Method for detecting type of vascular occlusion and related devices

Abstract

This application discloses a method and related equipment for detecting vascular occlusion types, relating to the field of medical imaging technology. The method includes: performing multimodal preprocessing on non-enhanced CT images and arterial phase CTA images to obtain multi-channel image tensors; encoding structured clinical text data to obtain clinical text features; performing image generation processing based on the multi-channel image tensors to obtain CT-like perfusion parameter maps; performing cross-modal fusion of clinical text features, diffusion features extracted based on CT-like perfusion parameter maps, and image features extracted based on multi-channel image tensors to obtain fused features; and determining the classification probability for each vascular occlusion type based on the fused features. This application enables multi-dimensional feature collaborative analysis, improving the identification ability of different vascular occlusion types and the reliability of classification results, providing more accurate and comprehensive auxiliary decision-making basis for clinical diagnosis.

Description

Technical Field

[0001] This application relates to the field of medical imaging technology, and more specifically, to a method and related equipment for detecting the type of vascular occlusion. Background Technology

[0002] With the continuous development of medical imaging and artificial intelligence technologies, image-based auxiliary diagnostic techniques for cerebrovascular diseases have been widely applied in clinical stroke diagnosis and treatment. Acute ischemic stroke, a cerebrovascular disease with high incidence, high disability rate, and high mortality, relies heavily on the rapid and accurate identification of vascular occlusion types for its treatment outcomes. In clinical practice, non-contrast CT scans are typically used to rule out cerebral hemorrhage, combined with arterial phase CT angiography (CTA) to assess cerebral vascular occlusion and assist physicians in developing thrombolytic or endovascular treatment plans. With the improvement of medical informatization, patients' symptom descriptions, scoring indicators, and medical history are gradually being structured and stored, providing a richer data foundation for the intelligent identification of vascular occlusion types.

[0003] However, most existing technologies for detecting vascular occlusion types rely on single-image modalities for analysis, such as using only non-contrast CT or CTA images to extract features for classification. Since non-contrast CT primarily reflects changes in brain tissue structure, and CTA mainly reflects vascular visualization, the information provided by a single modality has limitations. It is all time-window information and cannot comprehensively characterize the patient's pathological state and brain perfusion information, which is particularly useful for identifying small and medium-sized vessel occlusions. Furthermore, some vascular occlusions, especially smaller ones, are not typical in their imaging presentation and are easily affected by individual anatomical differences, image quality, and collateral circulation, thus limiting the accuracy of the detection results. In other words, related technologies suffer from poor identification of different types of vascular occlusion, especially different degrees of occlusion, during the detection process. Summary of the Invention

[0004] In the summary section of this application, the relevant technical solutions are described in general terms, and a series of simplified concepts are introduced. These concepts will be further elaborated in the detailed embodiments section. This summary section should not be construed as limiting the key or essential technical features of the claimed solutions, nor is it intended to limit the scope of protection of the claimed solutions.

[0005] The vascular occlusion type detection method and related equipment provided in this application can construct a multimodal cross-modal analysis mechanism by integrating non-enhanced CT, arterial phase CTA, CT-like perfusion parameter maps and clinical text information. This enables comprehensive characterization and accurate identification of vascular occlusion types, thereby improving the accuracy, reliability and clinical decision-making reference value of classification results.

[0006] In a first aspect, this application provides a method for detecting vascular occlusion types, comprising: performing multimodal preprocessing on non-enhanced CT images and arterial phase CTA images to obtain multi-channel image tensors; encoding structured clinical text data to obtain clinical text features; performing image generation processing based on the multi-channel image tensors to obtain a CT-like perfusion parameter map, wherein the CT-like perfusion parameter map includes a cerebral blood flow map, a cerebral blood volume map, and a mean transit time map; performing cross-modal fusion on the clinical text features, diffusion features extracted based on the CT-like perfusion parameter map, and image features extracted based on the multi-channel image tensors to obtain fused features; and determining the classification probability for vascular occlusion types based on the fused features, wherein the vascular occlusion types include large vessel occlusion, medium vessel occlusion, and no occlusion, wherein large vessel occlusion is an occlusion involving a vessel diameter greater than a first preset threshold, and medium vessel occlusion is an occlusion involving a vessel diameter less than or equal to the first preset threshold and greater than a second preset threshold.

[0007] In some embodiments, the multimodal preprocessing of the non-enhanced CT image and the arterial phase CTA image to obtain a multi-channel image tensor includes: resampling the non-enhanced CT image and the arterial phase CTA image based on a preset isotropic voxel size to obtain a first resampled image and a second resampled image; rigidly registering the second resampled image with the first resampled image as a fixed reference image to obtain a registered CTA image; performing skull stripping on the first resampled image and the registered CTA image to obtain a first brain parenchyma image and a second brain parenchyma image; performing preset window width and window level truncation and linear normalization on the first brain parenchyma image and the second brain parenchyma image to obtain a first normalized image and a second normalized image; and superimposing the first normalized image and the second normalized image in the channel dimension to obtain the multi-channel image tensor.

[0008] In some implementations, encoding structured clinical text data to obtain clinical text features includes: extracting continuous and discrete fields from the structured clinical text data; normalizing the continuous fields to obtain a first processing result; performing one-hot encoding on the discrete fields to obtain a second processing result; concatenating the first and second processing results and mapping them through a preset multilayer perceptron to obtain a first text feature; performing natural language conversion on the structured clinical text data to obtain clinical descriptive text; linearly mapping the semantic embedding vector extracted from the clinical descriptive text through a preset adapter network to obtain a second text feature; and weightedly fusing the first and second text features to obtain the clinical text feature.

[0009] In some implementations, the image generation process based on the multi-channel image tensor to obtain a CT-like perfusion parameter map includes: inputting the multi-channel image tensor as a conditional input to the inverse denoising network of a preset conditional diffusion model; and generating the CT-like perfusion parameter map by performing multi-step iterative denoising from initial random Gaussian noise guided by the conditional input through the inverse denoising network, wherein the preset conditional diffusion model uses historical CT perfusion parameter maps as diffusion targets during the training phase, and the inverse denoising network is used to learn the conditional inverse mapping from the noise distribution to the perfusion parameter map distribution.

[0010] In some embodiments, before performing cross-modal fusion of the clinical text features, diffusion features extracted based on the CT-like perfusion parameter map, and image features extracted based on the multi-channel image tensor to obtain fused features, the vascular occlusion type detection method further includes: performing three-dimensional feature extraction based on the multi-channel image tensor using a first preset convolutional network to obtain the image features; performing three-dimensional feature extraction of the CT-like perfusion parameter map using a second preset convolutional network to obtain the diffusion features; the process of fusing the clinical text features, diffusion features extracted based on the CT-like perfusion parameter map, and image features extracted based on the multi-channel image tensor... Image features extracted from image tensors are fused across modally to obtain fused features, including: mapping the image features, diffusion features, and clinical text features to a preset unified feature dimension to obtain a first mapped feature, a second mapped feature, and a third mapped feature; using any one of the first, second, and third mapped features as a query, and the other two mapped features as keys and values, cross-modal attention calculations are performed to obtain attention-weighted features corresponding to each modality; and the attention-weighted features corresponding to each modality are concatenated along the channel dimension to obtain the fused features.

[0011] In some implementations, determining the classification probability for the type of vascular occlusion based on the fusion features includes: performing a global average pooling operation on the fusion features to obtain a pooled feature vector; inputting the pooled feature vector into a preset fully connected classification layer to obtain a classification score vector; and performing normalized exponential processing on the classification score vector to obtain the classification probability that the object to be detected belongs to large vessel occlusion, moderate vessel occlusion, or no occlusion.

[0012] In some implementations, after determining the classification probability for a type of vascular occlusion based on the fusion features, the vascular occlusion type detection method further includes: obtaining the target predicted type with the highest probability value among the classification probabilities; if the probability value corresponding to the target predicted type is greater than a preset confidence threshold, determining the target predicted type as the target classification result; weighting the feature map of the last convolutional layer in the first preset convolutional network used to extract the image features based on the target predicted type to generate a spatial activation map; upsampling the spatial activation map to the original input image space to obtain an upsampled activation map; superimposing the upsampled activation map with the vascular enhancement image of the arterial phase CTA image to obtain a lesion localization heatmap; and encapsulating the target classification result, the classification probability of the object to be detected belonging to each type of vascular occlusion, and the lesion localization heatmap into a structured report.

[0013] Secondly, this application also provides a vascular occlusion type detection device, comprising: a data acquisition unit for performing multimodal preprocessing on non-enhanced CT images and arterial phase CTA images to obtain multi-channel image tensors; a text encoding unit for encoding structured clinical text data to obtain clinical text features; an image diffusion unit for performing image generation processing based on the multi-channel image tensors to obtain a CT-like perfusion parameter map, wherein the CT-like perfusion parameter map includes a cerebral blood flow map, a cerebral blood volume map, and a mean transit time map; a feature fusion unit for performing cross-modal fusion of the clinical text features, diffusion features extracted based on the CT-like perfusion parameter map, and image features extracted based on the multi-channel image tensors to obtain fused features; and a probability determination unit for determining the classification probability for a vascular occlusion type based on the fused features, wherein the vascular occlusion type includes large vessel occlusion, medium vessel occlusion, and no occlusion, wherein large vessel occlusion is an occlusion involving a vessel diameter greater than a first preset threshold, and medium vessel occlusion is an occlusion involving a vessel diameter less than or equal to the first preset threshold and greater than a second preset threshold.

[0014] Thirdly, this application also provides an electronic device, including: a memory and a processor, wherein the processor is configured to execute a computer program stored in the memory to implement the steps of the vascular occlusion type detection method described in the first aspect.

[0015] Fourthly, this application also provides a computer-readable storage medium storing computer-executable instructions or a computer program, which, when executed by a processor, implement the steps of the blood vessel blockage type detection method described in the first aspect.

[0016] Fifthly, this application also provides a computer program product, including a computer program or computer executable instructions, which, when executed by a processor, implement the steps of the blood vessel blockage type detection method provided in the embodiments of this application.

