0021] Secondly, in terms of clinical application and commercialization, this invention addresses the extremely high time sensitivity required in the diagnosis and treatment of acute stroke, where time is of the essence. It constructs a highly efficient automated segmentation architecture. This solution can accurately generate voxel-level lesion segmentation masks within seconds, greatly reducing the workload of radiologists who traditionally manually delineate lesions and significantly lowering the rate of missed and misdiagnosed small lesions due to fatigue. Simultaneously, due to the extremely low computational overhead of the intelligent feature modules embedded in the network, this technology possesses extremely high hardware and software compatibility, enabling seamless integration into existing hospital PACS systems or MRI post-processing workstations, forming an integrated intelligent diagnostic and treatment product from image acquisition to quantitative assessment.

[0022] Third, at the technical level, traditional multimodal segmentation methods mostly treat different MRI sequences as homogeneous channels and directly mix them, ignoring the unique physical imaging mechanisms of each sequence. This invention innovatively divides multimodal data into a mainstream highly sensitive to lesions (DWI, ADC) and an auxiliary stream providing structural background (FLAIR), achieving deep decoupling at the physical feature level. This invention introduces a sagittal mirror flip and feature difference (SDM) mechanism, explicitly incorporating the brain's natural anatomical symmetry prior into the neural network.

[0023] Fourth, this invention successfully overcomes two core challenges that have long plagued the field: accurate segmentation of minute lesions in complex anatomical contexts, and semantic interference during multimodal feature fusion. Because minute lesions account for a very small percentage of the total voxels, traditional networks easily lose their detailed features during stepwise downsampling. By using three-dimensional connected component analysis and volume constraints to screen target lesions, and utilizing morphologically eroded brain parenchyma security region masks to achieve accurate intelligent instance enhancement (SCP), coupled with deep multi-scale supervision and joint loss functions (Dice loss and cross-entropy loss) at the decoder end, the problem of insufficient representation ability for minute lesions is completely solved. Regarding the signal dilution problem during multimodal fusion, a cross-modal channel calibration module (CMCC) is used to adaptively generate auxiliary channel weights based on global context, performing a coarse-to-fine filtering and fusion of background features, effectively eliminating semantic conflicts caused by forced splicing of heterogeneous modalities.

[0024] Fifth, this invention overcomes the bias that general data augmentation methods can be directly and indiscriminately applied to medical images. Conventional random cut-and-paste or mixed augmentation can easily produce pseudo-lesions in the ventricular fluid or outside the skull that violate physiological common sense. This invention, however, defines safe areas of brain parenchyma through connected components and morphological constraints, ensuring amplification diversity while upholding anatomical legitimacy. This invention also overcomes the technical bias that early fusion of multimodal data at the input end yields the best results. Through reverse thinking, this invention demonstrates that blind fusion can actually cause mutual masking of features between sequences; by employing a strategy of early physical decoupling extraction and later adaptive directional calibration, it can significantly improve the network's ability to identify complex lesions. Attached Figure Description

[0025] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the disclosure of this invention and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a flowchart of the dual-flow stroke lesion segmentation method based on symmetry difference learning provided by the present invention; Figure 2 A schematic diagram illustrating the principle of the dual-flow stroke lesion segmentation method based on symmetry difference learning provided by this invention; Figure 3 This is a diagram of the overall architecture of the SD-DualNet network of the present invention; Figure 4 This is a single-stage data flow diagram of the dual-stream fusion network of the present invention, including the independent feature extraction paths of the main stream and auxiliary streams and the detailed internal structure of the cross-modal channel calibration module (CMCCBlock); Figure 5 This is a schematic diagram of the symmetric difference module (SDM) of the present invention; Figure 6The images show a qualitative comparison of ablation experiments of this invention on the ISLES2022 dataset, where (a) represents a typical scenario with small lesions and (b) represents a typical scenario with large-area complex boundary lesions.

[0026] Figure 7 The image shows a comparison of the visualization segmentation results of the present invention and existing representative methods in three typical clinical scenarios. (a) is a scenario of tiny lesions with a high risk of being missed, (b) is a scenario of multiple lesions with scattered boundaries, and (c) is a scenario of large-area lesions with complex morphology and easily affected by ventricular artifacts. Detailed Implementation

[0027] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0028] To address the issues of multimodal feature interference in stroke lesions and the tendency to miss small lesions, the innovations of this invention are as follows: First, this invention utilizes three-dimensional connected domains and morphological constraints to construct a brain parenchyma-safe region, achieving anatomically compliant intelligent instance enhancement (SCP). Second, multimodal MRI is decoupled according to physical imaging semantics into lesion-sensitive mainstream (DWI / ADC) and anatomical background auxiliary stream (FLAIR), and a dual-stream three-dimensional encoder path is constructed. Based on this, a symmetry difference module (SDM) is introduced into the mainstream path, explicitly incorporating brain symmetry priors through sagittal plane mirror flipping and feature difference calculation. Simultaneously, a cross-modal channel calibration module (CMCC) is embedded at skip connections, adaptively weighting auxiliary stream features based on global context and performing residual fusion with the mainstream. Finally, by combining a unified three-dimensional decoder with a multi-scale deep supervision joint loss function, voxel-level precise segmentation of multi-scale and small stroke lesions is achieved.

[0029] Example 1: As Figure 1 As shown, the dual-flow stroke lesion segmentation method based on symmetric difference learning provided in this embodiment of the invention includes: S1. Perform intelligent instance enhancement processing on the input multimodal brain magnetic resonance images; For example, input multimodal brain magnetic resonance images ( The system performs Smart Copy-Paste (SCP) processing on three modalities (DWI, ADC, and FLAIR, respectively), including: resolving lesion masks using three-dimensional connected component analysis, accurately screening easily missed small lesion instances through a volume-constrained candidate pool, and constructing brain parenchyma safe region masks using an adaptive cross-thresholding method for multimodal sequences. After determining the legal spatial anchor points, a mask-guided voxel replacement mechanism is used to generate fused images and corresponding labels. and : ; ; In the formula, For the target image, This is a binary mask for cropping candidate microlesion instances and their local regions. The source lesion image contains candidate lesion instances. The original segmentation label corresponding to the candidate lesion instance. ⊙ represents the native segmentation label of the target image, and ⊙ represents the Hadamard product.

[0030] Corresponding tags and While employing a common mask fusion operation in its basic mathematical form, the technological innovations in the specific application of this invention are reflected in the following two points: (1) The constraints of fusion are different (overcoming the anatomical damage problem of existing technologies). Existing copy-paste technology is usually blind and random pasting. If used directly for medical images, it is very easy to paste lesions into the ventricles, skull, or even outside the brain, producing artifacts that violate the laws of physics. However, the candidate mask involved in the calculation in this invention and target area It is the brain parenchyma security zone masking from the previous step. Those that are allowed to be fused only after strict screening and physical volume constraints represent intelligent fusion results that meet the premise of anatomical safety.

