An emergency early warning triage method and computer program product

Abstract

The application discloses an emergency early warning triage method and a computer program product, and relates to the technical field of medical image processing. The method comprises the following steps: acquiring image data and clinical text data of a patient; performing at least two levels of magnification processing on a plurality of two-dimensional slices in the image data to generate a multi-scale image slice sequence, performing feature extraction and fusion on the multi-scale image slice sequence to obtain image features; performing medical element recognition on the clinical text data and generating a text semantic feature vector by using a medical field pre-training language model; generating a comprehensive feature vector by a cross-modal attention fusion module through projection mapping, attention weight calculation and weighted aggregation; and performing risk assessment on a preset emergency type based on the comprehensive feature vector and outputting a triage result.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to an emergency early warning triage method and computer program product. Background Technology

[0002] Acute abdomen is a common critical illness in the emergency department, and the accuracy and timeliness of triage directly affect the patient's survival prognosis. Currently, clinical emergency CT imaging-assisted diagnosis mainly relies on manual image interpretation, which poses a risk of triage delays and misdiagnosis when there are insufficient staff on duty at night or on holidays.

[0003] Existing AI-based CT-assisted diagnosis techniques for acute abdominal pain typically employ three-dimensional voxel analysis, requiring convolution operations and feature extraction on the complete three-dimensional volumetric data of CT scans. This method is computationally complex, and a single analysis can take several minutes to complete, making it difficult to meet the time-sensitive requirements of emergency care. Furthermore, traditional three-dimensional analysis methods perform poorly on thick-slice CT scan data commonly used in emergency care, easily leading to a decrease in accuracy.

[0004] Another approach, based on two-dimensional slice classification, classifies individual CT slices independently. However, this method employs a single-scale image processing approach, which lacks sufficient sensitivity to early lesions such as minute ruptures and focal inflammation, leading to missed diagnoses. Furthermore, this approach completely ignores the contextual continuity between slices, failing to establish spatial relationships.

[0005] Furthermore, most existing auxiliary diagnostic systems are purely visual models, relying solely on image data for analysis, and cannot effectively integrate and utilize the clinical history text information entered by doctors. When faced with critical cases where imaging findings are atypical but clinical symptoms are typical, misdiagnosis or triage errors are likely to occur. Summary of the Invention

[0006] The purpose of this invention is to provide an emergency early warning and triage method and a computer program product to solve the technical problems existing in the prior art.

[0007] This invention provides an emergency triage and early warning method, comprising: acquiring patient image data; acquiring patient clinical text data; extracting features from the image data to generate image features; performing semantic parsing on the clinical text data to generate a text semantic feature vector; having a cross-modal attention fusion module guide the image features through cross-modal attention based on the text semantic feature vector to generate a comprehensive feature vector; and performing a risk assessment for at least one preset emergency type based on the comprehensive feature vector, and outputting triage results. This method deeply integrates clinical history text information with image features through a cross-modal attention fusion mechanism, achieving semantically guided regional focusing and improving the comprehensive judgment ability for cases with atypical imaging manifestations.

[0008] Furthermore, feature extraction is performed on multiple two-dimensional slices in the CT image data to generate image features. This includes: performing at least two levels of magnification processing on the two-dimensional slices in the CT image data to generate a multi-scale image slice sequence; and performing feature extraction and fusion on the multi-scale image slice sequence to obtain multi-scale image features, which are then used as the image features. By directly extracting features from the two-dimensional slice sequence, the high computational complexity of traditional three-dimensional voxel analysis is avoided, reducing the single triage time from minutes to seconds, meeting the timeliness requirements of emergency scenarios. Simultaneously, this multi-scale processing mechanism, through pyramid-style local magnification, enables the network to observe image features at three levels: global view, regional view, and local fine view. This significantly enhances the detection sensitivity for lesions with extremely small proportions, such as tiny rupture points and focal inflammation, effectively reducing the missed diagnosis rate of early critical signs.

[0009] Furthermore, feature extraction and fusion are performed on the multi-scale image slice sequence, including: extracting features from slices of different scales in the multi-scale image slice sequence to obtain visual features of the corresponding scales; integrating the visual features of different scales across scales to generate multi-scale fused features; and establishing contextual associations for the multi-scale fused features along the slice sequence dimension to generate the image features. This scheme, through cross-scale integration and temporal fusion of slice sequences, preserves the fine features of multiple scales while establishing spatial continuity information between adjacent slices, forming a unified feature representation that includes both global context and maintains local finesse.

