A stroke image segmentation method and system based on deep learning

Abstract

This invention relates to the field of intelligent medical image analysis technology, and in particular to a deep learning-based method and system for stroke image segmentation. The method includes: acquiring diffusion-weighted, apparent diffusion coefficient, and fluid attenuation inversion recovery images to form a dataset; obtaining standardized images and quality grades through resampling and intensity normalization; training a prior model with normal samples to generate a potential distribution centroid library and quality calibration curve; alternately updating the deformation field and anomaly editing field of the images to be segmented to generate candidate probability maps; zeroing the anomaly editing field to generate counterfactual healthy images; and combining differential evidence and contralateral mirror evidence to output lesion masks and suspicious region masks, thereby improving cross-quality stability and verifiability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image intelligent analysis, and in particular to a stroke image segmentation method and system based on deep learning. BACKGROUND

[0002] Stroke has the characteristics of high incidence, high disability rate and high mortality. The acute stage diagnosis and treatment emphasizes rapid evaluation and classification decision within the "time window". Imaging plays a key role in early screening, lesion range evaluation, hemorrhage and ischemia differentiation, treatment scheme selection and prognosis evaluation. In clinical work, the lesion boundary often presents weak contrast, complex shape, large volume span, and is affected by motion artifacts, stripe artifacts, scanning parameter differences and different center imaging processes, and the image quality fluctuates obviously, resulting in time-consuming manual delineation and consistency affected by operator experience, which is difficult to meet the requirements of high throughput, standardization and traceability. Existing automatic methods are prone to over-segmentation, under-segmentation and decreased stability in cross-device, cross-center and cross-quality level scenarios, and the "evidence chain explanation" of the segmentation results is insufficient, which is difficult to form a verifiable and reviewable output in the clinical process. SUMMARY

[0003] The present application provides a stroke image segmentation method and system based on deep learning, which at least solves the problem of how to realize stable output of stroke lesion segmentation and provide verifiable evidence under the conditions of image quality fluctuation and cross-center difference.

[0004] In a first aspect, the present application provides a stroke image segmentation method based on deep learning, comprising the following steps: Obtain diffusion weighted images, apparent diffusion coefficient images and liquid attenuated inversion recovery images to form an image data set; Perform spatial resampling and intensity normalization on the image data set to generate standardized image data and quality grading results, train a prior model based on the standardized image data of a normal subject set, generate a set of calibration information, and the set of calibration information includes a library of potential distribution centers and a set of quality calibration curves; For a subject to be segmented, based on the standardized image data, the quality grading results and the set of calibration information, alternately update the deformation field and the abnormal editing field to perform joint inference, obtain the abnormal editing field and the lesion candidate probability map; Generate counterfactual healthy images under the condition of zero abnormal editing field based on the generated prior model, obtain counterfactual difference evidence and contralateral mirror control evidence, and fuse the lesion candidate probability map to output a lesion mask and a suspicious region mask.

[0005] In a second aspect, the present application provides a stroke image segmentation system based on deep learning, which is used to implement the stroke image segmentation method based on deep learning, and the system comprises: The image acquisition module is used to acquire diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images to form an image dataset. The standard modeling module is used to spatially resample and normalize the intensity of the image dataset to generate standardized image data and quality grading results; and to train and generate a prior model based on the standardized image data of the normal subject set, generating a calibration information set, which includes a potential distribution centroid library and a quality calibration curve set. The joint inference module is used to perform joint inference on the subject to be segmented based on standardized image data, quality grading results and calibration information set, alternately updating the deformation field and the abnormal editing field to obtain the abnormal editing field and the lesion candidate probability map. The evidence output module is used to generate counterfactual health images based on the generative prior model under the condition of zeroing the abnormal editing field, obtain counterfactual difference evidence and contralateral mirror comparison evidence, and output lesion mask and suspicious area mask by fusing lesion candidate probability map.

[0006] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: By standardizing the preprocessing of diffusion-weighted, apparent diffusion coefficient, and liquid attenuation inversion recovery multi-sequence images, the consistency of input under cross-device and cross-parameter conditions was improved. By training a prior model based on normal subjects and constructing a potential distribution centroid library and a set of quality calibration curves, adaptive gating and threshold calibration for different image quality levels were achieved. By alternately updating the deformation field and the abnormal editing field for joint inference, robust alignment against artifacts and registration biases and stable output of lesion candidate probabilities were achieved. By zeroing the abnormal editing field to generate counterfactual healthy images and introducing contralateral mirror comparison evidence, the segmentation results were made verifiable and retrievable, thereby reducing missed and false detections and improving the ability to grade suspicious areas. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of the method of the present invention; Figure 2 This is an example image of a diffusion-weighted image in a specific embodiment of the present invention; Figure 3 This is an example image of the apparent diffusion coefficient in a specific embodiment of the present invention; Figure 4 This is an example image of a liquid attenuation inversion recovery image in a specific embodiment of the present invention; Figure 5 This is a schematic diagram showing the spatial alignment of the three modes in a specific embodiment of the present invention; Figure 6 This is a statistical diagram illustrating the distribution of quality grades in a specific embodiment of the present invention; Figure 7This is a schematic diagram of the Dice coefficient comparison box curve in a specific embodiment of the present invention; Figure 8 This is a schematic diagram of Dice statistics by quality grade in a specific embodiment of the present invention; Figure 9 This is a statistical diagram illustrating the relationship between lesion volume and Dice in a specific embodiment of the present invention; Figure 10 This is a structural block diagram of the system of the present invention. Detailed Implementation

[0008] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation.

[0009] Deep learning is a data-driven modeling method based on multi-layer neural networks. Through end-to-end training, it automatically learns hierarchical feature representations of images from a large number of samples, enabling it to perform complex pattern recognition and structured output without relying on manually designed rules. Compared to traditional algorithms based on thresholds, morphology, or handcrafted features, deep learning is more likely to form joint representations of local textures, edge transitions, and spatial context when processing medical images with weak contrast, varied morphology, and significant noise interference. Therefore, it has the potential to directly map "image input" into clinically required results such as "segmentation masks."

[0010] However, in real clinical settings, the performance of deep learning models is highly dependent on the consistency of data distribution and imaging quality: differences in scanning parameters, artifact types, signal-to-noise ratio levels, and sequence combinations from different centers can introduce significant domain shifts, leading to decreased stability in model output. Simultaneously, models often lack clinically verifiable comparative evidence, making it difficult to explain "why this area is identified as a lesion." Based on these characteristics, this invention focuses on the standardization and quality grading of multi-sequence images, generative modeling of normal priors, joint inference of deformation and abnormalities, and the fusion output of counterfactual healthy images and contralateral mirror-image comparative evidence. It constructs a stroke lesion segmentation scheme that maintains consistency and possesses a verifiable chain of evidence even under conditions of quality fluctuation.