[0017] In summary, this application first performs multimodal preprocessing on non-enhanced CT images and arterial phase CTA images, and constructs multi-channel image tensors. This allows anatomical and vascular imaging information from both types of image data to be co-expressed in a unified space. Compared to analysis using only single image data, multi-channel image tensors can simultaneously preserve brain parenchymal changes and vascular status features, thus providing more comprehensive imaging evidence for subsequent vascular occlusion type determination and improving the ability to distinguish between different vascular occlusion types. Based on multi-channel image tensors, image generation processing is performed to obtain CT-like perfusion parameter maps, including cerebral blood flow maps, cerebral blood volume maps, and mean transit time maps. Since these CT-like perfusion parameter maps can reflect the perfusion status of brain tissue from another perspective, hemodynamic features are further supplemented on the original structural and vascular image information. By extracting diffusion features from the CT-like perfusion parameter maps, the model can consider multi-dimensional information such as structure, blood vessels, and perfusion when determining the type of vascular occlusion, thereby enhancing the ability to express the differences between different occlusion types. By utilizing image data, structured clinical text data is encoded to obtain clinical text features. These clinical text features are then fused across modalities with image and diffusion features. Since clinical text data includes information such as patient symptoms and scoring indicators, it can supplement the image analysis results from a clinical perspective. Through the cross-modal fusion mechanism, correlation modeling between information from different sources can be achieved, enabling the model to comprehensively utilize image and clinical evidence when determining the type of vascular occlusion, thereby improving the reliability and stability of the classification results. Based on the fused features, the classification probability of the vascular occlusion type corresponding to the subject is determined, and the detection results are subdivided into three types: large vessel occlusion, medium vessel occlusion, and no occlusion. Among them, the fused features formed by fusing image, diffusion, and clinical text features can more comprehensively represent the state, thus improving the ability to identify different types of occlusion. At the same time, outputting classification probabilities rather than single classification results can also reflect the degree of matching between different types, providing richer reference information for subsequent diagnostic decisions and improving the practical value of the detection results. In summary, the vascular occlusion type detection method provided in this application integrates non-enhanced CT, arterial phase CTA, CT-like perfusion parameter maps, and clinical text information to construct a multimodal cross-modal analysis mechanism, which can achieve comprehensive characterization and accurate identification of vascular occlusion types, thereby improving the accuracy, reliability, and clinical decision-making reference value of the classification results. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A schematic flowchart illustrating a method for detecting the type of vascular occlusion provided in this application embodiment; Figure 2 This is a schematic diagram of the composition structure of a blood vessel blockage type detection device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the composition structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] The terms used in the specification, claims, and drawings of this application, such as "first," "second," "third," "fourth," etc. (if any), are used to distinguish similar objects and not to describe a specific order or sequence. Therefore, it is to be understood that these terms can be used interchangeably where appropriate, allowing the described embodiments to be used in different orders, unless specifically required by the illustrations or description. Furthermore, the terms "is" and "has," and any variations thereof, are intended to cover, non-exclusively, all possible constituent elements. For example, a process, method, system, product, or apparatus comprising several steps or units is not necessarily limited to the steps or units explicitly listed, but may also include other steps or units not explicitly listed, or steps or units inherent to the process, method, product, or apparatus.

[0020] In this application, a "module" or "unit" refers to a computer program or part of a computer program that has a specific function and works in conjunction with other related parts to achieve a predetermined goal. These modules or units can be implemented by software, hardware (e.g., processing circuitry or memory), or a combination of both. One or more processors or memories can implement one or more modules or units. Furthermore, each module or unit can also be part of a larger module or unit.

[0021] The technical solutions of this application will be described in detail below with reference to the accompanying drawings of the embodiments. It should be noted that the described embodiments are only a part of this application, and not all embodiments. In the following description, the "some embodiments" mentioned are only a subset of all possible embodiments, which may be the same or different subsets, and different embodiments can be combined with each other without conflict.

[0022] Figure 1 This is a schematic flowchart of a method for detecting the type of vascular occlusion provided in an embodiment of this application. For example, see [link to example]. Figure 1 The method for detecting the type of vascular occlusion provided in this application embodiment may include the following steps 101 to 105: Step 101: Perform multimodal preprocessing on the non-enhanced CT images and arterial phase CTA images to obtain multi-channel image tensors.

[0023] In some examples, non-contrast CT (Computed Tomography) refers to plain brain scans without iodine contrast agent injection. These images are acquired using a hospital's clinical spiral CT scanner, and the raw images are stored in the hospital's image archiving and communication system in a medical digital imaging and communication format. For example, it could be a complete tomographic sequence of images obtained from a patient's first plain brain scan upon admission to the emergency department. Arterial phase CTA (Computed Tomography Angiography) images are tomographic images of the brain's blood vessels acquired during the peak contrast agent filling period of intracranial arteries after intravenous injection of iodine contrast agent. Clinical scans use a contrast agent infusion rate of 4 to 5 ml / s, and the raw images are simultaneously archived to the hospital's image storage system. For example, it could be brain arterial imaging images completed immediately after a non-contrast CT scan of the same patient. Multimodal preprocessing is a standardized image processing workflow performed on non-enhanced CT images and arterial phase CTA images. It mainly includes spatial correction, irrelevant tissue removal, and image grayscale normalization. The entire process relies on the SimpleITK and ANTs open-source image processing tool libraries. In practice, the MNI152 standard human brain template is used as the registration and normalization benchmark. The multi-channel image tensor is a three-dimensional digital feature array obtained by superimposing and fusing two types of preprocessed single-channel cranial images along the channel dimension. It is generated by stitching together normalized non-enhanced CT images and arterial phase CTA images. In practice, it is a dual-channel three-dimensional image tensor with unified voxel space and pixel grayscale values ​​regularized to the 0-1 range.

[0024] For example, a dedicated in-hospital data interface can be used to automatically retrieve the patient's original non-contrast computed tomography (CT) images and arterial phase CT angiography images from the image archiving and communication system. Using the MNI152 standard human brain template as a unified benchmark, resampling and affine transformation are performed on both types of images to achieve spatial standardization and precise anatomical alignment. Subsequently, the removal of irrelevant tissues such as the skull, scalp, and ventricles is automatically completed to eliminate interference from non-brain tissue regions on the detection results. Finally, dedicated window width and level truncation and linear normalization are performed on the two types of images respectively to unify the image grayscale value range. Finally, a standard multi-channel image tensor is obtained by superimposing and synthesizing the images through channel dimensions, completing the preprocessing operation of this step.

[0025] By implementing step 101, multimodal preprocessing is performed on non-enhanced CT images and arterial phase CTA images, and a multi-channel image tensor is constructed. This enables the unified expression of brain parenchymal structural information and vascular imaging information, making different modal images complementary. It can simultaneously retain structural features reflecting the state of brain tissue and vascular features reflecting vascular patency, thereby providing more comprehensive imaging evidence for subsequent identification of vascular occlusion types and improving the distinguishability between different types of vascular occlusion.

[0026] Step 102: Encode the structured clinical text data to obtain clinical text features.

[0027] In some examples, structured clinical text data refers to standardized, structured diagnostic and treatment field data for patients with acute ischemic stroke stored in the hospital's Electronic Medical Record (EMR) system. This data can be automatically retrieved through the hospital's RESTful Application Programming Interface (API) without manual input. It mainly includes fixed fields such as the time from onset to imaging, the National Institutes of Health Stroke Scale score, the patient's clinical symptoms, level of consciousness, and past medical history. For example, it could be standardized text data recording the time from onset to CT scan, stroke score, hemiplegia symptoms, atrial fibrillation, and history of hypertension.

[0028] Clinical text features are quantitative feature vectors generated after standardizing and encoding the original structured clinical text data. They are used to quantitatively represent the patient's clinical condition and stroke pathogenesis. These features are obtained by the system through multi-dimensional encoding and fusion processing of the original clinical text data. For example, they can be digital text features with unified dimensions that can be recognized by the model, which integrate the patient's quantitative score, onset time, clinical symptoms and past medical history information.

[0029] For example, after image preprocessing is completed, the system automatically links with the hospital's electronic medical record system to capture all structured clinical text data of the corresponding patient; it automatically identifies and distinguishes between continuous and discrete data fields in the text data, and performs standardized preprocessing operations accordingly; at the same time, it converts the structured data into standard clinical descriptive text, mines deep semantic information, and completes feature mapping and fusion; finally, it generates clinical text features with unified dimensions and complete feature information, providing standardized clinical data support for subsequent multimodal feature fusion and vascular occlusion classification.

[0030] By implementing step 102, the structured clinical text data is encoded to obtain clinical text features, which can convert the patient's symptoms, scoring indicators and other clinical information into feature expressions that can be used by the model. Since clinical information can reflect the characteristics of the disease from the perspective of the patient's functional status and disease manifestations, it can serve as an important supplement to imaging information, so that subsequent analysis not only relies on imaging manifestations, but also combines clinical evidence for comprehensive judgment, thereby improving the reliability of the vascular occlusion type identification results.

[0031] Step 103: Perform image generation processing based on multi-channel image tensors to obtain a CT-like perfusion parameter map.

[0032] Among them, CT-like perfusion parameter maps can include cerebral blood flow maps, cerebral blood volume maps, and mean transit time maps.

[0033] In some examples, the Computed Tomography Perfusion (CTP) image is a composite parametric image obtained by simulating the perfusion imaging effect of standard clinical computed tomography scans. The entire image set does not require additional perfusion scans on the patient and is calculated using the multi-channel image tensors processed previously combined with the image generation algorithm. For example, it can be a set of three-dimensional images suitable for the diagnosis of acute ischemic stroke, reflecting the characteristics of intracranial blood flow distribution. The image generation algorithm can employ various deep learning image generation models, including but not limited to generative adversarial networks, diffusion models, diffusion language models, and flow matching models. Furthermore, any permutation and combination of the above models can be used to complete the image generation task. The embodiments in this application are not limited to a single diffusion model implementation.

[0034] Cerebral blood flow (CBF) is a component of CT-like perfusion parameter maps. It is used to quantify the total amount of blood flowing through the corresponding brain tissue region per unit time. It is a key imaging basis for determining whether ischemic changes have occurred in brain tissue. This image is generated synchronously during the diffusion generation operation; for example, it can be a pixelated image that can intuitively distinguish between normally blood-supplying brain regions and low-flow ischemic brain regions.