[0031] (2) Existing technologies are usually designed to increase the overall data volume; while the present invention incorporates volume constraints. arrive This method specifically addresses the clinical pain point of easily missing small infarcts by targeting and increasing the distribution density of small lesions at the data source. It is an indispensable final execution step in the Intelligent Instance Enhancement (SCP) strategy of this invention and is a prerequisite based on safe zone constraints.

[0032] S2. Decouple the processed multimodal images physically according to modal semantics; For example, the main stream containing DWI and ADC sequences and the auxiliary stream containing FLAIR sequences are fed into their respective independent parallel 3D encoder paths to extract multi-scale feature pyramids step by step. and Specifically, each of the three-dimensional encoders consists of four feature extraction stages (Stage 1 to Stage 4) stacked together, plus a bottleneck layer at the bottom. In any given stage... Each feature extraction stage ( (Representing the network depth level index), both the main and auxiliary streams use two consecutive 3D convolutional modules for feature encoding. The data flow of a single convolutional module passes through: one... A three-dimensional convolutional layer (Conv3D) is used to extract local spatial and pathological / anatomical features; a three-dimensional instance normalization layer (InstanceNorm) is used to accelerate network convergence and mitigate data bias caused by differences in different MRI equipment and individual patient differences; and a LeakyReLU activation function layer is used to introduce non-linear expressive power and maintain the stability of feature propagation. As the network depth increases, each encoder level progressively increases the number of feature channels while performing feature downsampling; the number of channels is set to 32, 64, 128, 256, and 320 from shallow to deep. After the above progressively independent encoding, the main stream and auxiliary stream output multi-scale feature pyramids corresponding to different spatial resolutions and abstract semantic levels. and (in, Representing the first The number of channels, height, width, and depth of the features are defined, thus providing a high-quality and physically isolated feature base for subsequent cross-modal channel calibration fusion.

[0033] S3. Embed the Symmetric Difference Module (SDM) in each feature layer of the mainstream encoder to generate a refined spatial symmetric difference mask, perform adaptive enhancement on the mainstream features, and output the features. Symmetrical Discrepancy Modules (SDMs) are embedded in the feature layers of the mainstream encoder at each level. Mirroring of the mainstream features is performed along the sagittal plane. The original difference map is generated by calculating the tensor absolute difference between the original feature map and the mirror reference feature map. This original difference map is then passed through a lightweight convolutional bottleneck layer to generate a refined spatially symmetric difference mask. Adaptive enhancement is applied to the mainstream features, and the output features are then optimized. Defined as: ; in, Activation function; the SDM achieves non-intrusive integration based on a dynamic feature interception mechanism using a forward hook function, without making any static modifications to the underlying topology of the three-dimensional encoder-decoder network. In the specific application scenario of this invention, the physical mechanism and mask represent... The generation and application logic is an innovative concept proposed in this invention.

[0034] Existing attention mechanisms typically generate masks based solely on the spatial dependencies of a single feature map. The Spatial Symmetry Difference Mask (SDM) is generated by the Symmetry Difference Module (SDM) proposed for the first time in this invention. It extracts the mirror inconsistency between the affected and healthy hemispheres. It explicitly transforms the medical anatomical symmetry prior into a mathematical mask and applies it to the mainstream pathological feature flow. The technical advantage lies in: giving the model a doctor-like bilateral contrast perspective, assigning high activation weights only to pathological signals that disrupt anatomical symmetry, thereby effectively suppressing false-positive interference from naturally symmetrical physiological structures (such as the normal lateral ventricle margin and high signal intensity in white matter).

[0035] S4. At each jump connection between the mainstream encoder and decoder, a cross-modal channel calibration module (CMCC) is embedded to obtain adaptive fusion features. Will and After concatenation along the channel dimension, cross-modal context information is extracted using global average pooling, and auxiliary flow channel attention weights are generated through a linear bottleneck structure. After applying weights to the auxiliary flow features, the residuals are superimposed with the mainstream features to obtain adaptive fusion features: ; In the formula, For the first The adaptive fusion feature map is generated after each network layer is calibrated and fused by CMCC. To add element by element, For the first Auxiliary stream encoder in each network layer Extracted feature maps, The channel attention weights are dynamically generated by the CMCC module based on the global context. These are the channel-level gating weights learned adaptively by the network.

[0036] In the multimodal stroke dual-stream segmentation network, the adaptive fusion feature formula, used as the feature fusion formula for cross-modal channel calibration (CMCC), completely solves the problem of indiscriminate propagation of intermodal noise caused by traditional multimodal early channel fusion. It achieves efficient complementarity and noise resistance of multimodal features. Specifically, its technical role is reflected in: dynamic gating and noise interception: the weight w acts as a dynamic valve for the fusion of the FLAIR modality into the mainstream (DWI / ADC modality). When the FLAIR modality extracts old lesions or artifact noise, the network adaptively outputs extremely low weights. This approach mathematically truncates the auxiliary flow output, protecting the lesion-sensitive mainstream features from contamination. Adaptive incorporation of anatomical boundary information: weights are strongly activated when the auxiliary flow provides a clear and undisturbed outline of brain structures. The structured high-frequency priors are fully preserved and incorporated into the mainstream through residual addition, which helps the model to refine the lesion edge.

[0037] S5. Adaptive fusion of the output features at each level. The data is fed into a unified 3D decoder, and voxel-level segmentation masks are generated through progressive upsampling. During the training phase, the segmentation loss is calculated simultaneously at multiple scale levels of the decoder to optimize the overall target. This is the weighted sum of losses at each level.

[0038] The output of adaptive fusion features at each level The data is fed into a unified 3D decoder and upsampled step by step to generate voxel-level segmentation masks. During the training phase, the segmentation loss is calculated simultaneously at multiple scale levels of the decoder to optimize the overall objective. The weighted sum of losses at each level: ; ; In the formula, To optimize overall losses, For cross-entropy loss, For Dice similarity coefficient loss, This represents the total number of decoder output levels. The weights used for deep supervision decay exponentially with decreasing resolution. It is the gold standard label for true voxel-level data. For scale-level indexing, Let be the segmentation loss calculated at the l-th resolution level of the decoder. This represents the predicted probability of the network.