[0011] like Figure 1 As shown, a deep learning-based stroke image segmentation method includes the following steps: Acquire diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images to form an image dataset; In one embodiment, after the subject enters the magnetic resonance imaging (MRI) procedure, diffusion-weighted images, apparent diffusion coefficient images, and fluid attenuation inversion recovery (FIR) images are acquired. The diffusion-weighted images characterize signal changes caused by restricted water molecule diffusion, the apparent diffusion coefficient images provide a quantitative representation of diffusion consistent with the diffusion-weighted images, and the FIR images suppress cerebrospinal fluid signals to highlight abnormal brain parenchymal signals. During acquisition, imaging parameters such as the same subject identifier, acquisition time, slice thickness, interslice spacing, matrix size, field of view, and phase encoding direction are written to each sequence, and a one-to-one sequence index is established for the three types of images. Subsequently, the diffusion-weighted images, apparent diffusion coefficient images, and FIR images, along with their imaging parameter information, are packaged into an image dataset for subsequent spatial resampling, intensity normalization, quality grading, and joint inference processing. This step reduces segmentation errors caused by cross-sequence inconsistencies through simultaneous acquisition and unified management of multiple sequences.

[0012] Acquiring diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images includes: acquiring diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images with the same subject identifier, and associating the diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images into an image dataset.

[0013] In one embodiment, when acquiring diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images with the same subject identifier, the imaging device or image archiving and communication system first outputs three types of sequence files and their accompanying imaging information files. The imaging information files include at least the subject identifier, examination number, sequence type identifier, acquisition time, slice location, slice thickness, interslice spacing, matrix size, field of view, orientation information, and phase encoding orientation information. The system performs a consistency check on the input files. The consistency check includes: verifying that the subject identifier is consistent; verifying that the examination number is consistent; verifying that the spatial coverage of the three types of sequences overlaps and that the slice location differences are within a preset tolerance; and verifying that the orientation information can be used to establish the same spatial coordinate system. If the consistency check passes, the diffusion-weighted image, apparent diffusion coefficient image, and liquid attenuation inversion recovery image are added to the buffer as a single data set, and an image dataset index is generated for this data set.

[0014] To ensure the traceability and reproducibility of subsequent processing, the image dataset is managed using a structured recording method. The image dataset includes at least: diffusion-weighted image paths, apparent diffusion coefficient image paths, liquid attenuation inversion recovery image paths, imaging information summaries, inter-sequence registration relationship identifiers, and data status identifiers. The inter-sequence registration relationship identifiers indicate whether the three types of sequences are in the same spatial coordinate system. If they are not in the same spatial coordinate system, a registration identifier is recorded within the image dataset to trigger subsequent spatial resampling and alignment processing. The data status identifiers indicate whether the image dataset has completed integrity verification, whether there are missing slices, and whether there are obvious artifacts, so that it can be directly used in subsequent quality grading.

[0015] When associating diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images into an image dataset, a correspondence strategy based on subject identification and slice location is adopted: a slice sequence index is established for diffusion-weighted images, and separate slice sequence indices are established for apparent diffusion coefficient images and liquid attenuation inversion recovery images; under the same subject identification, matching is performed according to slice location from near to far to obtain a cross-sequence slice correspondence table, which is then written into the image dataset index. If there is an inconsistency in the number of slices, the sequence with the largest spatial coverage is used as the reference sequence, and the remaining sequences are interpolated to complete or cropped to the slice set of the reference sequence. The results of the completion or cropping operations are also recorded in the image dataset index. Through the above association method, the image dataset forms input units at the data level that can be directly used for subsequent spatial resampling, intensity normalization, and joint inference, avoiding lesion region shifts caused by sequence mismatch, slice correspondence errors, or inconsistencies in orientation, thereby improving the spatial consistency between subsequent lesion candidate probability maps and lesion masks.

[0016] Spatial resampling and intensity normalization are performed on the image dataset to generate standardized image data and quality grading results. A prior model is trained based on the standardized image data of the normal subject set to generate a calibration information set, which includes a potential distribution centroid library and a quality calibration curve set. In one embodiment, after reading the image dataset, the system first unifies the diffusion-weighted image, the apparent diffusion coefficient image, and the liquid attenuation inversion recovery image to the same spatial coordinate system, and then resamples all three to a unified spatial sampling grid. Subsequently, intensity normalization processing is performed within the brain region. This intensity normalization process includes grayscale truncation, subtracting the brain region's statistical mean, and dividing by the brain region's statistical standard deviation to generate standardized image data. Based on the standardized image data, motion artifact indices, signal-to-noise ratio indices, and diffusion distortion risk indices are calculated, and quality grading results are obtained according to preset mapping rules for subsequent model selection and threshold configuration.

[0017] A prior model is generated by training standardized image data from a set of normal subjects. This prior model outputs a normal predicted image corresponding to the input modality, and a calibration information set is generated after training. The latent distribution centroid library is obtained by grouping the latent representations of normal subjects according to quality grading results and calculating the representative centroid of each group. The quality calibration curve set is obtained by statistically analyzing the reconstruction residual distribution and counterfactual difference distribution of normal subjects. It is used to determine the gating threshold under different quality grading results, thereby reducing the segmentation bias caused by different devices and different image qualities.

[0018] Spatial resampling and intensity normalization of image datasets include: performing spatial orientation uniformity on the image dataset; performing spatial resampling on the image dataset to obtain a uniform spatial sampling grid; and performing intensity normalization on the image dataset on the uniform spatial sampling grid to obtain standardized image data.

[0019] In one embodiment, when performing spatial resampling and intensity normalization on the image dataset, spatial orientation unification is first performed on the image dataset. Spatial orientation unification is used to unify the axial, left-right, and front-back directions of the diffusion-weighted image, apparent diffusion coefficient image, and liquid attenuation inversion recovery image under the same spatial convention, so as to ensure the stability of the pixel correspondence in subsequent cross-sequences. Specifically, the system reads the orientation information and pixel spacing information of the three types of images, rearranges the images into a unified axial order based on the orientation information, and corrects for any possible left-right flipping; if there is missing orientation information in the three types of images, the liquid attenuation inversion recovery image is used as a reference sequence, and the diffusion-weighted image and apparent diffusion coefficient image are mapped to the reference sequence space through rigid alignment to complete the orientation unification.

[0020] After spatial orientation unification, spatial resampling is performed on the image dataset to obtain a unified spatial sampling grid. This unified spatial sampling grid eliminates differences in slice thickness, inter-slice spacing, and planar resolution among different sequences, enabling subsequent quality grading, prior modeling, and joint inference for the three types of sequences on the same voxel grid. Specifically, the system uses the liquid attenuation inversion recovery image as a reference to determine the spatial range and sampling interval of the unified spatial sampling grid; spatial interpolation resampling is performed on the diffusion-weighted image and the apparent diffusion coefficient image, and consistent resampling verification is performed on the liquid attenuation inversion recovery image to ensure a one-to-one correspondence of voxel coordinates among the three types of sequences. Trilinear interpolation can be used, and the monotonicity of grayscale values ​​is maintained when resampling the apparent diffusion coefficient image to reduce abnormal gradients introduced by interpolation.

[0021] Intensity normalization is performed on the image dataset on a unified spatial sampling grid to obtain standardized image data. Intensity normalization is used to reduce grayscale drift caused by different devices, sequence gains, and scanning conditions, making the generated prior model and subsequent joint inference more stable to the input distribution. Specifically, the system performs brain region extraction on the liquid attenuation inversion recovery image to obtain a brain region mask, and maps the brain region mask to the unified spatial sampling grid where the diffusion-weighted image and the apparent diffusion coefficient image are located. Under the constraint of the brain region mask, grayscale truncation and standardization are performed on the three types of sequences respectively. Grayscale truncation is used to remove the influence of extreme high and low values ​​on the statistics, and standardization is performed by subtracting the brain region statistical mean and dividing by the brain region statistical standard deviation. Standardized image data formed by spatial orientation unification, spatial resampling, and intensity normalization ensures consistency of the three types of sequences in both space and grayscale. This provides reliable input for the stable generation of subsequent quality grading results, grouping statistics of the potential distribution centroid library, and joint inference of the abnormal editing field and lesion candidate probability map, thereby reducing lesion boundary shifts and false positives / false negatives caused by sequence misalignment and grayscale drift.