[0035] Cerebral blood volume (CBV) is also a component of CT-like perfusion parameter maps. It is used to reflect the total volume of blood contained in the vascular bed within local brain tissue and can help define the extent of the impact of vascular abnormalities. This image is generated simultaneously with the cerebral blood flow map; for example, it can be a specialized parameter image that can mark areas of decreased blood volume due to vascular occlusion.

[0036] Mean Transit Time (MTT) is an important component of CT-like perfusion parameter maps, used to reflect the average transit time of the tracer medium within the vascular network of brain tissue. Vascular occlusion areas will show the typical characteristic of prolonged transit time. This image is output based on a unified diffusion generation process; for example, it can be a functional image that can identify areas of delayed blood flow and help distinguish different degrees of occlusion.

[0037] For example, the multi-channel image tensor output in step 102 can be used as a fixed condition input to a pre-trained conditional diffusion model. The model performs multiple rounds of iterative denoising operations based on the input image information to gradually restore the corresponding blood flow parameter images. After the operation is completed, the cerebral blood flow map, cerebral blood volume map and mean transit time map are integrated to form a complete CT-like perfusion parameter map. The entire generation process can be achieved using only existing image data without adding clinical scanning procedures and examination consumables.

[0038] By implementing step 103, image generation processing is performed based on multi-channel image tensors to generate CT-like perfusion parameter maps, including cerebral blood flow maps, cerebral blood volume maps, and mean transit time maps. This allows for the acquisition of hemodynamic information reflecting the perfusion status of brain tissue based on the original structural and vascular image information. Since the CT-like perfusion parameter maps characterize the blood supply to brain tissue from a perfusion perspective, they help to enrich feature expression, enabling subsequent models to simultaneously utilize multi-dimensional information such as structure, blood vessels, and perfusion for analysis, thereby enhancing the ability to express the differential features between different types of vascular occlusion.

[0039] Step 104: Perform cross-modal fusion of clinical text features, diffusion features extracted based on CT-like perfusion parameter maps, and image features extracted based on multi-channel image tensors to obtain fused features.

[0040] In some examples, diffusion features are hemodynamic depth features mined from CT-like perfusion parameter maps generated by diffusion. These features can intuitively reflect the distribution and abnormal changes of blood flow perfusion in brain tissue. The features are obtained by performing three-dimensional feature extraction operations on the CT-like perfusion parameter maps through a dedicated convolutional network. This can filter out invalid background information in the images and retain the perfusion patterns related to vascular occlusion. For example, it can be a digital feature vector that can reflect pathological features such as reduced blood flow and delayed blood flow in local brain regions.

[0041] Image features are anatomical structures and vascular morphological features extracted from preprocessed multi-channel image tensors. They mainly carry imaging information such as the morphology of cranial tissues, the course of blood vessels, and changes in the morphology of blood vessel lumens. These features are generated by three-dimensional feature extraction based on multi-channel image tensors using independent convolutional networks, focusing on the abnormal morphological manifestations of blood vessels themselves. For example, they can be structured image feature vectors that can characterize occlusion morphologies such as intracranial vascular stenosis and vascular interruption.

[0042] Fusion features are integrated feature representations formed by combining information from three different sources. They comprehensively gather effective information from three dimensions: patient clinical condition, vascular morphology, and cerebral blood flow perfusion. This feature is obtained by performing cross-modal fusion processing on three independent features and has the ability to jointly characterize vascular lesions in multiple dimensions. For example, it can be a comprehensive feature vector that simultaneously includes information such as stroke score, abnormal vascular morphology, and local perfusion defects in a unified dimension.

[0043] The process of cross-modal fusion of clinical text features, diffusion features extracted based on CT-like perfusion parameter maps, and image features extracted based on multi-channel image tensors aims to break down information barriers between different modalities of data, establish the intrinsic relationship between clinical information, perfusion image information, and anatomical image information, and integrate the three independent features into a complete feature system. This operation completes feature association and recombination according to a preset fusion logic, thereby amplifying the feature differences between different types of vascular occlusion and providing comprehensive data support for subsequent classification and judgment.

[0044] For example, the corresponding convolutional network can be started first to extract diffusion features and image features respectively; then the clinical text features, diffusion features, and image features are mapped to the same feature dimension to ensure that the data formats of multiple features are compatible with each other; then cross-modal association and integration operations are carried out according to preset rules to eliminate the limitations of single modality information; after the operation is completed, a fusion feature with complete information and unified dimensions can be generated, thus successfully completing the entire processing flow of this step.

[0045] By implementing step 104, cross-modal fusion of clinical text features, diffusion features, and imaging features can establish the correlation between information from different sources, enabling the synergistic utilization of clinical information, imaging information, and perfusion information. Since various features reflect pathological states from different dimensions, the fused features obtained through cross-modal fusion can more comprehensively characterize the object to be detected, thereby reducing the impact of insufficient information from a single modality on the recognition results and improving the integrity and stability of feature expression.

[0046] Step 105: Based on the fusion features, determine the classification probability for the type of vascular occlusion.

[0047] The types of vascular occlusion can include large vessel occlusion, medium vessel occlusion, and no occlusion. Large vessel occlusion is occlusion involving a vessel diameter greater than a first preset threshold, while medium vessel occlusion is occlusion involving a vessel diameter less than or equal to the first preset threshold and greater than a second preset threshold.

[0048] In some examples, the classification probability for different types of vascular blockage is a quantitative result obtained by the detection model based on fusion feature calculations. It is used to characterize the likelihood of the object being detected belonging to each type of vascular blockage. The value ranges from zero to one, which can intuitively reflect the confidence level of the classification result. This probability is calculated by inputting the fusion features into the classification network and then performing feature mapping and probability conversion. For example, it can output the proportion of the object being detected corresponding to each of the three types of vascular states, thereby distinguishing the credibility of different judgment results.

[0049] The vascular occlusion type is a standardized diagnostic category system defined based on clinical diagnosis and treatment needs. It is divided into three main categories: large vessel occlusion, moderate vessel occlusion, and no occlusion. It is used to distinguish different degrees of intracranial vascular lesions and vascular status in patients with acute ischemic stroke. This category system is pre-defined based on the anatomical characteristics of intracranial vessels, the characteristics of occlusion lesions, and clinical diagnostic standards. For example, it can be a classification standard specifically used in stroke emergency scenarios to distinguish between severe vascular occlusion, moderate vascular occlusion, and completely patent vessels.

[0050] The first preset threshold is a pre-calibrated critical value for the inner diameter of blood vessels. Its main function is to delineate the boundary between large vessel occlusion and medium vessel occlusion, providing a quantitative basis for occlusion level determination. This threshold is comprehensively set by summarizing multi-center clinical anatomical data, stroke case vascular measurement data, and industry diagnosis and treatment standards, and is configured internally during the system deployment phase. For example, it can be a fixed inner diameter value that distinguishes between intracranial main vessels and secondary branch vessels, serving as the dividing line between the two occlusion types.

[0051] The second preset threshold is another pre-set critical value for the inner diameter of blood vessels. It is mainly used to delineate the boundary between moderate vascular occlusion and non-occlusion, and to distinguish between pathological branch vessel occlusion and normal blood vessel morphology. This threshold is calculated and determined in combination with the normal inner diameter range of small intracranial branch vessels and the criteria for judging vascular lesions, and is permanently fixed inside the detection system. For example, it can be the inner diameter boundary value for distinguishing between moderate vascular occlusion, physiologically thin blood vessels, and completely patent blood vessels.

[0052] Large vessel occlusion is a vascular occlusion lesion in which the diameter of the affected vessel is greater than a first preset threshold. It corresponds to a relatively typical large vessel occlusion symptom in clinical diagnosis and treatment. This category is determined by comparing the measured value of the vessel diameter with the first preset threshold. For example, it can be a lesion in which the lumen of the main intracranial vessels such as the internal carotid artery and the M1 segment of the middle cerebral artery is blocked.

[0053] Medium-sized vascular occlusion is defined as vascular occlusion in which the diameter of the affected vessel is less than or equal to a first preset threshold, while being greater than a second preset threshold. This corresponds to medium-sized vascular occlusion in clinical practice, which often presents with atypical imaging features and is easily missed. This category is determined by matching the vessel diameter to a numerical range formed by the two thresholds. For example, it could be a lesion involving the occlusion of intracranial secondary branch vessels such as the M2 to M4 segment of the middle cerebral artery or the A1 to A3 segment of the anterior cerebral artery.

[0054] "No occlusion" means that the target intracranial vessel of the subject does not meet the criteria for either type of occlusion, the vessel lumen remains in a normal and unobstructed state, and there are no pathological occlusion lesions that meet the definition requirements. When the inner diameter of the target vessel is less than or equal to the second preset threshold and no occlusion-related abnormal features are detected, it can be immediately identified as this category. For example, it can be a normal vascular state in which the patient's intracranial artery is continuous, the lumen is not significantly narrowed or interrupted, and there is no pathological occlusion.

[0055] For example, the fusion features obtained in the previous step can be imported into a preset classification operation module. After the module performs normalization processing on the fusion features, it generates initial classification score data through a fully connected network. The initial score is then subjected to normalized exponential operation, and the classification probabilities corresponding to the three types of blood vessel states are output sequentially. By combining the preset thresholds for the two levels of blood vessel diameter built into the system, the staff can complete the final differentiation and determination of the type of blood vessel blockage.

[0056] By implementing step 105, the classification probability of vascular occlusion type is determined based on fusion features, and large vessel occlusion, medium vessel occlusion, and no occlusion are classified, which enables automated identification of different vascular occlusion states. Since the fusion features integrate imaging features, perfusion features, and clinical features, it can improve the identification ability between different types of occlusion. At the same time, by outputting the classification probability corresponding to each category, not only can the classification results be given, but the matching degree of different categories can also be reflected, providing richer reference information for subsequent clinical diagnosis and decision-making, and improving the practical value of the detection results.