[0039] The technical application of (composite loss function): cross-entropy loss Applying independent classification penalties at the voxel level helps stabilize gradients in the early stages of model training; while Dice loss... The volume overlap of the mask is measured from a global perspective. Since small stroke lesions account for a very small proportion of the entire image (due to extreme class imbalance), relying solely on CE loss will cause the lesion gradient to be submerged by the background. Combining it with Dice loss can effectively overcome this problem. The technical role of (deep supervision): Because this invention employs a deep two-stream encoder-decoder network, gradients are prone to vanishing during backpropagation. By calculating the segmentation loss at multiple different scale levels of the decoder and distributing weights exponentially decreasing with resolution, (…). This forces the intermediate hidden layers of the network to also learn highly discriminative semantic features, accelerating network convergence and improving the final segmentation accuracy.

[0040] Example 2: The dual-flow stroke lesion segmentation system based on symmetric difference learning provided in this embodiment of the invention includes: The intelligent instance enhancement processing module is used to perform intelligent instance enhancement processing on the input multimodal brain magnetic resonance images; The physical decoupling module is used to physically decouple the processed multimodal images according to modal semantics; The symmetric difference module is used to embed in the feature layers of the mainstream encoder at each level to generate a fine-grained spatial symmetric difference mask, perform adaptive enhancement on the mainstream features, and output features; A cross-modal channel calibration module is used to embed adaptive fusion features at various jump connections between the main encoder and decoder. The weighted summation module for the loss at each level is used to feed the adaptive fusion features of each level into a unified 3D decoder and upsample them level by level to generate voxel-level segmentation masks. During the training phase, the segmentation loss is calculated simultaneously at multiple scale levels of the decoder, and the overall optimization objective is the weighted summation of the loss at each level. Example 3: As another embodiment of the present invention, the dual-flow stroke lesion segmentation method based on symmetric difference learning provided in this embodiment of the present invention, such as... Figure 2 It includes six stages: data preprocessing, intelligent instance enhancement, dual-stream feature encoding, cross-modal channel calibration fusion, feature decoding and segmentation output, and loss function calculation and model training.

[0041] Step 1: Data Acquisition and Preprocessing. Acquire the brain MRI image sequence data to be segmented, store it in three-dimensional voxel format, and denote it as the input tensor. Where C=3, corresponding to the three modes of diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC), and liquid attenuation inversion recovery sequence (FLAIR), respectively; H, W, and D are the height, width, and depth of the three-dimensional space, respectively. Given the set of real numbers, input X is a number that exists in... A four-dimensional real tensor in a continuous space, where every voxel eigenvalue belongs to the real number field.

[0042] Data preprocessing includes the following operations: using the DWI sequence as a fixed reference template, high-precision rigid resampling is performed on the FLAIR image using the third-order B-spline interpolation algorithm to ensure the physical alignment of the three input channels at the voxel level; bounding box cropping is performed based on the self-configured preprocessing pipeline of the nnU-Netv2 framework to eliminate background areas with no diagnostic value; and modality-specific Z-score normalization is performed on the three input modalities respectively.

[0043] Step 2: Intelligent Instance Augmentation (SCP). To address the issue of extremely uneven distribution of stroke lesion volume, intelligent instance augmentation is performed on the input image. This involves masking the lesions in the input image. Performing 3D connected component labeling operations using a 3×3×3 cube structuring element: ; In the formula, For the labeled voxel matrix, different instances are assigned independent integer labels. This represents the total number of instances with independent lesions. For any instance... Its physical volume Defined as: ; in, In three-dimensional voxel coordinates, This is an indicator function.

[0044] Constructing volume-constrained candidate lesion pools : ; In the formula, voxels, The voxel is set to trigger probability p=0.5.

[0045] This invention is specifically designed for anchoring small lesions that are easily missed during diagnosis. Volume-constrained candidate pool / lesion pool This condition is set. and This is a targeted screening logic specifically designed for this task, taking into account the actual physical volume distribution of small infarct foci (lacunar infarcts) in clinical practice. When the enhancement operation is triggered, target instances are randomly selected from the candidate pool P, and a safe zone in the brain parenchyma is constructed using the physical characteristics of the ADC and FLAIR sequences. This applies to any modality. voxels The local adaptive cross threshold (dynamic threshold) at the point is calculated as follows: ; This invention leverages the robustness of an adaptive cross-threshold formula to local signal heterogeneity, specifically designed to address contrast inconsistencies in MRI images. It is a concrete mathematical model for generating a dynamic threshold matrix, defining an adaptive threshold calculation method for individual voxels in the image. By traversing the entire 3D brain image and calculating for each voxel, a set of local thresholds with the same spatial dimensions as the original input image is obtained. This set of thresholds, arranged according to the original spatial coordinates, constitutes the dynamic threshold matrix. The dynamic threshold matrix is ​​a spatially adaptive intermediate feature map (or parameter matrix), representing a segmentation boundary reference standard tailored to a specific input image.

[0046] voxels Local dynamic threshold at that location and These represent the signal mean and standard deviation within a local three-dimensional sliding window centered on the voxel, respectively. This is a parameter for adjusting contrast sensitivity. This represents the dynamic range constant of the image signal intensity. After obtaining the dynamic threshold matrix, morphological erosion operations are used to shrink the physical boundaries of the safe zone. This innovative invention yields the final brain parenchyma safe zone mask. The generation process is defined as follows: ; In the formula, and For the three-dimensional feature tensor of the corresponding mode, and This is the adaptive threshold matrix calculated based on the mean and variance of the local sliding window. Represents voxel-level logical intersection operation. This indicates a morphological erosion operation performed using three-dimensional structuring elements.

[0047] The three-dimensional structural elements For example, a size of A three-dimensional cubic or spherical structure matrix of individual elements. Preferably, the specific size of this structural element can be configured according to the spatial resolution of the input magnetic resonance image and a preset boundary safety margin; the purpose is to forcibly shrink the physical boundary of the brain parenchyma safety region mask to ensure that the subsequently mapped lesion instances maintain a sufficient safe physical distance from the cerebrospinal fluid at the ventricular margin or in the sulci, preventing semantic confusion in the model. Only when the randomly generated three-dimensional offset vector To ensure that the candidate lesion falls completely into The spatial anchor point is deemed valid when it is internal and maintains a safe distance from the primary lesion. In a specific embodiment, the minimum spacing of the safe distance is exemplarily set to 3 to 5 voxels (e.g., after the image is resampled to...). In scenes with equal voxel resolution, the corresponding real physical spacing is 3 to 5 millimeters. (Regarding local masks...) Perform cross-modal synchronous cropping and generate fused images through a mask-guided voxel replacement mechanism. With corresponding tags : ; ; In the formula, For the target image, This is a binary mask for cropping candidate microlesion instances and their local regions. The source lesion image contains candidate lesion instances. The original segmentation label corresponding to the candidate lesion instance. ⊙ represents the native segmentation label of the target image, and ⊙ represents the Hadamard product.