[0022] The generation of quality grading results includes: calculating motion artifact index, signal-to-noise ratio index, and diffusion distortion risk index based on standardized image data, and mapping the motion artifact index, signal-to-noise ratio index, and diffusion distortion risk index to quality grading results based on grading mapping rules.

[0023] When generating quality grading results for standardized image data, the system assesses the quality of diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images within the same spatial sampling grid, and outputs the quality grading results for subsequent calibration information set retrieval and gating threshold configuration. Motion artifact indicators reflect inter-slice discontinuities and ghosting caused by subject movement. Specifically, the similarity of adjacent slices in the diffusion-weighted image and the liquid attenuation inversion recovery image is calculated and summarized using the following expression: in, As an indicator of motion artifacts, Number of slices For the first Gray-scale matrix of brain regions in Zhang slices, This is the function for calculating the correlation coefficient. A higher value indicates poorer consistency between slices. The signal-to-noise ratio (SNR) metric reflects the effective contrast and noise level of an image. The system calculates the brain region mean within the brain region mask and estimates the noise standard deviation in non-brain regions or low-signal regions, using the following calculation expression: in, This refers to the signal-to-noise ratio (SNR) metric. The mean gray level of the brain region. The noise grayscale standard deviation is used. The diffusion distortion risk index is used to reflect the geometric inconsistency risk between diffusion sequences and structural sequences. The system uses the liquid attenuation inversion recovery image as a reference, performs rigid alignment on the diffusion-weighted image and calculates the edge consistency residual. Then, it combines the phase encoding direction information in the imaging parameter information to generate a distortion risk score. The larger the residual and the more sensitive the phase encoding direction, the higher the risk.

[0024] The quality classification mapping rule adopts a "worst-case-dominated approach combined with consistency verification" method: when the motion artifact index reaches the degradation condition, the quality level is directly reduced; when the signal-to-noise ratio index reaches the degradation condition, the quality level is reduced; when the diffusion distortion risk index reaches the degradation condition, the quality level is reduced; when there is a conflict among the three indices, the worse grade of the diffusion distortion risk index and the motion artifact index is used as the quality classification result. The quality classification result is output along with the standardized image data and is used to select the corresponding group in the potential distribution centroid library and to determine the gating threshold set in the quality calibration curve set, thereby maintaining the stability of joint inference and the consistency of lesion mask boundaries even when image quality fluctuates.

[0025] The training of the prior model includes: training a latent representation encoding network to encode standardized image data into latent representations; training a modal conditional rendering network to generate normal prediction images from the latent representations; the modal conditional rendering network generates a quality conditional vector based on the quality grading results and modulates the generation of normal prediction images based on the quality conditional vector.

[0026] When training the prior model, the system first constructs a set of normal subjects, which consists of image samples from subjects clinically confirmed to have no acute stroke lesions. The system performs the same preprocessing procedure on the normal subject set as on the subjects to be segmented to obtain standardized image data and retains the corresponding quality grading results for subsequent quality conditional training. The prior model consists of a latent representation encoding network and a modal conditional rendering network. The latent representation encoding network extracts latent representations related to anatomical structures from the multimodal standardized image data, while the modal conditional rendering network generates normal predicted images given the latent representations and quality conditional vectors, ensuring that the generated results are consistent with the input modality in terms of grayscale distribution and texture detail.

[0027] In one implementation, the latent representation encoding network employs an encoder structure. The input consists of standardized image data including diffusion-weighted images, apparent diffusion coefficient images, and fluid attenuation inversion recovery images. The output is the latent representation. During the training phase, the system applies a masking perturbation to the input images. This perturbation randomly masks certain brain region voxels, forcing the latent representation encoding network to learn stable anatomical priors rather than memorizing local noise. The training objective of the latent representation encoding network is to minimize the reconstruction difference between the normally predicted images and the corresponding standardized image data. During training, amplitude constraints are imposed on the latent representation to prevent uncontrolled drift of the latent representation across different subjects.

[0028] The input to the modal conditional rendering network includes a latent representation and a quality condition vector, and the output is a normal predicted image corresponding to the input modality. The quality condition vector is generated from the quality grading results, which are discrete grade identifiers. The system maps the quality grading results to a fixed-dimensional quality condition vector and injects this vector into the multi-layer feature transformation of the modal conditional rendering network to modulate texture details and noise levels at different quality levels. To ensure feasibility, the modulation can be implemented using channel scaling and offsetting. Specifically, the quality condition vector is linearly transformed to obtain channel scaling coefficients and channel offsets, and channel-level scaling and offsetting are performed on the intermediate feature maps of the rendering network, enabling the same latent representation to generate a normal predicted image matching the corresponding quality level under different quality conditions.

[0029] After training, the generated prior model has two uses: First, in joint inference, it provides the "image appearance that should be presented under normal circumstances" for the subject to be segmented as a reference, so that the abnormal editing field can capture regions that are inconsistent with the normal predicted image; Second, it provides the reconstruction residual distribution source for the quality calibration curve set under different quality grades, so that the gating threshold can maintain a consistent discrimination scale when the image quality changes, thereby reducing false detections and false negatives caused by low-quality images.

[0030] The process of generating a potential distribution centroid library includes: grouping the normal subject set based on imaging protocol information and quality grading results of the image dataset; generating a priori model by inputting standardized image data for each group to obtain a potential representation; calculating the distribution centroid for the potential representation of each group and establishing a correspondence between group identifiers and distribution centroids to obtain a potential distribution centroid library.

[0031] When generating the potential distribution centroid library, the imaging protocol information and quality grading results are first extracted from the image dataset of the normal subject set, and the normal subject set is then grouped accordingly. The imaging protocol information reflects the differences in image appearance caused by different devices and scanning settings. Specifically, it can be parsed from the imaging information file output by the image archiving and communication system. The imaging protocol information includes at least device manufacturer information, magnetic field strength information, sequence type identifier, spatial resolution information, slice thickness information, phase encoding direction information, and echo plane direction information. The system encodes the imaging protocol information into a protocol information identifier and combines the protocol information identifier with the quality grading results to generate a group identifier. This ensures that normal subject samples within the same group have consistency in scanning settings and image quality, thereby reducing the impact of cross-domain differences on subsequent prior constraints.

[0032] After grouping, the system generates a prior model from the standardized image data input of each group to obtain a latent representation. The latent representation is output by the latent representation encoding network and serves as a low-dimensional representation of the subject's brain anatomy, ensuring comparability within the same group. To guarantee the stability of the latent representation, the system uses a consistent input organization method for the standardized image data of each subject, such as stitching together images in a fixed channel order (diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images), and inputting the stitched multimodal data into the latent representation encoding network. The latent representation output by the latent representation encoding network can be a feature tensor of the bottleneck layer. The system performs global averaging and aggregation on the feature tensor to obtain a latent representation vector, and stores the latent representation vector associated with the corresponding group identifier. To avoid perturbation of the distribution centroid by abnormal samples or residual artifacts, the system can perform consistency screening within the group: calculating the similarity between each latent representation vector and the initial centroid vector within the group, removing latent representation vectors with similarity below a preset threshold, and then calculating the distribution centroid based on the set of screened latent representation vectors.