[0057] In summary, this application first performs multimodal preprocessing on non-enhanced CT images and arterial phase CTA images, and constructs multi-channel image tensors. This allows anatomical structural information and vascular imaging information from both types of image data to be co-expressed in a unified space. Compared to analysis using only single image data, multi-channel image tensors can simultaneously preserve brain parenchymal change features and vascular state features, thus providing more comprehensive image evidence for subsequent vascular occlusion type determination and improving the ability to distinguish between different vascular occlusion types. Based on the multi-channel image tensors, image generation processing is performed to obtain CT-like perfusion parameter maps, including cerebral blood flow maps, cerebral blood volume maps, and mean transit time maps. Since these CT-like perfusion parameter maps can reflect the perfusion state of brain tissue from another perspective, hemodynamic features are further supplemented on the basis of the original structural and vascular image information. By extracting diffusion features from the CT-like perfusion parameter maps, the model can simultaneously consider multi-dimensional information such as structure, blood vessels, and perfusion when determining the type of vascular occlusion, thereby enhancing the ability to express the differences between different occlusion types. This study utilizes not only image data but also encodes structured clinical text data to obtain clinical text features. These clinical text features are then fused cross-modally with image and diffusion features. Since the clinical text data includes information such as patient symptoms and scoring indicators, it can supplement the image analysis results from a clinical perspective. Through the cross-modal fusion mechanism, correlation modeling between information from different sources can be achieved, enabling the model to comprehensively utilize both image and clinical evidence when determining the type of vascular occlusion, thereby improving the reliability and stability of the classification results. Based on the fused features, the classification probability of the vascular occlusion type corresponding to the subject is determined, and the detection results are subdivided into three types: large vessel occlusion, medium vessel occlusion, and no occlusion. The fused features formed by fusing image, diffusion, and clinical text features can more comprehensively represent the state, thus improving the ability to identify different types of occlusion. At the same time, outputting classification probabilities rather than single classification results can also reflect the degree of matching between different types, providing richer reference information for subsequent diagnostic decisions and improving the practical value of the detection results. In summary, the vascular occlusion type detection method provided in this application integrates non-enhanced CT, arterial phase CTA, diffusion-generated CT-like perfusion parameter maps, and clinical text information to construct a multimodal cross-modal analysis mechanism. This enables comprehensive characterization and accurate identification of vascular occlusion types, thereby improving the accuracy, reliability, and clinical decision-making reference value of the classification results.

[0058] In some embodiments, step 101 may include: resampling a non-enhanced CT image and an arterial phase CTA image based on a preset isotropic voxel size to obtain a first resampled image and a second resampled image; rigidly registering the second resampled image with the first resampled image as a fixed reference image to obtain a registered CTA image; performing skull stripping on the first resampled image and the registered CTA image to obtain a first brain parenchyma image and a second brain parenchyma image; performing preset window width and window level truncation and linear normalization on the first brain parenchyma image and the second brain parenchyma image to obtain a first normalized image and a second normalized image; and superimposing the first normalized image and the second normalized image in the channel dimension to obtain a multi-channel image tensor.

[0059] In some examples, the preset isotropic voxel size is a pre-defined spatial standard size in which the length, width, and height of a single voxel in a 3D medical image are completely equal. Its function is to unify the spatial resolution of images under different scanning devices and different scanning parameters, and to ensure that the spatial scale of multiple sets of images is consistent. This size is combined with the MNI152 standard human brain template parameters and clinical cranial image processing specifications, and is configured and solidified during the system deployment phase. For example, it can be a standard voxel specification adapted to cranial image analysis, which can give all input images a unified spatial reference.

[0060] The first resampled image is obtained by interpolating a non-enhanced CT image according to a preset isotropic voxel size. This image is generated by the image processing module calling the resampling algorithm to perform spatial interpolation on the original non-enhanced CT image. For example, it can be a plain CT image of the brain with uniform spatial resolution and spatial coordinates matching the MNI152 standard human brain template. The second resampled image is obtained by resampling an arterial phase CTA image according to the same set of preset isotropic voxel sizes. This image is obtained by processing the original arterial phase CTA image using the same interpolation algorithm and spatial parameters as the first resampled image. For example, it can be a cerebral angiography image with the same voxel specifications and spatial scale as the first resampled image. The fixed reference image is the target image used as the spatial reference in the multimodal image registration process. All other images are corrected in position using this image as the standard. The first resampled image can be designated as the fixed reference image. This setting is predefined in the algorithm's running logic and does not require manual secondary selection. For example, it can be a resampled non-enhanced CT image, used as a reference template for anatomical alignment of the two images.

[0061] Rigid registration is an image processing method that aligns two images by relying solely on basic spatial transformations such as translation and rotation. The entire process does not change the internal structure and morphology of the images. This processing can be implemented using open-source tool libraries such as ANTs and SimpleITK. The computation process uses mutual information as the loss function and combines it with gradient descent algorithm to complete parameter optimization. For example, it can complete the registration operation of aligning anatomical points of two cranial images without distorting the morphology of blood vessels and brain tissue.

[0062] A registered CTA image is an arterial phase CTA image that is rigidly registered with a fixed reference image after the second resampled image has undergone rigid registration processing. This image is automatically output after the first resampled image is used as a reference and a complete rigid registration operation is performed. For example, it can be a cranial angiography image in which the spatial coordinates, anatomical position and non-enhanced CT image correspond one-to-one.

[0063] Skull stripping is a specialized process for removing non-brain tissue regions such as the skull, scalp, and eyeballs from cranial images. The aim is to eliminate the interference of irrelevant tissues on subsequent feature extraction. This process uses a three-dimensional U-shaped network trained on the Hammersmith dataset to generate a brain mask, and then relies on the brain mask to filter out invalid regions to achieve stripping. For example, it can be an image operation that removes the bony and soft tissues around the skull and only retains the effective brain regions inside the skull.

[0064] The first brain parenchyma image is the image of the intracranial brain parenchyma region retained after the first resampled image has undergone skull dissection processing. This image is generated by filtering regions of the first resampled image using a brain mask to remove invalid pixels such as the skull and scalp; for example, it can be a pure brain parenchyma plain scan image that completely preserves the anatomical structure and grayscale information of the brain by removing external interfering tissues. The second brain parenchyma image is the pure brain parenchyma image obtained after registering the CTA image and performing skull dissection processing. This image reuses the same set of brain masks and dissection logic to remove non-brain tissues from the registered arterial phase CTA image; for example, it can be an angiographic image containing only the brain parenchyma region with an anatomical range completely consistent with the first brain parenchyma image.

[0065] Preset window width and level truncation is a preprocessing operation that removes pixels from the image that exceed the effective grayscale range by pre-setting window width and level parameters according to medical imaging diagnostic standards. Dedicated truncation parameters are configured and fixed in advance for two types of images, with non-contrast CT images and arterial phase CTA images corresponding to different grayscale ranges. For example, the grayscale of non-contrast CT images can be limited to -10 HJ units to 80 HJ units, while the grayscale of arterial phase CTA images can be limited to 100 HJ units to 600 HJ units. Linear normalization is an operation that maps the grayscale values ​​of the image pixels after window width and level truncation to a uniform numerical range using a fixed linear formula. This facilitates subsequent feature calculations by the model. This operation matches the corresponding conversion formula based on the grayscale truncation range of different images and is automatically executed by the image processing module.

[0066] The first normalized image is the standardized image obtained by sequentially truncating and linearly normalizing the first brain parenchyma image with preset window widths and levels. This image is generated by first truncating the grayscale range and then applying the corresponding linear conversion formula to complete the numerical mapping. For example, it can be a standardized non-contrast computed tomography (CT) brain parenchyma image with a regular grayscale range and invalid grayscale information removed. The second normalized image is the standardized image obtained by truncating and linearly normalizing the second brain parenchyma image with preset window widths and levels according to corresponding parameters. This image undergoes grayscale processing and numerical mapping with reference to the specific parameters of arterial phase CTA images. For example, it can be a standardized arterial phase CT angiography brain parenchyma image with a uniform pixel value range and standardized grayscale features.

[0067] The process of superimposing the first and second normalized images along the channel dimension to obtain a multi-channel image tensor involves stitching and integrating two independent single-channel normalized images along the data channel dimension to form a three-dimensional tensor data that fuses dual-modal image information. This operation relies on the tensor stitching function built into the deep learning framework to complete the data combination according to the channel dimension; for example, it can be a dual-channel three-dimensional cranial image tensor that fuses non-enhanced computed tomography (CT) images and arterial phase CT angiography images.

[0068] For example, the program can first call the built-in preset isotropic voxel size parameters to resample the original non-enhanced CT image and the arterial phase CTA image respectively, obtaining two sets of images with uniform spatial dimensions. The first resampled image is set as a fixed reference image, and rigid registration is performed on the second resampled image to achieve precise alignment of the anatomical positions of the two images. Then, a brain extraction model can be run to perform skull dissection processing on the two images, obtaining a first brain parenchyma image and a second brain parenchyma image containing only the effective region. According to preset parameters, window width and window level truncation and linear normalization are performed on the two brain parenchyma images sequentially, outputting two standardized single-channel images. Finally, the program superimposes and combines the first and second normalized images along the channel dimension to generate the multi-channel image tensor required by the process.

[0069] Through the implementation of the above embodiments, preprocessing operations such as resampling, rigid registration, skull dissection, intensity normalization, and channel overlay are performed sequentially to ensure that the non-enhanced CT images and CTA images are consistent in spatial location, resolution, and grayscale distribution. Among them, resampling can eliminate spatial differences caused by different equipment and scanning parameters, rigid registration can achieve accurate correspondence between brain parenchyma and vascular structures, skull dissection can reduce the interference of irrelevant tissues such as skull on model analysis, and normalization processing can reduce the impact of grayscale fluctuations caused by different patients and different scanning conditions. By constructing spatially aligned and intensity-consistent multi-channel image tensors, the accuracy of subsequent feature extraction can be improved, enabling the model to more accurately identify the vascular occlusion area and its surrounding tissue changes, thereby improving the ability to distinguish between different types of vascular occlusion.

[0070] In some embodiments, step 102 may include: extracting continuous and discrete fields from structured clinical text data; normalizing the continuous fields to obtain a first processing result; performing one-hot encoding on the discrete fields to obtain a second processing result; concatenating the first and second processing results and mapping them through a preset multilayer perceptron to obtain a first text feature; performing natural language conversion on the structured clinical text data to obtain clinical description text; linearly mapping the semantic embedding vector extracted from the clinical description text through a preset adapter network to obtain a second text feature; and performing weighted fusion of the first and second text features to obtain clinical text features.