[0048] Step 3: Dual-stream feature extraction and encoding. For example... Figure 3 As shown in the overall architecture diagram of the SD-DualNet network of this invention, the multimodal images processed in step two are physically decoupled according to modal semantics, divided into a main stream containing DWI and ADC sequences and an auxiliary stream containing FLAIR sequences. These are then fed into their respective independent and parallel 3D encoder paths to extract multi-scale feature pyramids step by step. and ,in, For network depth-level indexing, both encoders consist of four stacked stages (Stage 1 to 4), with the number of channels at each stage set to 32, 64, 128, 256, and 320 from shallow to deep, respectively. Nonlinear stability of feature propagation is maintained through three-dimensional instance normalization (InstanceNorm3d) and the LeakyReLU activation function.

[0049] Symmetric difference modules (SDMs) are embedded in each feature layer of mainstream encoders. For example... Figure 5 As shown in the structure of the Symmetric Difference Module (SDM) of this invention, the SDM first moves along the sagittal plane to the main feature map. Perform mirror flip operation Construct mirror reference features: ; In the formula, For mirror reference feature map; Calculate the absolute difference between the original feature map and its mirror image to generate the original difference map: ; In the formula, This is the original symmetrical difference diagram; Will The input consists of a lightweight convolutional bottleneck layer composed of two 1×1×1 convolutional layers and a ReLU activation function, which is then mapped by a Sigmoid function to generate a refined spatially symmetric difference mask. : ; in, and These are the weight matrices for dimensionality reduction and dimensionality increase convolutions, respectively. It is the ReLU activation function. This is the Sigmoid activation function. The mask... Bearing the asymmetric information caused by the lesion, and serving as the core indicator guiding spatial attention allocation, the final output feature is defined as: ; This mechanism assigns higher activation weights to candidate lesion regions that significantly disrupt bilateral symmetry, while symmetrical physiological structures are relatively suppressed due to their smaller differences, effectively reducing the false positive rate in complex regions. The aforementioned SDM module achieves non-invasive implantation based on a dynamic feature interception mechanism using the forward hook function. It captures and processes intermediate layer features in real time without any static modification to the basic topology of the 3D encoder-decoder network, fully preserving the deep supervision path and gradient propagation stability of the basic framework, with a parameter increment of only approximately 1.4M.

[0050] formula and The technological improvement lies in the transfer from image space to deep feature space. Existing technologies, which introduce symmetric priors, typically involve complex, non-rigid registration of the original image at the input, which is not only time-consuming but also allows even small registration errors to be amplified infinitely by the network. This invention abandons image-level registration and innovatively proposes dynamically performing mirror flipping and difference calculations in a deep abstract feature space, giving the model extremely high deformation robustness and effectively avoiding the registration error amplification problem of existing technologies.

[0051] formula The technological improvement (based on mask generation using a convolutional bottleneck layer) lies in its lightweight fault tolerance and noise reduction. Existing technologies often employ bi-branch Siamese networks with a significantly increased number of parameters when calculating bilateral differences. This invention designs an extremely lightweight dimensionality reduction-up convolutional bottleneck layer, specifically designed to smooth out non-pathological difference noise caused by minor head rotations or physiological asymmetries in the channel dimension, thereby accurately locating the true lesion with extremely low computational overhead.

[0052] Step 4: Cross-modal channel calibration and fusion. Figure 4 This is a single-stage data flow diagram of the dual-stream fusion network of the present invention, including the independent feature extraction paths of the main stream and auxiliary streams and the detailed internal structure of the cross-modal channel calibration module (CMCCBlock). The cross-modal channel calibration module (CMCC) is embedded at each jump connection between the main encoder and decoder. For the first... Layer feature map, and First, the data is spliced ​​together along the channel dimension. And extract the global context vector through global average pooling (GAP) operation: ; In the formula, For global context vectors, For the average pooling operation across the entire series, Let it be a 2C-dimensional real vector space. For the first Feature maps spliced ​​from each network layer; Subsequently, the compression ratio of the hidden layer channel is... Linear bottleneck structure generates auxiliary flow channel attention weights : ; in, and These are the learnable weight parameter matrices for the dimensionality-reduced and dimensionality-increasing convolutional layers, respectively. σ is the ReLU activation function, and σ is the Sigmoid activation function. For the generated channel attention mask. Final adaptive fusion features. By generating weights It acts on the secondary stream and is generated by residual superposition with the mainstream stream: ; In the formula, For the first The adaptive fusion feature map is generated after each network layer is calibrated and fused by CMCC. To add element by element, For the first Auxiliary stream encoder in each network layer Extracted feature maps, The channel attention weights are dynamically generated by the CMCC module based on the global context. These are the channel-level gating weights learned adaptively by the network.

[0053] When the auxiliary flow extracts background artifacts unrelated to the lesion, the network, by learning cross-modal contextual relevance, tends to output extremely low weights close to 0. Mainstream pathological features remain uncontaminated; conversely, if the auxiliary flow provides a clear outline of brain structures, the weights are strongly activated. The structured high-frequency priors are fully preserved and incorporated into the mainstream features.

[0054] In the feature decoding and reconstruction stage, such as Figure 4 As shown on the right, the fused high-level features It acts as a high-quality jump connection between the encoder and decoder. Decoding features from deeper layers... First, spatial upsampling and channel dimensionality reduction are performed using transposed convolution (TransposedConv3D). The upsampled features are then compared with... The data is concatenated along the channel dimension and then fed into a series of 3D convolutional modules for deep decoding of multimodal semantics and refinement of spatial details, ultimately outputting the decoded features of the current stage. This is done to gradually restore the original image resolution.

[0055] The global average pooling (GAP) or fully connected network (dimensionality reduction and expansion mechanism) and other basic mathematical operation formulas of this invention, under the specific architecture of this invention, have extremely significant technical improvements compared with the direct splicing and fusion of existing technologies.

[0056] The technical improvement of the Global Average Pooling (GAP) formula lies in its upgrade from single-modal self-attention to cross-modal global context. Existing channel attention techniques typically perform pooling only within a single feature map. The improvement of this invention lies in the fact that, before executing the GAP formula, the separated mainstream (DWI / ADC, representing pathology) and secondary flow (FLAIR, representing anatomical structure) features are forcibly concatenated along the channel dimension. Therefore, the z extracted by this formula is no longer a simple feature vector, but rather, for the first time, quantifies the high-order physical correlation (cross-modal context) between sensitive pathological features and background structural features. This allows the network to simultaneously perceive the state of both flows, providing a global decision-making basis for subsequent denoising.