[0033] The centroid of distribution is used to characterize the representative central position of a group in the latent space. The centroid of distribution can be calculated by taking the mean of the latent representation vectors within the group, and can be expressed by the following formula: in, The centroid vector of the group distribution. For the first in the group The latent representation vector corresponding to a sample of normal subjects This represents the number of potential representation vectors involved in the calculation. After completing the distribution centroid calculation, the system establishes a correspondence between group identifiers and distribution centroid vectors, and writes this correspondence into the potential distribution centroid library. The potential distribution centroid library can be stored using a key-value structure, where the key is the group identifier and the value is the distribution centroid vector and its version identifier. The version identifier is used for iterative management when expanding the normal subject set or updating the prior model, avoiding instability in joint inference caused by mixing old and new priors.

[0034] Through the above grouping and distribution centroid calculation methods, the potential distribution centroid library can provide prior center constraints for subsequent subjects to be segmented, matching their imaging protocol information and quality grading results. This enables the potential representation updates in the joint inference stage to converge around the corresponding distribution centroid, reducing potential representation drift caused by equipment differences or quality fluctuations. This reduces the probability of misjudging artifacts as lesions and provides a foundation for the stable generation of subsequent abnormal editing fields and lesion candidate probability maps.

[0035] The process of generating a set of quality calibration curves includes: statistically generating the reconstruction residual distribution of the prior model based on the set of normal subjects under each quality grading result; performing joint inference of alternating updates of the deformation field and the abnormal editing field on the set of normal subjects to obtain the abnormal editing field, and statistically analyzing the distribution of deformation field changes; generating counterfactual health images based on the prior model under the condition of zeroing the abnormal editing field, and statistically analyzing the counterfactual difference distribution; and determining the set of gating thresholds based on the reconstruction residual distribution, the deformation field change distribution, and the counterfactual difference distribution to form a set of quality calibration curves.

[0036] In one embodiment, when generating a set of quality calibration curves, the system uses the quality grading results as an index to manage the set of normal subjects in buckets, and statistically analyzes three types of "normal fluctuations" under each quality grading result for gating and backoff discrimination in subsequent joint inference. The first type is the reconstructed residual distribution: for each normal subject, standardized image data is input into the prior model to obtain normal predicted images, the voxel-by-voxel difference between the two is calculated, and the residual statistics are summarized in the brain region mask as residual statistics. The residual statistics may include residual mean, residual standard deviation, and residual high quantile. The system aggregates the residual statistics of all normal subjects under the same quality grading result to obtain the reconstructed residual distribution corresponding to that quality grading result, thereby characterizing "the fitting level that the prior model should achieve under normal circumstances".

[0037] The second category is the distribution of deformation field changes: The system performs a joint inference process on the normal subject set, which is consistent with the subject to be segmented, alternatingly updating the deformation field and the abnormal editing field to obtain the abnormal editing field and record the deformation field change in each iteration. The deformation field change is used to reflect the spatial adjustment intensity required to achieve cross-sequence alignment. The system summarizes the deformation field change within the brain region mask into a change statistic, and aggregates them under the same quality grading results to obtain the deformation field change distribution, thereby characterizing the "acceptable fluctuation range of the alignment process under normal circumstances" and avoiding misjudging alignment errors as lesion evidence in low-quality images.

[0038] The third category is the counterfactual difference distribution: Based on the generative prior model, the system generates counterfactual healthy images under the condition of zeroing out the abnormal editing field, and calculates the difference statistics between the counterfactual healthy images and the standardized image data. These difference statistics characterize the "background noise level of counterfactual difference evidence under normal circumstances," and are aggregated under the same quality grading results to obtain the counterfactual difference distribution. Subsequently, the system determines the gating threshold set based on the three distributions and organizes the gating threshold set into a quality calibration curve set according to the quality grading results. To ensure reproducibility, the gating threshold set can be determined using quantiles, calculated using the following expression: ; in, The threshold value is used for gate control. For quantile operators, To preset the quantile level, The system constructs a sample set of samples from any of the following distributions: residual distribution, deformation field change distribution, or counterfactual difference distribution. The system jointly writes the reconstruction residual gating threshold, deformation field change gating threshold, and counterfactual difference gating threshold obtained under different quality grading results into a quality calibration curve set, and records the generated prior model version and the normal subject set version as correlation information. Through this calibration method, in subsequent joint inference of the subjects to be segmented, the gating threshold set can adaptively adjust according to the quality grading results, ensuring that the abnormal edit field update and lesion candidate probability map generation maintain a consistent discrimination scale under different image quality and device conditions. This reduces false detections caused by artifacts and minimizes missed detections of low-contrast lesions.

[0039] For segmented subjects, based on standardized image data, quality grading results and calibration information set, deformation field and abnormal editing field are alternately updated to perform joint inference and obtain abnormal editing field and lesion candidate probability map; After receiving standardized image data of the subject to be segmented, the system retrieves matching latent distribution centroids and gating threshold sets from the calibration information set based on the quality grading results. It initializes the deformation field as a unit mapping and the anomaly editing field as a zero matrix. Then, it performs multiple alternating iterations: first, using the liquid attenuation inversion recovery image as a reference, it updates the deformation field through a similarity-driven differentiable registration network, aligning the diffusion-weighted image and the apparent diffusion coefficient image to the reference space; then, it uses the aligned multimodal input to generate a prior model to obtain a normal prediction image, calculates the residual between the standardized image data and the normal prediction image, and updates the anomaly editing field under the constraints of the gating threshold set to suppress high residuals caused by artifacts; simultaneously, it shrinks the latent representation towards the latent distribution centroid to limit extrapolation. The iteration ends when the deformation field change and the anomaly editing field increment are satisfied, using these as convergence criteria. After iteration, the anomaly editing field is normalized and mapped to a lesion candidate probability map, which is then output.

[0040] Joint inference is performed by alternately updating the deformation field and the anomalous editing field, including: determining the prior centroid corresponding to the subject to be segmented in the potential distribution centroid library based on the quality grading results, and determining the gating threshold set based on the quality calibration curve set; determining the phase encoding direction information based on the imaging parameter information of the diffusion-weighted image, and extracting stripe artifact features in the diffusion-weighted image; updating the deformation field based on the stripe artifact features to align the standardized image data during the alternating iteration process, generating a normal prediction image based on the generative prior model, updating the anomalous editing field under the constraints of the gating threshold set based on the residual between the standardized image data and the normal prediction image, and updating the potential representation based on the prior centroid constraints to obtain the anomalous editing field and the lesion candidate probability map.