[0071] In some examples, continuous fields are quantitative fields in structured clinical text data whose values ​​exhibit continuous variation. The field values ​​can fluctuate continuously within a reasonable range and are primarily used to record objective quantitative diagnostic and treatment indicators for patients. These fields can be automatically filtered and extracted from structured clinical text data pushed by the electronic medical record system according to data type rules. Examples include continuous data items such as the time from onset to imaging and the National Institutes of Health Stroke Scale score. Discrete fields are categorical fields in structured clinical text data whose values ​​are fixed classification options and do not exhibit continuous variation. They are used to record qualitative information such as patient status, medical history, and symptoms. These fields are filtered simultaneously with continuous fields, relying on the system's built-in data classification rules to extract data from the original text data. Examples include categorical data items such as patient consciousness status, history of atrial fibrillation or diabetes, and type of limb symptoms.

[0072] The first processing result is standardized data generated by performing normalization on the extracted continuous fields. This standardization eliminates differences in numerical magnitude between different quantitative indicators. This result is obtained by uniformly processing the continuous fields using the system's built-in Z-score normalization algorithm. For example, it could be standardized data after converting varying durations of illness or stroke scale scores to a unified numerical range. The second processing result is coded data generated by performing one-hot encoding on the extracted discrete fields. This coded data converts textual classification information into a model-recognizable digital form. This result is obtained by encoding each category of discrete fields according to the system's preset encoding rules. For example, it could be binary coded data converted from states of consciousness such as wakefulness and drowsiness, or past medical histories such as atrial fibrillation and hypertension.

[0073] The pre-defined multilayer perceptron is a fully connected neural network that has been pre-trained and whose structure and parameters are fixed. Mapping refers to the operation process of using this network to perform dimensional transformation, feature extraction and information condensation on combined data. The pre-defined multilayer perceptron mapping operation is automatically executed by directly calling the pre-deployed multilayer perceptron model during the process. For example, it can be a network operation behavior of performing feature dimensionality reduction and information integration on concatenated multi-field data.

[0074] The first text feature is a structured text feature vector formed by fusing continuous and discrete field information. It carries quantitative and classification indicators from the original clinical data. This feature is generated by field concatenation and layer-by-layer operation and transformation by a multilayer perceptron. For example, normalized data and one-hot encoded data can be merged first, and then fed into a multilayer perceptron for feature transformation, finally outputting the complete operation chain of structured path text features.

[0075] Natural language processing (NLP) is the process of reconstructing structured clinical fields in key-value pair format into coherent and natural sentences according to a preset text template. This transforms formatted data into descriptive text. The conversion relies on built-in sentence templates to fill in and combine fragmented structured fields; for example, it can integrate individual fields such as score, duration, symptoms, and medical history into a complete sentence describing the patient's condition. Clinical descriptive text is the pure sentence text generated after NLP of structured clinical text data. It fully restores comprehensive information such as the patient's stroke condition, examination time, and basic medical history. This text is directly generated through field matching and sentence concatenation; for example, it can be a natural descriptive sentence that completely records the patient's stroke score, duration of illness, clinical symptoms, and past medical history.

[0076] The pre-trained adapter network is a lightweight linear network module with pre-trained and fixed parameters. It is used to adapt vector dimensions and fine-tune features, ensuring consistency in the dimensions of features from different sources. This network is trained and configured during the system deployment phase and directly called upon during the process. For example, it can be a linear transformation network that corrects the dimensions of the language model's output vector to match the overall feature system. The semantic embedding vector is the vector data encoded and output by the pre-trained clinical language model after clinical descriptive text is input. It can deeply represent the semantic information such as the patient's condition and symptoms contained in the text. This vector is obtained by encoding and extracting the clinical descriptive text through a dedicated pre-trained clinical language model. For example, it can be a high-dimensional digital vector carrying semantic information such as the patient's overall disease characteristics and clinical manifestations. The pre-trained clinical language model can be the Clinical Bidirectional Encoder Representations from Transformers (ClinicalBERT). This model is pre-trained based on the MIMIC-III large-scale clinical notes dataset and can accurately capture clinical professional expressions, symptom features and disease correlation information. In the text feature extraction process of this embodiment, the main parameters of the model are kept frozen and are only used to output the deep semantic representation vector corresponding to the clinical description text.

[0077] The second text feature is the feature vector obtained by linearly mapping the semantic embedding vector through a preset adapter network. It carries the deep semantic information of the clinical description text. This feature is generated by the semantic embedding vector input to the adapter network to complete dimensional transformation and feature optimization. For example, it can be a special text feature vector that has been dimensionally normalized and extracted to extract the semantic information of the text.

[0078] The process of weighted fusion of the first and second text features to obtain clinical text features is an integration operation of dual-path text features. By assigning fixed weights to the two sets of features and performing a summation operation, structured data information and natural language semantic information are fused. This fusion operation is automatically calculated according to the system's preset weight coefficients. For example, it can be an integrated comprehensive text feature vector that simultaneously includes objective quantitative indicators, categorical medical history information, and clinical semantic information.

[0079] For example, the process can begin by reading the original structured clinical text data, automatically distinguishing and extracting continuous and discrete fields; performing normalization and one-hot encoding on the two types of fields respectively, generating a first processing result and a second processing result. The two sets of processed data are then concatenated and fed into a pre-defined multilayer perceptron to complete feature mapping, thus obtaining the first text feature; simultaneously, natural language processing is performed on the original structured fields to obtain the clinical description text, extracting the corresponding semantic embedding vectors and generating the second text feature through linear mapping via an adapter network; finally, a weighted summation of the two sets of text features can be performed according to preset weights, fusing the dual-path information to obtain the final clinical text feature.

[0080] Through the implementation of the above embodiments, a dual-path text feature construction approach combining structured coding and semantic coding is adopted. On the one hand, by standardizing the coding of continuous and discrete fields, structured clinical indicators such as stroke scale scores, onset time, and underlying diseases can be preserved. On the other hand, through natural language conversion and semantic embedding extraction, the implicit semantic relationships in clinical information such as symptom descriptions can be further mined. Subsequently, weighted fusion is used to form clinical text features, enabling the model to utilize not only explicit quantitative indicators but also richer clinical semantic information. For cases of different degrees of vascular occlusion with similar imaging manifestations but different clinical symptoms, this scheme can provide additional discriminative criteria, thereby improving the ability to identify different occlusion types.

[0081] In some embodiments, step 103 may include: inputting a multi-channel image tensor as a conditional input to the inverse denoising network of a preset conditional diffusion model; and generating a CT-like perfusion parameter map by performing multi-step iterative denoising from initial random Gaussian noise guided by the conditional input through the inverse denoising network, wherein the preset conditional diffusion model uses historical CT perfusion parameter maps as diffusion targets during the training phase, and the inverse denoising network is used to learn the conditional inverse mapping from the noise distribution to the perfusion parameter map distribution.

[0082] In some examples, the pre-defined diffusion model can specifically be a denoising diffusion probabilistic model (DDPM), which is a dedicated generative deep learning model that is trained offline and has its parameters fixed. This model can convert noisy data into target medical images under the constraints of external image conditions. During the training phase, the model uses historical CT perfusion parameter maps accumulated in clinical practice as the diffusion learning target and continuously fits the mapping relationship between conventional cranial images and perfusion images. After training, the model is deployed inside the system and the parameters are not updated during operation. For example, it can be a cranial diffusion model specifically adapted to acute ischemic stroke scenarios, which can infer perfusion images based on non-enhanced CT images and arterial phase CTA images.

[0083] The inverse denoising network is a component sub-network of the pre-defined conditional diffusion model. It adopts a conditional U-shaped network architecture and is mainly used to perform inverse denoising operations. It is also the carrier for the model to learn distribution transformation rules. This network is jointly trained with the conditional diffusion model. During the training process, it continuously learns the conditional inverse mapping from noise distribution to perfusion parameter map distribution. After training, it is embedded into the system call chain as an independent functional module. For example, it can be a dedicated convolutional network module that combines external condition information to gradually remove random noise and restore the hemodynamic features of brain tissue.

[0084] The process of inputting multi-channel image tensors as conditional inputs into the inverse denoising network of a pre-defined conditional diffusion model involves setting the previously processed multimodal image tensors as global constraints to provide anatomical and vascular morphological basis for subsequent image generation. This operation is completed automatically by the system. The program reads the generated multi-channel image tensors, adapts them according to the data format specified by the inverse denoising network, and then transmits them to the network's conditional input port. For example, a dual-channel three-dimensional tensor that fuses two types of cranial image information can be used as guiding data to be completely fed into the inverse denoising network.

[0085] The system employs an inverse denoising network, starting with initial random Gaussian noise and guided by conditional input, to perform multi-step iterative denoising to generate a CT-like perfusion parameter map. It uses pure random Gaussian noise as the starting point and, relying on the inverse denoising network, completes multiple rounds of cyclic denoising under the continuous guidance of multi-channel image tensors, ultimately outputting an image with realistic blood flow perfusion characteristics. Based on the model training logic, the preset conditional diffusion model consistently uses historical computed tomography (CT) perfusion parameter maps as the diffusion target during the training phase. This allows the inverse denoising network to stably grasp the conditional inverse mapping relationship between noise distribution and perfusion parameter map distribution. This entire computational process is automatically executed by the system according to a preset number of iterations, and the target image is directly output after the iterations are completed. For example, it could start with initial Gaussian noise, undergo 1000 iterations of denoising, and finally output a complete CT-like perfusion parameter map including cerebral blood flow map, cerebral blood volume map, and mean transit time map.

[0086] For example, a locally deployed pre-defined conditional diffusion model can be loaded first, and the inverse denoising network within the model can be enabled. The multi-channel image tensor produced in the previous step can be retrieved, and after data format adaptation, it can be used as conditional information input into the inverse denoising network. The program automatically generates initial random Gaussian noise samples and starts iterative denoising operations according to the model's preset total number of steps. Each iteration relies on the multi-channel image tensor for feature guidance, gradually removing noise and restoring perfusion image details. After all iterative steps are completed, the inverse denoising network can output a complete CT-like perfusion parameter map, successfully completing the image generation work of this step.