[0057] Auxiliary channel attention weight The technological improvement lies in constructing a dynamic physical valve to resist the propagation of multimodal noise. Existing technologies often employ static channel stitching when fusing multimodal images (assuming each modality is absolutely effective). The innovation lies in utilizing cross-modal vectors. Through a linear bottleneck layer with a compression ratio of 1 / 4 ( Attention weights for the auxiliary flow channels were dynamically generated. (Gating weights specific to auxiliary flow features). It directly addresses the industry's technical pain point that non-specific high-signal artifacts in the FLAIR modality easily contaminate DWI features. When the auxiliary flow is determined to contain artifact noise, the output is close to 0 (noise truncation); when the auxiliary flow is determined to be a clear brain boundary, the output is close to 1 (allowing features).

[0058] Step 5: Segmentation Output and Loss Function Calculation. The feature maps decoded in Step 4 are mapped to segmentation probability maps through a 1×1×1 convolutional layer. The network is trained using a composite loss function that combines cross-entropy loss and Dice loss. In a single output layer of the network, the composite loss function is defined as: ; Where Y represents the true voxel-level gold standard label. This represents the predicted probability of the network. Applying independent classification penalties at the voxel level helps stabilize the gradient in the early stages of training; This approach measures the volume overlap between the predicted mask and the gold standard from a global perspective, effectively overcoming the gradient submersion problem caused by excessive background coverage of small lesions. To further mitigate gradient vanishing that may occur in deep networks, segmentation loss is calculated at multiple scale levels of the decoder, with the final overall optimization objective being... The weighted sum of losses at each level: ; Where L is the total number of output levels of the decoder. The weights for deep supervision decay exponentially as the resolution decreases. This mechanism forces the intermediate hidden layers of the network to learn highly discriminative semantic features. The network parameters are updated by minimizing this loss function using the backpropagation algorithm until the model converges.

[0059] The training hyperparameters are configured as follows: the optimizer is stochastic gradient descent (SGD) with Nesterov momentum, the initial learning rate is set to 0.01, and a poly learning rate decay strategy (exponent set to 0.9) is used to ensure smooth convergence in the later stages of training; the training batch size is set to 8, and the preprocessed multimodal 3D images are randomly cropped into 80×112×112 3D image patches and fed into the network; the training is carried out for a total of 300 epochs, and each epoch contains 250 forward propagations.

[0060] See Figure 3 As shown, the SD-DualNet network uses a three-dimensional encoder-decoder architecture as its backbone: the left encoder path consists of four stacked stages (Stages 1-4), including a main stream and auxiliary stream encoders that run in parallel with physical decoupling; the right decoder path also corresponds to four stages, restoring the spatial resolution of the feature maps layer by layer; the encoder and decoder are adaptively fused at the skip connections of each stage through the CMCC module; the output of each stage of the main encoder embeds the SDM module in a non-intrusive manner; the final decoded features at the output are mapped through a 1×1×1 convolutional layer to output a voxel-level lesion segmentation prediction map. .

[0061] Experiment 1: To verify the effectiveness and advancement of the proposed method, an experimental evaluation was conducted on the publicly available medical imaging benchmark dataset ISLES2022 (Ischemic Stroke Lesion Segmentation 2022). This dataset contains 250 stroke patient cases that underwent rigorous clinical desensitization. For each patient, the dataset provides three highly diagnostically complementary 3D image sequences: DWI, ADC, and FLAIR. The infarct core gold standard for all cases was manually delineated at the voxel level by experienced neuroradiologists. This example employs a standard 5-fold cross-validation strategy, and the final model performance is the average of the five validation results. All experiments were performed on a computing node configured with a single NVIDIA GeForce RTX 4090 (24GB) GPU, using PyTorch 2.7.1 as the underlying deep learning framework.

[0062] The baseline comparisons cover three mainstream architectures: classic CNN architectures (SEAN, DeepLabV3+, nnUNetv2), Transformer-based architectures (UNETR, SwinUNETR, SegFormer, TransBTS), and emerging state-space models (Mamba-UNet), as well as advanced algorithms proposed in recent years for this task (St-RegSeg, F3-Net, LBMNet, FRPNet, Tri-UNetX-2D). All comparison models use the same underlying training protocol and hyperparameter configuration to ensure the fairness of the comparative experiments. Evaluation metrics include Dice similarity coefficient (DSC), intersection-over-union ratio (IoU), recall, F2 score, 95% Hausdorff distance (HD95), and absolute volume error (AVD), and the statistical significance of performance improvements is verified using a nonparametric paired Wilcoxon signed-rank test (p<0.05).

[0063] Experimental results show that the proposed method, SD-DualNet, achieves optimal performance across all evaluation metrics: the Dice similarity coefficient reaches 86.15%, significantly outperforming the general high-order benchmark nnUNetv2 (78.89%) in medical image segmentation, and surpassing recent advanced models Tri-UNetX-2D (86.00%) and FRPNet (85.72%); the recall rate reaches 86.45%, and the F2-Score reaches 86.38%, both ranking first among all compared methods; HD95 is significantly reduced to 4.33mm, and AVD is reduced to 11.45%, and the above performance improvements are statistically significant compared to the second-best models, fully verifying the comprehensive advantages of the proposed method in multi-dimensional clinical indicators. Figure 7 The images show a comparison of the visualization segmentation results of this invention and existing representative methods in three typical clinical scenarios, from top to bottom. Figure 7 The image in middle (a) shows a scenario with a high risk of being missed in diagnosis for tiny lesions. Figure 7 Image (b) shows a scene of multiple lesions with scattered boundaries. Figure 7 The middle (c) image shows a large-area lesion with complex morphology that is easily affected by ventricular artifacts.

[0064] The visualization results in three typical clinical scenarios further demonstrate the significant advantages of the method of the present invention in capturing small lesions, refining the segmentation of multiple lesions, and delineating the boundaries of large-area lesions.

[0065] Table 1 Model DSC↑ IoU↑ Recall↑ F2-Score↑ HD95 (mm) ↓ AVD↓ UNETR 56.32% 46.85% 51.10% 52.52% 18.73 58.10% SwinUNETR 61.08% 49.75% 61.37% 60.61% 12.72 46.45% SEAN 63.34% 53.47% 59.46% 61.34% 11.45 33.14% SegFormer 64.64% 52.24% 62.37% 62.97% 9.95 28.61% TransBTS 67.96% 54.24% 61.43% 64.36% 9.25 33.38% DeepLabV3+ 72.72% 61.06% 67.83% 69.48% 7.07 24.96% nnUNetv2 78.89% 68.28% 75.50% 76.57% 7.14 21.52% Mamba-UNet 80.15% 67.54% 78.82% 79.35% 5.98 15.20% St-RegSeg 81.30% 69.12% 79.50% 80.21% 5.52 13.85% F3-Net 79.92% 67.85% 78.22% 79.28% 6.25 16.50% LBMNet 82.03% 61.50% 84.30% 85.50% 5.10 14.20% FRPNet 85.72% 79.58% 84.93% 85.14% 4.86 12.15% Tri-UNetX-2D 86.00% 75.40% 83.20% 84.15% 8.30 13.60% SD-DualNet (Ours) 86.15%∗ 76.21% 86.45%∗ 86.38%∗ 4.33∗ 11.45%∗ Note: Optimal values ​​are shown in bold. * indicates that the performance improvement of the method of this invention (SD-DualNet) compared with the suboptimal model is statistically significant (p<0.05).