[0041] When performing joint inference on the segmented subject by alternately updating the deformation field and the anomalous editing field, the system first reads the standardized image data, quality grading results, and calibration information set. Based on the quality grading results, prior centroids are retrieved from the latent distribution centroid library. When the same quality grading result corresponds to multiple prior centroid records, the system further determines the prior centroids based on the imaging protocol information of the standardized image data and the grouping identifier matching relationship of each prior centroid record. Simultaneously, the system retrieves a set of gating thresholds from the quality calibration curve set. The gating threshold set is used to constrain the intensity range of subsequent residual interpretation and anomalous editing field updates. The deformation field is initialized as an identity map, the anomalous editing field is initialized as an all-zero matrix, and the latent representation is initialized as the forward output of the latent representation encoding network on the standardized image data.

[0042] The system determines the phase encoding direction information based on the imaging parameters of the diffusion-weighted image and extracts stripe artifact features from the diffusion-weighted image. Stripe artifact features characterize the periodic stripe intensity fluctuations along the phase encoding direction. The system calculates a local intensity profile along the phase encoding direction and generates a stripe artifact intensity map based on the ratio of the first-order difference energy to the low-frequency energy of the profile. After normalization within the brain region mask, the stripe artifact intensity map serves as a weight map for updating the deformation field. Low-value regions in the weight map represent significant stripe artifact areas. In these regions, the system reduces the weight of the registration similarity term to prevent artifacts from driving unrealistic deformations in the deformation field.

[0043] During the alternating iteration process, the system first updates the deformation field to align with the standardized image data. Specifically, using the liquid attenuation inversion recovery image as the reference sequence, the system performs differentiable registration optimization on the diffusion-weighted image and the apparent diffusion coefficient image. The registration objective consists of a similarity term and a smoothness term. The similarity term is calculated under weighting of the fringe artifact intensity map, reducing the contribution of salient fringe artifact regions to the deformation field update. The smoothness term is used to constrain the spatial continuity of the deformation field to avoid folding and sharp jumps. After the deformation field is updated, the system maps the diffusion-weighted image and the apparent diffusion coefficient image to the reference sequence space according to the deformation field, forming the alignment input for the current iteration.

[0044] The system then generates normal predicted images based on the generative prior model and updates the anomalous edit field and latent representations. Specifically, the system feeds the aligned input into the latent representation encoding network to obtain latent representations and applies prior centroid constraints to the latent representations, causing them to shrink towards the prior centroid to suppress cross-device extrapolation. Under the constraints of the quality condition vector corresponding to the latent representation and the quality grading results, the modal conditional rendering network outputs normal predicted images corresponding to each modality. The system calculates the residual map between the standardized image data and the normal predicted images and performs gated updates on the residual map according to a set of gate thresholds: when the residual amplitude is lower than the gate threshold, the residual is included in the background fluctuations and used to suppress the expansion of the anomalous edit field; when the residual amplitude is higher than the gate threshold and the cross-modal consistency constraint is satisfied, the residual is included in the anomalous interpretation and written into the anomalous edit field. Neighborhood consistency constraints are also introduced during the anomalous edit field update to make the anomalous edit field spatially connected and suppress isolated noise.

[0045] The iterative convergence criterion is jointly determined by the change in the deformation field and the increment of the abnormal editing field. The iteration ends when the preset convergence condition is met. The system normalizes the final abnormal editing field and maps it to a lesion candidate probability map, using the following calculation expression: ; in, This is a probability map of candidate lesions. An amplitude map of the abnormal editing field. This is a threshold map of the set of gated thresholds in space. For the Sigmoid function. By incorporating stripe artifact features into the deformation field update, incorporating the quality calibration curve set into the gating threshold set, and constraining the potential representation update with prior centroid, the joint inference process remains stable under different image qualities and scanning conditions, making the output abnormal editing field and lesion candidate probability map less sensitive to artifacts and possessing a reproducible update path.

[0046] Based on the generative prior model, counterfactual health images are generated under the condition of zeroing the abnormal editing field, and counterfactual difference evidence and contralateral mirror comparison evidence are obtained. The lesion candidate probability map is fused to output lesion mask and suspicious area mask.

[0047] After joint inference, all anomalous editing fields are zeroed out, and standardized image data is input to generate a priori models, producing counterfactual healthy images corresponding to diffusion-weighted images, apparent diffusion coefficient images, and fluid attenuation inversion recovery images. The difference between the standardized image data and the counterfactual healthy images is calculated voxel-by-voxel, and normalization and smoothing are performed within the brain region to obtain a counterfactual difference evidence map. Based on the midsagittal plane localization results, candidate regions are determined from the lesion candidate probability map and mapped to the contralateral mirror control region. The intensity difference and structural consistency between the candidate regions and the contralateral mirror control region are calculated to obtain the contralateral mirror control evidence map. The counterfactual difference evidence map and the contralateral mirror control evidence map are used to gating and classify the lesion candidate probability map, outputting lesion masks. Regions with low confidence but spatial continuity are output as suspicious region masks, while isolated noise points are filtered through three-dimensional connected components.

[0048] The process of fusing the lesion candidate probability map to output lesion masks and suspicious region masks includes: performing spatial alignment on standardized image data based on the deformation field to obtain spatially aligned image data; generating counterfactual healthy images based on the generative prior model under the condition of zeroing out the anomalous editing field; obtaining counterfactual difference evidence based on the spatially aligned image data and the counterfactual healthy images; determining candidate regions based on the lesion candidate probability map and generating corresponding contralateral mirror control regions to obtain contralateral mirror control evidence; and classifying the lesion candidate probability map into hierarchical categories based on the counterfactual difference evidence and the contralateral mirror control evidence to output lesion masks and suspicious region masks.

[0049] In one embodiment, when outputting lesion masks and suspicious region masks, spatial alignment is first performed on standardized image data based on the deformation field obtained through joint inference, forming spatially aligned image data. This spatially aligned image data is used to map diffusion-weighted images, apparent diffusion coefficient images, and fluid attenuation inversion recovery images to the same reference space, ensuring consistent spatial correspondence in subsequent evidence calculations at the voxel level. The alignment process uses the deformation field to resample the data for each modality and performs interpolation consistency checks at brain region boundaries to avoid spurious differences introduced by boundary interpolation.

[0050] The anomalous edit field is then zeroed out, and the spatially aligned image data is input to generate the prior model, yielding counterfactual health images corresponding to each modality. The counterfactual health images represent the normal appearance that the prior model should produce under the current subject's anatomical structure and image quality conditions, without introducing anomalous edit effects. Counterfactual difference evidence is calculated based on the spatially aligned image data and the counterfactual health images. This evidence measures the strength of the deviation between the actual image and the counterfactual health image. For ease of reproduction, counterfactual difference evidence can be calculated using a multimodal weighted difference, expressed as follows: ; in, This is a counterfactual discrepancy evidence diagram. For modal indexing, For spatially aligned image data, the first The voxel grayscale of each modality, The first in counterfactual health images The voxel grayscale of each modality, Modal weights are determined by the quality grading results. Counterfactual discrepancy evidence maps are normalized and smoothed within a brain region mask, focusing evidence on spatially continuous anomalous areas while suppressing isolated noise.