[0087] Through the implementation of the above embodiments, the conditional diffusion model generates CT-like perfusion parameter maps based on non-enhanced CT images and CTA images, enabling the model to obtain hemodynamic information reflecting the perfusion status of brain tissue without the need for additional CT perfusion examination. Since parameters such as cerebral blood flow, cerebral blood volume, and mean transit time can reflect the blood supply status and functional impairment of ischemic areas, they can provide important supplementary information beyond structural and vascular imaging for determining the type of vascular occlusion. In particular, in some cases of moderate vascular occlusion, the direct imaging signs may not be obvious due to the small occlusion area, while perfusion abnormalities have often already appeared. By introducing perfusion features generated by diffusion, the model's sensitivity to different degrees of occlusion can be enhanced, and the accuracy of vascular occlusion type detection can be improved.

[0088] In some embodiments, prior to step 104, the aforementioned method for detecting vascular occlusion type may further include: extracting three-dimensional features based on multi-channel image tensors using a first preset convolutional network to obtain image features; extracting three-dimensional features from CT-like perfusion parameter maps using a second preset convolutional network to obtain diffusion features; step 104 may include: mapping image features, diffusion features, and clinical text features to preset unified feature dimensions to obtain first mapping features, second mapping features, and third mapping features; using any one of the first, second, and third mapping features as a query, and the other two mapping features as keys and values, performing cross-modal attention calculations to obtain attention-weighted features corresponding to each modality; and concatenating the attention-weighted features corresponding to each modality along the channel dimension to obtain fusion features.

[0089] In some examples, the first preset convolutional network is a three-dimensional convolutional neural network that is trained offline and whose network structure and weight parameters are all fixed. It is specifically designed to parse the spatial structure information in multi-channel image tensors and adapt to the feature extraction requirements of cranial anatomical images and vascular images. After training, the network is directly deployed to the system operating environment without the need for online parameter tuning. For example, it can be a dedicated feature extraction network that uses a three-dimensional residual network architecture and is optimized for the combined tensor of non-enhanced CT images and arterial phase CTA images.

[0090] The process of extracting three-dimensional features from multi-channel image tensors through a first preset convolutional network to obtain image features involves inputting the previously generated multi-channel image tensors completely into the first preset convolutional network. The network then captures three-dimensional spatial features such as the morphology of cranial tissues, the course of blood vessels, and the morphology of lumens through three-dimensional convolutional operations, and finally outputs standardized image features. This process is automatically executed by the system according to preset calling logic without manual intervention. For example, it can extract high-dimensional image feature vectors that can characterize occlusion morphologies such as vascular stenosis and vascular interruption from the fused image tensors.

[0091] The second pre-set convolutional network is another three-dimensional convolutional neural network with pre-trained and fixed parameters. It is mainly designed for CT-like perfusion parameter maps and focuses on mining hemodynamic features. This network independently completes model training and deployment, and works with the first pre-set convolutional network to extract features from different image modalities. For example, it can be a three-dimensional convolutional network adapted to cerebral blood flow maps, cerebral blood volume maps, and average transit time maps, which can accurately identify feature information corresponding to local perfusion abnormalities in brain tissue.

[0092] The process of extracting three-dimensional features from the CT-like perfusion parameter map through a second preset convolutional network to obtain diffusion features involves inputting the CT-like perfusion parameter map into the second preset convolutional network, using three-dimensional convolution to capture information such as blood flow distribution, perfusion delay, and abnormal blood flow, and then outputting diffusion features after calculation. This process is automatically triggered and executed along with the overall detection link; for example, it can extract feature vectors that can reflect the pathological patterns such as local cerebral ischemia and abnormal blood flow duration.

[0093] The pre-defined uniform feature dimension is a fixed feature vector length standard set in advance. Since image features, diffusion features, and clinical text features come from different modalities, their original dimensions differ. This standard is used to unify the vector size of the three types of features to ensure that subsequent cross-modal attention calculations can be carried out normally. This dimension parameter is fixedly written in the system configuration file and remains unchanged during operation. For example, it can be to set a uniform feature vector length so that features from three different sources have the same data structure and dimension specifications.

[0094] The first, second, and third mapping features are standardized feature data obtained after the three types of original features have undergone dimensional adaptation. The first mapping feature is obtained by dimensional transformation of image features, the second mapping feature is obtained by dimensional transformation of diffusion features, and the third mapping feature is obtained by dimensional transformation of clinical text features according to the same standard. All three types of mapping features are generated by the system's built-in dimensional transformation module and correspond to the standardized features of the three major modalities. For example, they can be three sets of standardized feature vectors with completely consistent dimensions, respectively carrying vascular morphology information, blood flow perfusion information, and clinical diagnosis and treatment information.

[0095] The process of performing cross-modal attention computation using any one of the first, second, and third mapping features as a query, and the remaining two mapping features as keys and values, to obtain attention-weighted features corresponding to each modality, is the computational process for achieving multimodal information association and fusion. The query is used to locate the currently focused feature subject, and the keys and values ​​are used to provide association reference information. Cross-modal attention computation can uncover the inherent associations between different modalities and weight and strengthen the original features according to the strength of the association. This operation relies on the system's built-in attention computation module to automatically complete multiple rounds of computation. For example, it can take turns using the three types of mapping features as query items, associate them with the other two types of features to carry out the operation, amplify the cross-modal association features related to vascular occlusion, and generate single-modal features with optimized weights.

[0096] The process of concatenating the attention-weighted features corresponding to each modality along the channel dimension to obtain the fused features is to combine multiple independent attention-weighted features sequentially according to the data channel arrangement rules to form an integrated feature that gathers information from all modalities. This concatenation operation is implemented by the tensor concatenation function built into the deep learning framework; for example, it can integrate three types of weighted features: vascular morphology, blood flow perfusion, and clinical condition, to form a comprehensive feature vector with complete dimensions and comprehensive information.

[0097] For example, the first and second preset convolutional networks can be invoked in parallel to perform three-dimensional feature extraction on the multi-channel image tensor and CT-like perfusion parameter map, respectively, to obtain image features and diffusion features. The image features, diffusion features, and previously obtained clinical text features are then uniformly converted to a preset unified feature dimension to generate three sets of corresponding mapping features. Subsequently, each set of mapping features can be selected as a query in turn, and cross-modal attention calculation can be completed in combination with the other two sets of mapping features to obtain attention-weighted features for each modality. After all attention operations are completed, channel dimension splicing can be performed on all attention-weighted features. The final spliced ​​integrated data is the fusion feature, successfully completing all processing work in this step.

[0098] Through the implementation of the above embodiments, image features, diffusion features, and clinical text features are extracted respectively, and multimodal feature fusion is achieved through a cross-modal attention mechanism. Compared with the simple stitching method, the cross-modal attention mechanism can dynamically focus on key information in other modalities based on the current modal features, enabling image features to establish a deeper association with perfusion features and clinical features. For example, when a minor vascular abnormality appears in the image, the model can further focus on the corresponding perfusion abnormality area and related clinical symptoms, thereby amplifying the differences between different types of vascular occlusion. By constructing a cross-modal information interaction mechanism, the discriminative ability of fused features can be improved, further enhancing the accuracy of vascular occlusion type detection.

[0099] In some embodiments, step 105 may include: performing a global average pooling operation on the fused features to obtain a pooled feature vector; inputting the pooled feature vector into a preset fully connected classification layer to obtain a classification score vector; and performing normalized exponential processing on the classification score vector to obtain the classification probability that the object to be detected belongs to large vessel occlusion, medium vessel occlusion, or no occlusion.

[0100] In some examples, global average pooling is a holistic feature compression operation performed on high-dimensional fused features. By averaging across all spatial dimensions of the fused features, it reduces feature size, filters redundant information, and fully retains features relevant to vascular occlusion determination. This operation is automatically executed by the system calling the pooling function module built into the deep learning framework. For example, it can be used to calculate the global mean of high-dimensional fused features formed by multimodal concatenation, achieving overall simplification of the feature space dimensions. The pooled feature vector is the one-dimensional feature data generated after the fused features undergo global average pooling. This vector eliminates invalid spatial information and concentrates the effective features of three modalities: image morphology, blood perfusion, and clinical text. It serves as the foundation for subsequent classification calculations and is directly output after the global average pooling operation. For example, it can be a one-dimensional feature vector with regular dimensions and highly condensed feature information, adapted to the input requirements of classification networks.

[0101] The preset fully connected classification layer is a fully connected neural network layer that has been trained offline and whose network structure and weight parameters are fixed. It is specifically designed to receive pooled feature vectors and complete feature mapping, outputting raw judgment scores for different classification dimensions. This network layer is trained and deployed together with the overall detection model, and its parameters are not adjusted during actual operation; it is directly called and executed by the system. For example, it can be a classification network layer specifically designed for the three judgment scenarios of large vessel occlusion, medium vessel occlusion, and no occlusion. The classification score vector is the raw score sequence obtained by the forward operation of the network after the pooled feature vector is input into the preset fully connected classification layer. Each score in the sequence corresponds to a type of vessel occlusion. The score value can initially reflect the degree of matching between the feature and the corresponding category. At this time, the score has not yet been converted into standard probability form. This vector is generated by the operation of the fully connected classification layer after the pooled feature vector is input into it. For example, it can be a numerical sequence composed of three sets of raw scores, which correspond to the three vessel states of large vessel occlusion, medium vessel occlusion, and no occlusion, respectively.

[0102] Normalization index processing is a numerical transformation operation based on the normalization index function. It can transform the original classification scores with variable value ranges into probability values ​​with values ​​between 0 and 1, and keep the sum of the probabilities of all categories at 1. This processing is automatically completed by the system's built-in dedicated calculation module. For example, it can convert three sets of original judgment scores with large differences into standardized values ​​that conform to the probability distribution rules.

[0103] The classification probability of the object to be detected belonging to large vessel occlusion, moderate vessel occlusion, and no occlusion is the final quantitative result obtained after the classification score vector is processed by the normalization index. It respectively represents the probability of the intracranial vascular status belonging to large vessel occlusion, moderate vessel occlusion, and no occlusion. After the normalization index processing is completed, this set of probability data can be split by category and output sequentially.

[0104] For example, a global average pooling operation can be performed on the fusion features obtained in the previous step to generate a simplified pooled feature vector. The pooled feature vector is then input into a preset fully connected classification layer, and the corresponding classification score vector is calculated through network mapping. Subsequently, a normalization exponential function can be called to perform a unified transformation on the original scores to complete the probability normalization process. After the transformation is completed, the system outputs the classification probabilities corresponding to the three types of vascular occlusion according to the preset classification rules, thus completing the entire processing flow of this step.