[0066] Experiment Example 2: Ablation Experiment. To further verify the independent contributions of the three core mechanisms—Intelligent Instance Enhancement Strategy (SCP), DualStream Feature Decoupling Architecture (DualStream), and Symmetric Differential Module (SDM)—to segmentation performance, a systematic ablation and combination experiment was conducted on the ISLES2022 dataset. All variant models were trained and evaluated under the same five-fold cross-validation protocol.

[0067] Starting with the baseline model nnUNetv2, a comparative experimental group was constructed by gradually introducing various core modules: After introducing the SCP strategy alone, the model recall increased from 75.50% to 80.85%, and the Dice coefficient jumped to 81.12%, proving that the SCP strategy forces the encoder to establish perceptual sensitivity to small targets from the data source, effectively alleviating the problem of missed diagnoses in clinical practice; After introducing the two-stream decoupling architecture alone, the absolute volumetric error (AVD) decreased from 21.52% to 16.80%, verifying that physically isolated multimodal feature flows can successfully suppress the FLAIR sequence. Non-specific artifact interference; after the SDM module was connected alone, the boundary accuracy index HD95 dropped sharply from 7.14mm to 5.45mm, confirming the independent effectiveness of the anatomical symmetry prior in refining the boundary of unilateral lesions; after the simultaneous introduction of SCP and dual-stream architecture, the Dice coefficient jumped to 83.75%, and the model complemented each other in the detection of small lesions and multimodal noise resistance; after integrating all three core mechanisms, Dice steadily improved to 86.15%, and HD95 was significantly improved to 4.33mm, fully demonstrating that the three modules produced a significant synergistic effect after deep integration.

[0068] Table 2 Model SCP DualStream SDM DSC↑ IoU↑ Recall↑ F2-Score↑ HD95 (mm) ↓ AVD↓ Baseline model 78.89% 68.28% 75.50% 76.57% 7.14 21.52% Experiment 1 81.12% 71.35% 80.85% 81.02% 6.24 19.12% Experiment 2 82.05% 72.45% 78.15% 79.20% 6.05 16.80% Experiment 3 81.65% 71.80% 77.20% 77.95% 5.45 18.25% Experiment 4 83.75% 74.88% 83.10% 83.35% 5.52 15.40% Ours 86.15% 76.21% 86.45% 86.38% 4.33 11.45% like Figure 6 As shown, the qualitative visualization analysis further confirms the above quantitative conclusions: in scenarios containing microlesions ( Figure 6 In Figure (a), the baseline network exhibits significant missed diagnoses due to the loss of small target features caused by continuous downsampling. After introducing the SCP strategy, the network's sensitivity to small targets is effectively activated, successfully detecting hidden lacunar infarct lesions; in scenarios containing large-area complex boundary lesions ( Figure 6 In Figure (b), the baseline model produces significant boundary spillover and false positive predictions in the healthy hemisphere. With the complete integration of the dual-stream architecture and SDM, the cross-hemispheric mirror difference operation completely eliminates the false positive artifacts caused by the symmetrical physiological structure, and the contour boundary of the lesion area is accurately restored, which is closely corroborated by the significant decrease in HD95 in the quantitative assessment.

[0069] Experiment Example 3: Module Universality and Fine-Grained Performance Verification Experiment. This example further verifies the effectiveness of the method of the present invention from two dimensions.

[0070] On the one hand, to verify the universality of Symmetric Differential Modules (SDM) across different basic topologies, baseline models representing three major architectural paradigms were selected for module porting experiments: nnUNetv2 based on a pure convolutional architecture (CNN), SwinUNETR based on a Transformer architecture, and Mamba-UNet based on a State-Space Model (SSM). During the porting process, SDM was directly embedded as a lightweight companion layer into the encoder feature output at the corresponding network depth in a non-intrusive manner, without any static modifications to the original cross-layer connections and basic computation graph. Experimental results show that after introducing SDM, the Dice coefficients of the three baseline models increased to 81.95% (nnUNetv2), 64.55% (SwinUNETR), and 82.68% (Mamba-UNet), respectively, with only a slight increase of approximately 1.4M parameters. Furthermore, the HD95 of all models integrating SDM showed a significant decrease, verifying that the anatomical symmetry prior extracted by SDM, as a universal medical physical constraint, can seamlessly empower various deep learning segmentation frameworks.

[0071] Table 3 Model Architectural Paradigm Params(M)↓ DSC↑ HD95 (mm) ↓ nnUNetv2 CNN 31.2 78.89% 7.14 nnUNetv2+SDM CNN 32.6 81.95% 5.08 SwinUNETR Transformer 62.2 61.08% 12.72 SwinUNETR+SDM Transformer 63.6 64.55% 9.85 Mamba-UNet SSM 45.8 80.15% 5.98 Mamba-UNet+SDM SSM 47.3 82.68% 4.85 On the other hand, to comprehensively evaluate the ability of the method of this invention to perceive lesions of different scales, the ISLES2022 dataset was divided into three subsets based on the actual infarct volume: small lesions (volume < 5 mL), medium lesions (volume 5–50 mL), and large lesions (volume > 50 mL). These subsets were then compared horizontally with the baseline model nnUNetv2 and the current best proprietary network FRPNet. Experimental results show that on the small lesion subset, the recall rate of the baseline model nnUNetv2 was only 52.30%, and the recall rate of FRPNet was only 64.15%. In contrast, the recall rate of the method of this invention, SD-DualNet, jumped significantly to 76.85%, and the DSC reached 71.20%, fully demonstrating that the SCP strategy significantly improves the model's ability to prevent missed diagnoses in extreme long-tail clinical scenarios through active mining and adaptive reconstruction of small instances. On the large lesion subset, the DSC of the method of this invention reached 89.45%, and the recall rate reached 91.05%, also ranking first among all methods.