[0051] When determining candidate regions based on lesion candidate probability maps, thresholding and three-dimensional connected component extraction are first performed on the lesion candidate probability maps to obtain a set of candidate regions. For each candidate region, the centroid and principal axis direction are calculated, and a contralateral mirror control region is generated under the constraint of the midsagittal plane localization result. The contralateral mirror control evidence is used to re-validate the candidate regions using bilateral symmetry. Specifically, the region statistics and structural consistency statistics of counterfactual difference evidence can be extracted from the candidate regions and the contralateral mirror control regions, respectively, and the two are combined to obtain the contralateral mirror control evidence score. When the difference between the candidate region and the contralateral mirror control region is significant and the direction is consistent, the contralateral mirror control evidence is enhanced; otherwise, the contralateral mirror control evidence is reduced to suppress false detections caused by symmetry artifacts or background noise.

[0052] Finally, based on counterfactual difference evidence and contralateral mirror comparison evidence, the lesion candidate probability map is classified into different levels: evidence gating and confidence re-evaluation are performed on the lesion candidate probability map within the candidate region. Regions that simultaneously meet the criteria of high lesion candidate probability, strong counterfactual difference evidence, and strong contralateral mirror comparison evidence are output as lesion masks; regions with moderate lesion candidate probability but insufficient support from counterfactual difference evidence and contralateral mirror comparison evidence, and which meet spatial continuity constraints, are output as suspicious region masks. Before output, hole filling and small connected component removal are performed on the lesion mask and suspicious region mask to obtain topologically continuous segmentation results that are easy to read clinically. By introducing dual evidence constraints of counterfactual difference evidence and contralateral mirror comparison evidence, the segmentation result not only depends on the network probability output but also introduces an interpretable comparison mechanism, thereby reducing false detections caused by stripe artifacts, low signal-to-noise ratio, and alignment residuals, and retaining suspicious regions for further verification in scenarios with unclear boundaries or low contrast lesions.

[0053] In one specific embodiment, diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images are acquired for each subject, and the three types of images are associated using the subject's identifier to form an image dataset. For example... Figure 2 The image shown is an example of an axial image of the same subject under a diffusion-weighted sequence. In the central region of the image, approximately circular low-signal structures are visible on both sides; adjacent to the outer edge of the circular structure on the right, an elliptical high-signal area is visible, forming a clear contrast with the surrounding tissue grayscale, used to simulate the diffusion restriction associated with acute ischemia; the outer periphery of the image shows a background area with gradual grayscale changes and dark edge bands, reflecting the field of view and boundary attenuation during acquisition. Figure 3 As shown, for the same as Figure 2 Examples of apparent diffusion coefficient images corresponding to the same layer. (And...) Figure 2 The elliptical high-signal regions in the image correspond spatially. Figure 3 An elliptical low-signal region is presented at the same location, reflecting the decreasing trend of diffusion-limited areas on the apparent diffusion coefficient map; a nearly circular low-signal structure is still visible at the image center, with relatively uniform surrounding tissue texture, facilitating subsequent cross-modal difference calculations and evidence construction after registration and standardization. Figure 4 As shown, for the same as Figure 2 Example of fluid attenuation inversion recovery image of the same subject at the same level. The approximately circular structure around the center of the image appears as a low signal in this sequence, consistent with the appearance of suppressed fluid signal; a similar structure is visible in the adjacent area on the right. Figure 2 , Figure 3 Elliptical anomalous signal regions corresponding to the same spatial location have identifiable differences in grayscale compared to the surrounding tissues. These differences serve to demonstrate cross-sequence consistency clues and support the joint determination of subsequent "counterfactual difference evidence" and "contralateral mirror comparison evidence."

[0054] In engineering, DICOM sequences output from MRI equipment can be directly received. Pre-registration sequence matching is completed based on subject identification, sequence type labels, and acquisition timestamps. Samples with missing sequences are directly discarded to avoid input inconsistencies in subsequent inference stages. To reduce cross-device differences, brain tissue regions are extracted from each sequence first, then uniformly converted into volumetric data representations in the same spatial coordinate system, and imaging protocol information is recorded for subsequent grouping and calibration.

[0055] Subsequently, spatial orientation uniformity, spatial resampling, and intensity normalization are performed on the image dataset to output standardized image data, and quality grading results are generated simultaneously. For example... Figure 5 As shown, the diffusion-weighted image, apparent diffusion coefficient image, and fluid attenuation inversion recovery image are aligned in the same spatial coordinate system and displayed side by side. All three sub-images in the figure show the same anatomical level with a generally consistent outline and central structural position. The elliptical abnormal region on the right appears in the same spatial location in all three modalities: high signal in the diffusion-weighted image, low signal in the apparent diffusion coefficient image, and a distinguishable abnormal signal from surrounding tissue in the fluid attenuation inversion recovery image. This provides a basis for subsequent calculation of difference maps, generation of candidate regions, and construction of mirror-image control regions on a unified grid. Spatial orientation unification unifies the three images to the same anatomical orientation using a sequence orientation matrix; spatial resampling uses trilinear interpolation to obtain a unified spatial sampling grid, and the three images are resampled to this grid. Intensity normalization is performed within the brain tissue region, using quantile clipping and linear scaling to make the grayscale ranges of different devices comparable. The quality grading results are obtained by combining three achievable indicators: first, the motion artifact indicator is calculated by the decrease in gray-level correlation between adjacent brain slices; second, the signal-to-noise ratio indicator is calculated by the ratio of the brain region mean to the standard deviation of background noise; and third, the diffusion distortion risk indicator is calculated by the degree of boundary misalignment between the diffusion-weighted image and the liquid-attenuated inversion recovered image. These three indicators are mapped using a pre-defined rule to obtain the quality level, which serves as the conditional input for generating the prior model and as one of the index keys for the calibration information set.

[0056] A prior model is trained using standardized image data from a set of normal subjects, and a calibration information set is generated. The prior model is generated using a combination of an encoding network and a rendering network: the encoding network encodes the standardized image data into latent representations; the rendering network uses these latent representations as input to generate normal predicted images. To allow the model to explicitly adapt to different quality levels, the rendering network introduces a quality condition vector at its input. This quality condition vector is obtained from the quality grading results through embedding mapping and modulates the feature channels during rendering, ensuring that the same latent representation can still output normal predicted images consistent with the noise morphology at different quality levels. The calibration information set includes a potential distribution centroid library and a quality calibration curve set. The potential distribution centroid library groups the normal subject set according to imaging protocol information and quality grading results. Within each group, the mean vector of the potential characteristics is calculated as the distribution centroid, and a correspondence between the group identifier and the distribution centroid is established. The quality calibration curve set statistically analyzes three types of distributions on a group basis and generates a gating threshold set. The three types of distributions are the reconstruction residual distribution, the deformation field change distribution, and the counterfactual difference distribution. The gating threshold set is used to constrain the update intensity and update area of ​​the subsequent abnormal editing field to avoid the introduction of large-area false positives by low-quality input.

[0057] For segmented subjects, based on standardized image data, quality grading results, and calibration information sets, the deformation field and anomaly editing field are alternately updated to perform joint inference, resulting in an anomaly editing field and a lesion candidate probability map. In implementation, the deformation field is initialized as a spatial identity mapping, and the anomaly editing field is initialized as an all-zero map. Each iteration includes three key updates: The first stage determines the phase encoding direction information based on the imaging parameters of the diffusion-weighted image and extracts stripe artifact features from the diffusion-weighted image. These stripe artifact features are added as weights to the cost function of the deformation field update, allowing the deformation field update to prioritize the elimination of misalignments in the stripe direction, reducing the misclassification of stripes as lesions. The second stage uses the standardized image data aligned with the current deformation field as input, calls the prior model to obtain a normal prediction image, and calculates the residual map as anomaly cues. The third stage updates the anomaly editing field under the constraints of a gating threshold set, ensuring that the anomaly editing field only grows in regions with significant residuals and consistent with the quality grade. Simultaneously, the prior centroids provided by the potential distribution centroid library constrain the drift of potential representations, preventing the model from being led by real lesions and "learning lesions." The lesion candidate probability map can be obtained by fusing the residual map, cross-modal consistency features and the output of a lightweight segmentation network. In engineering, it can be implemented using a two-dimensional or three-dimensional convolutional network with three-modal channel input.