[0105] Through the implementation of the above embodiments, global average pooling and classification calculation are performed on the fused features, and the classification probabilities corresponding to large vessel occlusion, moderate vessel occlusion, and no occlusion are output, enabling the model to complete the vessel occlusion type judgment based on global features. Global average pooling can effectively integrate key information from various spatial locations and reduce the impact of local noise on the classification results. The fully connected classification layer can learn the feature distribution patterns corresponding to different vessel occlusion types. Finally, the classification results in probabilistic form are output, which not only provide detection conclusions but also reflect the model's confidence level for each category, thereby improving the reliability and reference value of the identification results of different degrees of vessel occlusion.

[0106] In some embodiments, after step 105, the aforementioned vascular occlusion type detection method may further include: obtaining the target predicted type with the highest probability value among the classification probabilities; if the probability value corresponding to the target predicted type is greater than a preset confidence threshold, determining the target predicted type as the target classification result; weighting the feature map of the last convolutional layer in the first preset convolutional network used to extract image features based on the target predicted type to generate a spatial activation map; upsampling the spatial activation map to the original input image space to obtain an upsampled activation map; superimposing the upsampled activation map with the vascular enhancement image of the arterial phase CTA image to obtain a lesion localization heatmap; and encapsulating the target classification result, the classification probability of the object to be detected belonging to each vascular occlusion type, and the lesion localization heatmap into a structured report.

[0107] In some examples, the target prediction type is selected from the classification probabilities corresponding to the three types of vascular occlusion, and the category with the highest probability value is selected. This is a preliminary diagnostic result determined by the model based on the fusion features. This content is obtained by the system automatically traversing all classification probabilities and completing the numerical comparison. For example, the three classification probabilities may correspond to large vessel occlusion, medium vessel occlusion, and no occlusion, respectively. Among them, medium vessel occlusion has the highest probability value, so medium vessel occlusion is the target prediction type.

[0108] The pre-set confidence threshold is a fixed probability critical value that is pre-set during the deployment phase by combining a large amount of clinical test data with stroke diagnostic criteria. It is used to judge whether the preliminary prediction results of the model have effective clinical reference value. This threshold is fixed during the system configuration phase and will not be adjusted during equipment operation. For example, the system can set 0.7 as the confidence critical standard. Only when the prediction probability is higher than this value is the prediction result considered reliable.

[0109] The target classification result is the final diagnostic conclusion of the vascular occlusion type determined after the confidence rule verification. It is also the formal judgment content output externally. When the probability value corresponding to the target prediction type is greater than the preset confidence threshold, the target prediction type can be directly determined as the target classification result. For example, the target prediction type can be large vessel occlusion, and the corresponding probability reaches 0.85, which exceeds the preset confidence threshold. In this case, large vessel occlusion can be determined as the target classification result.

[0110] Spatial activation maps are feature distribution maps that visually represent the brain regions that the model focuses on when making category judgments. They are generated through weighted operations on convolutional network feature maps. For example, they can highlight brain regions where intracranial vascular lesions are located, reflecting the feature images of the core areas of focus when the model identifies vascular occlusion. The process of generating spatial activation maps by weighting the feature map of the last convolutional layer in the first preset convolutional network used to extract image features based on the target prediction type is as follows: the feature map of the last convolutional layer retained during the operation of the first preset convolutional network is retrieved, and different weights are assigned to each region of the feature map in combination with category gradient information, ultimately completing feature reorganization and generating spatial activation maps. The entire set of operations is automatically executed by the gradient calculation module called by the system; for example, for samples judged as moderate vascular occlusion, the convolutional feature map is weighted to highlight the brain region range corresponding to moderate vascular occlusion.

[0111] The original input image space is the complete spatial system corresponding to the non-contrast CT image and the arterial phase CTA image obtained from the patient's initial scan. It includes all spatial parameters such as the resolution, voxel size, and anatomical coordinates of the original image. This spatial standard uses the original parameters recorded during the image acquisition and preprocessing stages as a unified benchmark for subsequent image transformations. For example, it could be the image size, spatial coordinates, and resolution standard corresponding to the patient's first cranial scan image after admission. The upsampled activation map is the image obtained by transforming the size and coordinates of the spatial activation map using an interpolation algorithm and adapting it to the original input image space. This image is obtained by enlarging and aligning the spatial activation map using a trilinear interpolation algorithm. For example, it could be a smaller feature activation map scaled up to the same resolution and image size as the original cranial image.

[0112] Angiography-enhanced images are obtained by specifically processing arterial phase CTA images. They enhance the display of intracranial vessel contours and courses, and are generally generated using maximum intensity projection. This image is obtained by the system performing projection operations on registered and normalized arterial phase CTA images. For example, it can be a projection image clearly showing the morphology of major intracranial arteries and branches such as the internal carotid artery, middle cerebral artery, and anterior cerebral artery. Lesion localization heatmaps are composite images formed by overlaying and fusing upsampled activation images and angiography-enhanced images from arterial phase CTA images with set transparency. Different regions of interest are distinguished by color, visually marking the location of vascular occlusion lesions. This image is generated by the system performing layer overlay and pseudo-color mapping operations. For example, it can be a visual diagnostic image formed by marking vascular occlusion points with colored highlights and overlaying them onto a vascular projection image.

[0113] The process of packaging the target classification results, the classification probability of the object to be tested belonging to each type of vascular occlusion, and the lesion localization heat map into a structured report is an operation that integrates diagnostic conclusions, quantitative data, and visual images into a format. The content is packaged according to a standard template compatible with the clinical system to form a complete and viewable diagnostic document. This operation is automatically completed by the system calling the report generation module, filling in the text conclusion, probability values, and image files in sequence; for example, it can integrate the final occlusion type, the three classification probabilities, and the lesion heat map to generate a standardized diagnostic report that can be stored and viewed in the hospital's image archiving and communication system.

[0114] For example, the system can first compare the numerical values ​​of the probabilities of the three categories to select the corresponding target prediction type, and then compare its probability value with a preset confidence threshold to determine the official target classification result. Next, it can retrieve the feature map of the last layer of the first preset convolutional network, perform a weighted operation based on the target prediction type, and generate the corresponding spatial activation map. Then, it can perform upsampling processing on the spatial activation map using trilinear interpolation to match the original input image space, and then overlay it with the enhanced vascular image of the arterial phase CTA image to obtain a lesion localization heatmap. Finally, it can integrate the content according to a preset template, uniformly encapsulating the classification result, the probabilities of each category, and the lesion localization heatmap to output a complete structured report.

[0115] Through the implementation of the above embodiments, after obtaining the classification results, a lesion localization heatmap is further generated based on the target prediction type, and the classification results, classification probabilities, and heatmaps are encapsulated into a structured report. The probability threshold screening avoids the direct use of low-confidence results in clinical decision-making, improving the reliability of the detection results. The generated lesion localization heatmap visually displays the vascular occlusion areas and related brain tissue regions of interest to the model, enabling doctors to quickly verify the model's judgment basis. For cases prone to diagnostic discrepancies, such as large vessel occlusion and moderate vessel occlusion, the heatmap provides visual reference evidence, improving the interpretability and clinical acceptability of the model's output results, thereby further enhancing the practical application value of vascular occlusion type detection.

[0116] Furthermore, as an implementation of the aforementioned method embodiments, this application also provides a vascular blockage type detection device for implementing the aforementioned method embodiments. This device embodiment corresponds to the aforementioned method embodiments. For ease of reading, this vascular blockage type detection device embodiment will not repeat the details of the aforementioned method embodiments one by one, but it should be clear that the device in this application embodiment can correspondingly implement all the contents of the aforementioned method embodiments. Figure 2 As shown, the vascular occlusion type detection device 20 includes: a data acquisition unit 201, a text encoding unit 202, an image diffusion unit 203, a feature fusion unit 204, and a probability determination unit 205. The data acquisition unit 201 performs multimodal preprocessing on non-enhanced CT images and arterial phase CTA images to obtain multi-channel image tensors. The text encoding unit 202 encodes structured clinical text data to obtain clinical text features. The image diffusion unit 203 performs image generation processing based on the multi-channel image tensors to obtain a CT-like perfusion parameter map, wherein the CT-like perfusion parameter map may include cerebral blood flow... The system includes a flow rate map, a cerebral blood volume map, and a mean transit time map; a feature fusion unit 204, used to perform cross-modal fusion of the aforementioned clinical text features, diffusion features extracted based on CT-like perfusion parameter maps, and image features extracted based on multi-channel image tensors to obtain fused features; and a probability determination unit 205, used to determine the classification probability for the type of vascular occlusion based on the fused features, wherein the type of vascular occlusion may include large vessel occlusion, medium vessel occlusion, and no occlusion, large vessel occlusion is occlusion involving a vessel diameter greater than a first preset threshold, and medium vessel occlusion is occlusion involving a vessel diameter less than or equal to the first preset threshold and greater than a second preset threshold.

[0117] In some embodiments, the data acquisition unit 201 is further configured to resample the non-enhanced CT image and the arterial phase CTA image based on a preset isotropic voxel size to obtain a first resampled image and a second resampled image; rigidly register the second resampled image with the first resampled image as a fixed reference image to obtain a registered CTA image; perform skull stripping processing on the first resampled image and the registered CTA image to obtain a first brain parenchyma image and a second brain parenchyma image; perform preset window width and window level truncation and linear normalization on the first brain parenchyma image and the second brain parenchyma image to obtain a first normalized image and a second normalized image; and superimpose the first normalized image and the second normalized image in the channel dimension to obtain a multi-channel image tensor.

[0118] In some embodiments, the text encoding unit 202 is further configured to extract continuous fields and discrete fields from structured clinical text data; normalize the continuous fields to obtain a first processing result; perform one-hot encoding on the discrete fields to obtain a second processing result; concatenate the first processing result and the second processing result, and then map them through a preset multilayer perceptron to obtain a first text feature; perform natural language conversion on the structured clinical text data to obtain clinical description text; perform linear mapping on the semantic embedding vector extracted based on the clinical description text through a preset adapter network to obtain a second text feature; and perform weighted fusion of the first text feature and the second text feature to obtain clinical text features.