[0072] Table 4

[0073] Experiment Example 4: Generalization Performance Verification Experiment. To verify the cross-domain generalization ability of the method of this invention, independent generalization experiments were conducted on the ATLASv2.0 dataset. The dataset focuses on the segmentation of chronic stroke (softening lesions) in the T1-weighted (T1w) unimodal scenario. The lesions exhibit low signal characteristics and are extremely similar in grayscale to normal ventricles and cerebrospinal fluid (CSF), which is a recognized high-difficulty task in the current field of medical image segmentation.

[0074] To address the physical limitation of single-modal input causing the failure of the dual-stream architecture, a pseudo-dual-stream feature extraction mechanism based on frequency domain decoupling is constructed: the main stream (StreamA) uses the original T1w image as input, responsible for capturing low-frequency anatomical structures and global contextual semantics; the auxiliary stream (StreamB) inputs the high-frequency gradient map extracted by the 3D edge operator, forcing the network to explicitly focus on the lesion edges where local grayscale changes drastically. The remaining preprocessing, SDM module embedding, CMCC module fusion, and loss function calculation steps are the same as in Implementation Method 1. In a preferred embodiment, the 3D edge operator is a 3D Sobel operator (3DSobelOperator), and its kernel function size is set to... Individual elements. In practice, this operator is executed along the three orthogonal spatial directions (X, Y, Z axes) of the 3D image. The fixed-weight convolution is used to calculate the sum of squares of the gradients in each direction and take the square root to generate a three-dimensional spatial gradient magnitude map.

[0075] In this embodiment, the specific tool used to extract the high-frequency gradient map is the 3D Sobel operator (3DSobelOperator), whose kernel function size is 3x3x3. The reason for choosing this operator and size is that it can extract the edge gradients in three-dimensional space (X, Y, Z axes) while providing a certain smoothing effect, effectively resisting discrete noise in single-modal MRI images.

[0076] Experimental results show that the method of this invention achieves a Dice coefficient of 68.92%, a recall of 71.19%, and an HD95 of 22.75 mm, surpassing all comparable methods, including the latest state-of-the-art algorithms specifically designed for this dataset, such as GFA-UNet (DSC 61.25%) which incorporates 3D global attention, Shang et al.'s method (DSC 64.50%) which employs a novel labeling strategy, and LBMNet (DSC 67.57%) which uses a CNN-Mamba hybrid architecture. This fully demonstrates that the method of this invention has excellent cross-domain generalization and robustness across different modalities and pathological stages.

[0077] Table 5 Model DSC↑ Recall↑ HD95 (mm) ↓ UNETR 44.50% 35.20% 46.20 SegFormer 46.87% 36.03% 44.78 SwinUNETR 49.58% 41.50% 44.15 TransBTS 52.15% 45.30% 39.80 DeepLabV3+ 57.03% 47.04% 34.64 nnUNetv2 57.34% 55.14% 33.87 Mamba-UNet 60.85% 60.33% 29.78 GFA-UNet 61.25% 58.40% 29.10 F3-Net 62.80% 61.50% 28.50 Shangetal. 64.50% 59.00% 27.45 Tri-UNetX-2D 65.85% 66.40% 26.30 FRPNet 66.30% 67.15% 24.90 LBMNet 67.57% 69.38% 24.15 SD-DualNet (Ours) 68.92%∗ 71.19%∗ 22.75∗ The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for segmenting lesions in dual-flow stroke based on symmetric difference learning, characterized in that, The method includes the following steps: S1. Perform intelligent instance enhancement processing on the input multimodal brain magnetic resonance images; S2. Decouple the processed multimodal images physically according to modal semantics; S3. Embed the Symmetric Difference Module (SDM) in each feature layer of the mainstream encoder to generate a refined spatial symmetric difference mask, perform adaptive enhancement on the mainstream features, and output the features. S4. At each jump connection point between the mainstream encoder and decoder, a cross-modal channel calibration module (CMCC) is embedded to obtain adaptive fusion features. S5. The adaptive fusion features at each level are fed into a unified 3D decoder and upsampled level by level to generate voxel-level segmentation masks. During the training phase, the segmentation loss is calculated simultaneously at multiple scale levels of the decoder. The overall optimization objective is the weighted sum of the losses at each level.

2. The method for segmenting dual-flow stroke lesions based on symmetric difference learning according to claim 1, characterized in that, In step S1, performing intelligent instance enhancement processing on the input multimodal brain magnetic resonance images includes: S1.1 Data Acquisition and Preprocessing: Acquire the brain MRI image sequence data to be segmented, store it in three-dimensional voxel format, and denote it as the input tensor. ,in, These correspond to three modes: diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC), and liquid attenuation inversion recovery sequence (FLAIR). These represent the height, width, and depth in three-dimensional space, respectively. It is the set of real numbers; S1.2, Intelligent Instance Enhancement Processing (SCP): Lesion Masking of Input Images Performing 3D connected component labeling operations using a 3×3×3 cube structuring element: ； In the formula, For the labeled voxel matrix, different instances are assigned independent integer labels. This represents the total number of independent lesion cases. This is a connected component labeling function, representing the execution of a three-dimensional connected component labeling operation; This is the lesion mask for the input image. It is a 3×3×3 cube structural element; For any instance Physical volume Defined as: ； In the formula, In three-dimensional voxel coordinates, For indicator functions, Independent positive integer labels for different instances; Constructing volume-constrained candidate lesion pools : ； In the formula, voxels, voxel, trigger probability set to ; When the enhancement operation is triggered, the candidate lesion pool is constrained by volume. Randomly selected target instances are used to construct a brain parenchyma safe zone using the physical properties of ADC and FLAIR sequences; for arbitrary modalities voxels The local adaptive cross threshold at a given location, and the dynamic threshold calculation are as follows: ； In the formula, voxels Local dynamic threshold at that location These represent the signal mean and standard deviation within a local three-dimensional sliding window centered on the voxel, respectively. This is a parameter for adjusting contrast sensitivity. This is the dynamic range constant of the image signal intensity; After obtaining the dynamic threshold matrix, morphological erosion operations are used to shrink the physical boundaries of the safe zone and mask the brain parenchyma safe zone. The generation process is defined as follows: ； In the formula, and For the three-dimensional feature tensor of the corresponding mode, and This is the adaptive threshold matrix calculated based on the mean and variance of the local sliding window. Represents voxel-level logical intersection operations. This indicates a morphological erosion operation performed using three-dimensional structuring elements. For local masks Perform cross-modal synchronous cropping and generate fused images through a mask-guided voxel replacement mechanism. With corresponding tags : ； ； In the formula, For the target image, This is a binary mask for cropping candidate microlesion instances and their local regions. The source lesion image contains candidate lesion instances. The original segmentation label corresponding to the candidate lesion instance. ⊙ represents the native segmentation label of the target image, and ⊙ represents the Hadamard unit.