[0058] In the output stage, after zeroing the anomalous editing field, a counterfactual healthy image is generated by the generative prior model. Based on this, two types of evidence are constructed, and a lesion candidate probability map is fused to output a lesion mask and a suspicious region mask. The counterfactual difference evidence is obtained from the difference map between standardized image data and the counterfactual healthy image, used to characterize the necessary modification amount "from anomalous to healthy". The contralateral mirror comparison evidence is obtained by determining candidate regions from spatially aligned image data, then generating a contralateral mirror comparison region around the midline and calculating the difference, used to suppress false alarms caused by symmetrical structures and imaging noise. The two types of evidence are first normalized according to the quality calibration curve set to obtain a standardized evidence score, which is then fused with the lesion candidate probability map. ; in, This is a probability map of the merged lesions. This is a probability map of candidate lesions. For counterfactual difference evidence scores, The evidence score is the contralateral mirror comparison score. , , To integrate weights, This is the Sigmoid function. Based on the two-level thresholds given by the set of quality calibration curves, [the following is applied]: Thresholding yields lesion masks and suspicious area masks, allowing high-confidence areas to enter the lesion mask and boundaries and low-confidence areas to enter the suspicious area mask, facilitating clinical review.

[0059] The efficacy was validated using 60 samples with lesion masking. The quality grade distribution was as follows: Grade 1 (15 cases), Grade 2 (34 cases), and Grade 3 (11 cases). The Dice coefficient, 95th percentile Hausdorff distance, and false positive volume percentage were calculated. Figure 6 The figure shows the statistical results of the quality level distribution of the 60 samples used for effect verification in this embodiment. The horizontal axis represents the quality level, and the vertical axis represents the number of cases; quality level 1 corresponds to 15 cases, quality level 2 corresponds to 34 cases, and quality level 3 corresponds to 11 cases, reflecting that the samples are mainly concentrated in quality level 2, while a certain proportion of low-quality samples are retained to verify the robustness of the gating threshold set and evidence fusion strategy in low-quality scenarios.

[0060] Using the control segment as a baseline, the average Dice coefficient obtained in this embodiment increased from 0.619 to 0.717, an absolute increase of 0.098 and a relative increase of approximately 15.8%. Figure 7The figure shows the box plot statistics of the Dice coefficients for the control segmentation and the segmentation of this invention on the same test set. The horizontal axis represents the two segmentation result categories, and the vertical axis represents the Dice coefficient. The control segmentation's box positions are generally lower, with greater dispersion, and the lower bound extends to a lower value range. The segmentation of this invention's box positions are generally higher, with a significantly increased median and convergent box height, indicating improved overlap and stability on most samples. The mean 95th percentile Hausdorff distance decreased from 12.11 mm to 8.03 mm, a relative decrease of approximately 33.7%. The false positive volume ratio decreased from 0.247 to 0.194, a relative decrease of approximately 21.5%. When observing by quality level, the Dice for quality level 1 increased from 0.704 to 0.748; the Dice for quality level 2 increased from 0.618 to 0.724; and the Dice for quality level 3 increased from 0.506 to 0.654. Figure 8 As shown, this is a statistical result of the mean and fluctuation range of the Dice coefficient grouped by quality level. The horizontal axis represents the quality level, and the vertical axis represents the mean Dice coefficient. The control segmentation curve shows a significant downward trend as the quality level deteriorates, and the error bars are longer at quality levels 2 and 3, indicating greater fluctuations in low-quality scenarios. The segmentation curve of this invention is higher than the control segmentation at all quality levels, especially at quality levels 2 and 3, demonstrating that the combination of gating threshold set to suppress noise propagation and counterfactual evidence and contralateral mirror evidence to suppress false positives can still maintain high consistency under low-quality input. It can be seen that the gain is more significant in poor-quality scenarios, which is consistent with the design goals of gating threshold set to suppress noise propagation and counterfactual evidence and contralateral evidence to suppress false positives. Figure 9 The figure shows the distribution and segmented mean trend of the Dice coefficient of the segmentation results of this invention in different lesion volume ranges. The horizontal axis represents the lesion volume on a logarithmic scale, and the vertical axis represents the Dice coefficient of the segmentation results of this invention. The scatter plots represent the results of a single case. It can be seen that the scatter plots are relatively more dispersed in the small volume range, while the scatter plots are more concentrated in the medium volume range. The broken line connects the mean points after segmentation by volume. Overall, it fluctuates in the range of approximately 0.70 to 0.75, indicating that this invention can maintain a relatively stable segmentation overlap at different volume scales. At the same time, for a very small number of abnormal cases, the "suspicious area mask" can still guide the review.

[0061] like Figure 10 As shown, a deep learning-based stroke image segmentation system is used to implement a deep learning-based stroke image segmentation method. The system includes: The image acquisition module is used to acquire diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images to form an image dataset. Hardware-wise, the image acquisition module typically consists of a magnetic resonance imaging (MRI) device, a sequence control unit, and a data interface unit. The MRI device includes a main magnet, gradient coils, RF transmitting and receiving coils, and an RF power amplifier and receiving link. The sequence control unit schedules diffusion-weighted acquisition, apparent diffusion coefficient imaging, and liquid attenuation inversion recovery acquisition under the same subject identification, and reconstructs the original acquisition data of each sequence into image data. The data interface unit writes the images and their scanning parameters to a storage array or transmits them via the hospital network to a back-end processing server, thereby forming an image dataset organized by subject, sequence type, and timestamp.

[0062] The standard modeling module performs spatial resampling and intensity normalization on the image dataset to generate standardized image data and quality grading results. It then trains a prior model based on standardized image data from a set of normal subjects, generating a calibration information set, including a latent distribution centroid library and a set of quality calibration curves. The standard modeling module is typically deployed on an image post-processing workstation or server node, consisting of a central processing unit (CPU), a graphics processing unit (GPU), memory, and high-speed storage. The CPU is responsible for image reading / writing, spatial coordinate and orientation information parsing, and resampling and normalization task scheduling. The GPU accelerates tensor computation during convolutional network inference and training. Memory caches standardized images and intermediate features. High-speed storage stores quality grading results, generated prior model parameters, and the calibration information set. This module can access the image archiving and communication system via a local area network to read the set of normal subjects and complete model training, calibration curve statistics, and result persistence on the server side.