[0119] In some embodiments, the image diffusion unit 203 is further configured to input the multi-channel image tensor as a conditional input to the inverse denoising network of the preset conditional diffusion model; through the inverse denoising network, starting from the initial random Gaussian noise, multi-step iterative denoising is performed guided by the conditional input to generate a CT-like perfusion parameter map, wherein the preset conditional diffusion model uses historical CT perfusion parameter maps as diffusion targets during the training phase, and the inverse denoising network is used to learn the conditional inverse mapping from the noise distribution to the perfusion parameter map distribution.

[0120] In some embodiments, the feature fusion unit 204 is further configured to perform three-dimensional feature extraction based on multi-channel image tensors through a first preset convolutional network to obtain image features; perform three-dimensional feature extraction on CT-like perfusion parameter maps through a second preset convolutional network to obtain diffusion features; map the image features, diffusion features, and clinical text features to preset unified feature dimensions to obtain first mapping features, second mapping features, and third mapping features; use any one of the first, second, and third mapping features as a query, and the other two mapping features as keys and values, respectively, to perform cross-modal attention calculation to obtain attention-weighted features corresponding to each modality; and concatenate the attention-weighted features corresponding to each modality along the channel dimension to obtain fused features.

[0121] In some embodiments, the probability determination unit 205 is further configured to perform a global average pooling operation on the fused features to obtain a pooled feature vector; input the pooled feature vector into a preset fully connected classification layer to obtain a classification score vector; and perform normalized exponential processing on the classification score vector to obtain the classification probability that the object to be detected belongs to large vessel occlusion, medium vessel occlusion, or no occlusion.

[0122] In some embodiments, the vascular occlusion type detection device 20 further includes a report generation unit, used to obtain the target predicted type with the highest probability value in the classification probability; if the probability value corresponding to the target predicted type is greater than a preset confidence threshold, the target predicted type is determined as the target classification result; the feature map of the last convolutional layer in the first preset convolutional network used to extract image features is weighted based on the target predicted type to generate a spatial activation map; the spatial activation map is upsampled to the original input image space to obtain an upsampled activation map; the upsampled activation map is superimposed with the vascular enhancement image of the arterial phase CTA image to obtain a lesion localization heatmap; the target classification result, the classification probability of the object to be detected belonging to each vascular occlusion type, and the lesion localization heatmap are encapsulated into a structured report.

[0123] This application also provides a computer-readable storage medium storing computer-executable instructions or a computer program, which, when executed by a processor, will cause the processor to perform any step of the blood vessel occlusion type detection method provided in this application.

[0124] In some embodiments, the computer-readable storage medium may be a random access memory (RAM), a read-only memory (ROM), flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); or it may be a variety of devices that include one or any combination of the above-mentioned memories.

[0125] In some embodiments, computer-executable instructions may take the form of programs, software, software modules, scripts, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.

[0126] In some embodiments, computer-executable instructions may, but do not necessarily, correspond to files in a file system, and may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., files that store one or more modules, subroutines, or code sections).

[0127] In some embodiments, computer-executable instructions may be deployed to execute on an electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.

[0128] like Figure 3 As shown, this application also provides an electronic device 30, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor. When the processor 320 executes the computer program 311, it implements any step of the above-described method for detecting the type of vascular occlusion.

[0129] This application also provides a computer program product comprising a computer program or computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer program or computer-executable instructions from the computer-readable storage medium and executes the computer program or computer-executable instructions, causing the electronic device to perform any step of the blood vessel occlusion type detection method described above.

[0130] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting the type of vascular occlusion, characterized in that, include: Multimodal preprocessing was performed on non-contrast CT images and arterial phase CTA images to obtain multi-channel image tensors; Structured clinical text data is encoded to obtain clinical text features; Image generation processing is performed based on the multi-channel image tensor to obtain a CT-like perfusion parameter map, wherein the CT-like perfusion parameter map includes a cerebral blood flow map, a cerebral blood volume map, and a mean transit time map; Cross-modal fusion is performed on the clinical text features, the diffusion features extracted based on the CT-like perfusion parameter map, and the image features extracted based on the multi-channel image tensor to obtain fused features; Based on the fusion features, the classification probability for the type of vascular occlusion is determined, wherein the type of vascular occlusion includes large vessel occlusion, medium vessel occlusion and no occlusion, the large vessel occlusion is an occlusion involving a vessel diameter greater than a first preset threshold, and the medium vessel occlusion is an occlusion involving a vessel diameter less than or equal to the first preset threshold and greater than a second preset threshold.

2. The method for detecting the type of vascular occlusion according to claim 1, characterized in that, The multimodal preprocessing of non-contrast CT images and arterial phase CTA images to obtain multi-channel image tensors includes: Based on a preset isotropic voxel size, the non-enhanced CT image and the arterial phase CTA image are resampled respectively to obtain a first resampled image and a second resampled image. Using the first resampled image as a fixed reference image, rigid registration is performed on the second resampled image to obtain a registered CTA image; The first resampled image and the registered CTA image were subjected to cranial dissection processing to obtain a first brain parenchyma image and a second brain parenchyma image. The first brain parenchyma image and the second brain parenchyma image are respectively truncated with a preset window width and window level and linearly normalized to obtain a first normalized image and a second normalized image; The first normalized image and the second normalized image are superimposed along the channel dimension to obtain the multi-channel image tensor.

3. The method for detecting the type of vascular occlusion according to claim 1, characterized in that, The encoding of structured clinical text data to obtain clinical text features includes: Extract continuous and discrete fields from the structured clinical text data; The continuous field is normalized to obtain the first processing result; One-hot encoding is performed on the discrete field to obtain the second processing result; After concatenating the first processing result and the second processing result, the first text feature is obtained by mapping through a preset multilayer perceptron. Natural language conversion is performed on the structured clinical text data to obtain clinical description text; The second text feature is obtained by linearly mapping the semantic embedding vector extracted from the clinical description text through a preset adapter network; The first text feature and the second text feature are weighted and fused to obtain the clinical text feature.

4. The method according to claim 1, characterized in that, The image generation process based on the multi-channel image tensor to obtain a CT-like perfusion parameter map includes: The multi-channel image tensor is used as a conditional input and fed into the inverse denoising network of the preset conditional diffusion model. The inverse denoising network starts from initial random Gaussian noise and performs multi-step iterative denoising guided by the conditional input to generate the CT-like perfusion parameter map. The preset conditional diffusion model uses historical CT perfusion parameter maps as diffusion targets during the training phase, and the inverse denoising network is used to learn the conditional inverse mapping from noise distribution to perfusion parameter map distribution.

5. The method for detecting the type of vascular occlusion according to claim 1, characterized in that, Before performing cross-modal fusion of the clinical text features, the diffusion features extracted based on the CT-like perfusion parameter map, and the image features extracted based on the multi-channel image tensor to obtain the fused features, the vascular occlusion type detection method further includes: The image features are obtained by performing three-dimensional feature extraction based on the multi-channel image tensor through a first preset convolutional network; The diffusion features are obtained by extracting three-dimensional features from the CT-like perfusion parameter map using a second preset convolutional network. The cross-modal fusion of the clinical text features, the diffusion features extracted based on the CT-like perfusion parameter map, and the image features extracted based on the multi-channel image tensor yields fused features, including: The image features, diffusion features, and clinical text features are respectively mapped to a preset unified feature dimension to obtain a first mapped feature, a second mapped feature, and a third mapped feature; Using any one of the first, second, and third mapping features as a query, and the other two mapping features as keys and values, cross-modal attention calculations are performed to obtain the attention weighted features corresponding to each modality. The attention-weighted features corresponding to each modality are concatenated along the channel dimension to obtain the fused features.

6. The method for detecting the type of vascular occlusion according to claim 1, characterized in that, The determination of the classification probability for the type of vascular occlusion based on the fusion features includes: Perform a global average pooling operation on the fused features to obtain a pooled feature vector; The pooled feature vector is input into a preset fully connected classification layer to obtain a classification score vector; The classification score vector is normalized to obtain the classification probability of the object to be detected belonging to large vessel occlusion, medium vessel occlusion, and no occlusion.

7. The method for detecting the type of vascular occlusion according to claim 1, characterized in that, After determining the classification probability for a particular type of vascular occlusion based on the fusion features, the vascular occlusion type detection method further includes: Obtain the target prediction type with the highest probability value among the classification probabilities; If the probability value corresponding to the target prediction type is greater than the preset confidence threshold, the target prediction type is determined as the target classification result; Based on the target prediction type, the feature map of the last convolutional layer in the first preset convolutional network used to extract the image features is weighted to generate a spatial activation map. The spatial activation map is upsampled to the original input image space to obtain the upsampled activation map; The upsampling activation map is superimposed on the vascular enhancement image of the arterial phase CTA image to obtain a lesion localization heat map; The target classification results, the classification probability of the object to be detected belonging to each type of vascular occlusion, and the lesion localization heatmap are encapsulated into a structured report.

8. A device for detecting the type of blood vessel blockage, characterized in that, include: The data acquisition unit is used to perform multimodal preprocessing on non-contrast CT images and arterial phase CTA images to obtain multi-channel image tensors; Text encoding units are used to encode structured clinical text data to obtain clinical text features; An image diffusion unit is used to perform image generation processing based on the multi-channel image tensor to obtain a CT-like perfusion parameter map, wherein the CT-like perfusion parameter map includes a cerebral blood flow map, a cerebral blood volume map, and a mean transit time map. The feature fusion unit is used to perform cross-modal fusion of the clinical text features, the diffusion features extracted based on the CT-like perfusion parameter map, and the image features extracted based on the multi-channel image tensor to obtain fused features; The probability determination unit is used to determine the classification probability for the type of vascular occlusion based on the fusion features, wherein the type of vascular occlusion includes large vessel occlusion, medium vessel occlusion and no occlusion, the large vessel occlusion is an occlusion involving a vessel diameter greater than a first preset threshold, and the medium vessel occlusion is an occlusion involving a vessel diameter less than or equal to the first preset threshold and greater than a second preset threshold.

9. An electronic device, comprising: The memory and processor are characterized in that the processor, when executing a computer program stored in the memory, implements the steps of the method for detecting the type of vascular occlusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or the computer program are executed by a processor, the steps of the method for detecting the type of vascular occlusion as described in any one of claims 1 to 7 are implemented.

Images