3. The method for segmenting dual-flow stroke lesions based on symmetric difference learning according to claim 2, characterized in that, In step S2, the physical decoupling of the processed multimodal image according to modal semantics includes: The main stream containing DWI and ADC sequences and the auxiliary stream containing FLAIR sequences are fed into their respective independent parallel 3D encoder paths to extract multi-scale feature pyramids step by step. and The three-dimensional encoders are all composed of four feature extraction stages and a bottom bottleneck layer stacked together; in any given... Each feature extraction stage Representing the network depth-level index, both the mainstream and auxiliary streams employ two consecutive 3D convolutional modules for feature encoding; among them, For mainstream 3D encoders in the first Mainstream feature maps output at each scale level For the auxiliary flow three-dimensional encoder in the first Auxiliary flow feature maps output at each scale level; The data flow of a single convolutional module passes through the following layers in sequence: a 3×3×3 three-dimensional convolutional layer for extracting local spatial and pathological / anatomical features; a three-dimensional instance normalization layer to accelerate network convergence and reduce data bias caused by differences in different MRI equipment and individual patients; and a LeakyReLU activation function layer to introduce non-linear expressive power and maintain the stability of feature propagation; each encoder layer increases the number of feature channels step by step while performing feature downsampling. After the above-mentioned step-by-step independent encoding, the main stream and auxiliary stream respectively output multi-scale feature pyramids corresponding to different spatial resolutions and abstract semantic levels. and ,in, Representing the first The channel number, height, width, and depth of the level features provide physically isolated feature bases for cross-modal channel calibration fusion.

4. The method for segmenting dual-flow stroke lesions based on symmetric difference learning according to claim 3, characterized in that, In step S3, a refined spatially symmetric difference mask is generated, and adaptive enhancement is applied to the mainstream features. The output features include: Symmetric Dissimilarity Modules (SDMs) are embedded in the feature layers of the mainstream encoder at each level. Mirroring is performed on the mainstream features along the sagittal plane. The original difference map is generated by calculating the tensor absolute difference between the original feature map and the mirror reference feature map. This original difference map is then passed through a lightweight convolutional bottleneck layer to generate a refined spatially symmetric difference mask. Adaptive enhancement is applied to the mainstream features, and the output features are then optimized. Defined as: ； In the formula, For activation function, In the first The final adaptive fusion feature output by the cross-modal channel calibration module CMCC at each scale level.

5. The method for segmenting dual-flow stroke lesions based on symmetric difference learning according to claim 4, characterized in that, Generate refined spatially symmetric difference mask include: The Symmetric Difference Module (SDM) first moves along the sagittal plane to the main feature map. Perform mirror flip operation Construct mirror reference features: ； In the formula, For mirror reference feature map; Calculate the absolute difference between the original feature map and its mirror image to generate the original difference map: ； In the formula, This is the original symmetrical difference diagram; Will The input consists of a lightweight convolutional bottleneck layer composed of two 1×1×1 convolutional layers and a ReLU activation function, which is then mapped by a Sigmoid function to generate a refined spatially symmetric difference mask. : ； In the formula, and These are the weight matrices for dimensionality reduction and dimensionality increase convolutions, respectively. It is the ReLU activation function. This is the Sigmoid activation function.

6. The method for segmenting dual-flow stroke lesions based on symmetric difference learning according to claim 5, characterized in that, In step S4, the cross-modal channel calibration module CMCC is embedded to obtain adaptive fusion features, including: Will and After concatenation along the channel dimension, cross-modal context information is extracted using global average pooling, and auxiliary flow channel attention weights are generated through a linear bottleneck structure. After applying weights to the auxiliary flow features, the residuals are superimposed with the mainstream features to obtain adaptive fusion features: ； In the formula, For the first The adaptive fusion feature map is generated after each network layer is calibrated and fused by CMCC. To add element by element, For the first Auxiliary stream encoder in each network layer Extracted feature map, The channel attention weights are dynamically generated by the CMCC module based on the global context. These are the channel-level gating weights learned adaptively by the network.

7. The method for segmenting dual-flow stroke lesions based on symmetric difference learning according to claim 6, characterized in that, Generate auxiliary flow channel attention weights include: For the Layer feature map, and First, the data is spliced ​​together along the channel dimension. The global context vector is extracted through the global average pooling (GAP) operation: ； In the formula, For global context vectors, For the average pooling operation of the entire series, Let it be a 2C-dimensional real vector space. For the first Feature maps spliced ​​from each network layer; Through hidden layer channel compression ratio Linear bottleneck structure generates auxiliary flow channel attention weights : ； In the formula, and These are the learnable weight parameter matrices for the dimensionality-reduced and dimensionality-increasing convolutional layers, respectively. For the generated channel attention mask.

8. The method for segmenting dual-flow stroke lesions based on symmetric difference learning according to claim 7, characterized in that, In step S5, the overall optimization objective is... The weighted sum of losses at each level: ； ； In the formula, To optimize overall losses, For cross-entropy loss, For Dice similarity coefficient loss, This represents the total number of output layers for the decoder. The weights used for deep supervision decay exponentially with decreasing resolution. It is the gold standard label for true voxel-level data. For scale-level indexing, Let be the segmentation loss calculated at the l-th resolution level of the decoder. This represents the predicted probability of the network.

9. A dual-flow stroke lesion segmentation system based on symmetric difference learning, characterized in that, The system implements the dual-flow stroke lesion segmentation method based on symmetric difference learning as described in any one of claims 1-8, and the system comprises: The intelligent instance enhancement processing module is used to perform intelligent instance enhancement processing on the input multimodal brain magnetic resonance images; The physical decoupling module is used to physically decouple the processed multimodal images according to modal semantics; The symmetric difference module is used to embed in the feature layers of the mainstream encoder at each level to generate a fine-grained spatial symmetric difference mask, perform adaptive enhancement on the mainstream features, and output features; A cross-modal channel calibration module is used to embed adaptive fusion features at various jump connections between the main encoder and decoder. The weighted summation module for the loss at each level is used to feed the adaptive fusion features of each level into a unified 3D decoder and upsample them level by level to generate voxel-level segmentation masks. During the training phase, the segmentation loss is calculated simultaneously at multiple scale levels of the decoder, and the overall optimization objective is the weighted summation of the loss at each level.

10. The dual-flow stroke lesion segmentation system based on symmetric difference learning according to claim 9, characterized in that, The dual-flow stroke lesion segmentation system based on symmetry difference learning is mounted on a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can realize the functions of the dual-flow stroke lesion segmentation system based on symmetry difference learning.

Images