[0063] The joint inference module is used to perform joint inference on the deformation field and the anomaly editing field of the subject to be segmented, based on standardized image data, quality grading results, and calibration information set, to obtain the anomaly editing field and the lesion candidate probability map. In terms of hardware, the joint inference module typically runs as an inference service on a computing node equipped with a graphics processing unit (GPU), and is equipped with a task queue and caching components. After receiving the standardized image data, quality grading results, and calibration information set, the module performs forward computation to generate the prior model on the GPU, and completes iterative control, convergence determination, and anomaly rollback of the deformation field and the anomaly editing field during the alternating update process with the cooperation of the central processing unit (CPU). Memory is used to store intermediate tensors such as the deformation field, the anomaly editing field, and the candidate probability map. The caching component is used to reuse the calibration information and prior centroid index of multiple inferences for the same subject, thereby ensuring the stable operation and data consistency of the iterative inference process.

[0064] The evidence output module generates counterfactual health images based on a generative prior model under zero-based conditions in the anomalous editing field, obtaining counterfactual difference evidence and contralateral mirror comparison evidence. It then fuses lesion candidate probability maps to output lesion masks and suspicious region masks. Hardware-wise, the evidence output module is typically deployed on the same machine as the joint inference module or as a separate service. Its key resources are the graphics processing unit's computing power, storage, and visualization output interface. This module generates counterfactual health images and calculates difference evidence on the graphics processing unit, while simultaneously constructing the contralateral mirror comparison region, fusing evidence, and performing hierarchical mapping on the central processing unit, ultimately outputting lesion masks and suspicious region masks. The output interface can be an image archiving and communication system write-back interface, a reporting system interface, or a local display interface, used to overlay the masks back into the original sequence in a standard medical image format and record evidence scores, supporting clinical review and record keeping.

[0065] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A deep learning-based stroke image segmentation method, characterized in that, Includes the following steps: Acquire diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images to form an image dataset; Spatial resampling and intensity normalization are performed on the image dataset to generate standardized image data and quality grading results. A prior model is trained based on the standardized image data of the normal subject set to generate a calibration information set, which includes a potential distribution centroid library and a quality calibration curve set. For segmented subjects, based on the standardized image data, the quality grading results and the calibration information set, the deformation field and the abnormal editing field are alternately updated to perform joint inference, and the abnormal editing field and lesion candidate probability map are obtained. Based on the aforementioned prior model, a counterfactual health image is generated under the condition of zeroing the abnormal editing field, thereby obtaining counterfactual difference evidence and contralateral mirror comparison evidence. The lesion candidate probability map is then fused to output a lesion mask and a suspicious region mask.

2. The method according to claim 1, characterized in that, Obtaining the diffusion-weighted image, the apparent diffusion coefficient image, and the liquid attenuation inversion recovery image includes: Acquire the diffusion-weighted image, the apparent diffusion coefficient image, and the liquid attenuation inversion recovery image that have the same subject identifier, and associate the diffusion-weighted image, the apparent diffusion coefficient image, and the liquid attenuation inversion recovery image into the image dataset.

3. The method according to claim 1, characterized in that, Spatial resampling and intensity normalization of the image dataset includes: Spatial orientation uniformity is performed on the image dataset; Spatial resampling is performed on the image dataset to obtain a uniform spatial sampling grid; Intensity normalization is performed on the image dataset on a uniform spatial sampling grid to obtain the standardized image data.

4. The method according to claim 1, characterized in that, Generating the quality grading results includes: Motion artifact index, signal-to-noise ratio index, and diffusion distortion risk index are calculated based on the standardized image data, and the motion artifact index, the signal-to-noise ratio index, and the diffusion distortion risk index are mapped to the quality grading result based on the grading mapping rule.

5. The method according to claim 1, characterized in that, Training the generative prior model includes: Train a latent representation encoding network to encode the standardized image data into latent representations; A modal conditional rendering network is trained to generate normal prediction images from the latent representations. The modal conditional rendering network generates quality condition vectors based on the quality grading results and modulates the generation of normal prediction images based on the quality condition vectors.

6. The method according to claim 1, characterized in that, Generating the potential distribution centroid library includes: The normal subject set is grouped based on the imaging protocol information of the image dataset and the quality grading results. The standardized image data for each group is input into the generative prior model to obtain the potential representation; The distribution centroid is calculated for the potential representations of each group, and a correspondence is established between the group identifier and the distribution centroid to obtain the potential distribution centroid library.

7. The method according to claim 1, characterized in that, Generating the set of quality calibration curves includes: Under each of the aforementioned quality grading results, the reconstruction residual distribution of the generated prior model is statistically analyzed based on the set of normal subjects; On the set of normal subjects, perform joint inference of alternatingly updated deformation field and abnormal edit field to obtain abnormal edit field, and statistically analyze the distribution of deformation field change. Based on the aforementioned prior model, counterfactual health images are generated under the condition that the anomalous editing field is set to zero, and the counterfactual difference distribution is statistically analyzed. The set of gate thresholds is determined based on the reconstructed residual distribution, the deformation field change distribution, and the counterfactual difference distribution, forming the set of quality calibration curves.

8. The method according to claim 1, characterized in that, Joint inference is performed by alternately updating the deformation field and the anomaly editing field, including: Based on the quality grading results, the prior centroid corresponding to the subject to be segmented is determined in the potential distribution centroid library, and the set of gate thresholds is determined based on the set of quality calibration curves. Phase encoding direction information is determined based on the imaging parameter information of the diffusion-weighted image, and stripe artifact features are extracted from the diffusion-weighted image. During the alternating iteration process, the deformation field is updated based on the stripe artifact features to align the standardized image data. A normal prediction image is generated based on the generated prior model. The abnormal editing field is updated under the constraints of the gating threshold set based on the residual between the standardized image data and the normal prediction image. The potential representation is updated based on the prior centroid constraint to obtain the abnormal editing field and the lesion candidate probability map.

9. The method according to claim 1, characterized in that, The lesion candidate probability map is fused to output the lesion mask and the suspicious region mask, including: Spatial alignment is performed on the standardized image data based on the deformation field to obtain spatially aligned image data; The counterfactual health image is generated based on the aforementioned prior model under the condition that the abnormal editing field is set to zero. The counterfactual discrepancy evidence is obtained based on the spatially aligned image data and the counterfactual health image. Based on the lesion candidate probability map, candidate regions are determined and corresponding contralateral mirror control regions are generated to obtain the contralateral mirror control evidence. The candidate probability map of the lesion is classified and graded based on the counterfactual difference evidence and the contralateral mirror comparison evidence to output the lesion mask and the suspicious area mask.

10. A deep learning-based stroke image segmentation system, used to implement the deep learning-based stroke image segmentation method according to any one of claims 1-9, characterized in that, The system includes: The image acquisition module is used to acquire diffusion-weighted images, apparent diffusion coefficient images, and liquid attenuation inversion recovery images to form an image dataset. The standard modeling module is used to spatially resample and normalize the intensity of the image dataset to generate standardized image data and quality grading results; and to train and generate a prior model based on the standardized image data of the normal subject set, generating a calibration information set, which includes a potential distribution centroid library and a quality calibration curve set. The joint inference module is used to perform joint inference on the subject to be segmented based on standardized image data, quality grading results and calibration information set, alternately updating the deformation field and the abnormal editing field to obtain the abnormal editing field and the lesion candidate probability map. The evidence output module is used to generate counterfactual health images based on the generative prior model under the condition of zeroing the abnormal editing field, obtain counterfactual difference evidence and contralateral mirror comparison evidence, and output lesion mask and suspicious area mask by fusing lesion candidate probability map.

Images