[0010] Furthermore, contextual relationships are established for the multi-scale fused features at the slice sequence dimension, including: employing a sliding window attention mechanism to capture spatial continuity information between adjacent slices while maintaining computational efficiency. This mechanism avoids the high computational complexity of traditional global attention and effectively captures the temporal dependencies of the slice sequences while ensuring inference speed.

[0011] Furthermore, the clinical text data undergoes semantic parsing to generate text semantic feature vectors. This includes: identifying medical elements in the clinical text data and extracting at least one of anatomical location elements, organ name elements, symptom description elements, and medical history information elements; and mapping the extracted elements to text semantic feature vectors using a pre-trained language model in the medical field. The pre-trained language model can be based on the Transformer architecture. This approach accurately extracts key semantic elements through medical entity recognition and then uses a pre-trained language model to map discrete medical entities into high-dimensional continuous feature vectors containing rich clinical prior knowledge, providing a semantically guided foundation for subsequent cross-modal fusion.

[0012] Furthermore, the cross-modal attention fusion module guides the image features through cross-modal attention based on the text semantic feature vector to generate a comprehensive feature vector, including: calculating the attention weight of the text semantic feature vector to each spatial location in the image features, wherein the attention weight reflects the correlation between the text semantic feature vector and each spatial region in the image features; and weighting and aggregating the image features according to the attention weight to generate the comprehensive feature vector.

[0013] Specifically, the text semantic feature matrix (composed of multiple text semantic vectors) is mapped to a query matrix through linear projection, and the image features are mapped to a key matrix and a value matrix. The dot product of the query matrix and the transpose of the key matrix is ​​calculated and normalized to obtain the attention weight matrix of the text semantic feature matrix to each spatial location in the image feature map. The attention weight matrix reflects the correlation between the text semantic feature matrix and each spatial region in the image features. The value matrix is ​​weighted and aggregated according to the attention weight matrix to generate the comprehensive feature vector. This cross-attention mechanism enables the precise guidance of medical history semantics to image features, allowing anatomical regions with higher semantic weights in the medical history to receive stronger responses in the image features, while suppressing feature interference from regions unrelated to the medical history, simulating the diagnostic thinking of clinicians who "ask for medical history first, then look at the images."

[0014] Furthermore, the cross-modal attention fusion module is trained through the following pre-training steps: First, using image data, anatomical elements, and their corresponding spatial masks as training samples, strong-supervised spatial localization pre-training is performed using spatial localization loss; and / or using the pairing relationship between image data and clinical text data as training samples, weak-supervised global semantic contrast pre-training is performed using contrast loss. In this hierarchical pre-training strategy, the strong-supervised stage enables the model to accurately learn the localization ability of "entity word → image spatial region," while the weak-supervised stage supplements the implicit understanding ability of complex clinical semantics. The two complement each other and jointly support the efficient learning of cross-modal fusion.

[0015] Furthermore, the spatial mask is generated as follows: for anatomical orientation features, the corresponding region of interest is determined using a pre-set anatomical partition lookup table to generate a first spatial mask; for organ name features, a segmentation mask for the corresponding organ is generated using a pre-trained organ segmentation sub-network, serving as a second spatial mask. This dual-path spatial mapping mechanism ensures both rapid localization of anatomical orientation entities and accurate segmentation of organ name entities, providing high-quality pixel-level annotations for strongly supervised pre-training.

[0016] Furthermore, risk assessments are performed for at least one preset emergency type, and triage results are output, including: outputting classification judgment results for the preset emergency type; generating risk level scores based on the classification judgment results; and outputting visualized triage information, which includes at least one of lesion area localization results, triage reports, and priority markers. This multi-level output mechanism provides both quantitative risk scoring and classification judgments, and generates intuitive visualized localization and structured triage reports, comprehensively meeting the dual needs of emergency physicians for rapid decision-making and clinical record keeping.

[0017] Furthermore, the preset emergency types include at least one of acute aortic aneurysm rupture, organ rupture and hemorrhage, acute necrotizing pancreatitis, and strangulated intestinal obstruction. These four types of critical illnesses cover the most dangerous and rapidly progressing typical scenarios of acute abdominal conditions, specifically addressing the core pain points of emergency triage.

[0018] The present invention also provides a computer program product storing a computer program that, when executed by a processor, implements the method described in any of the preceding claims.

[0019] The beneficial effects of this invention are as follows: 1. By adopting a lightweight inference architecture based on two-dimensional slices, the traditional time-consuming three-dimensional voxel analysis is eliminated, reducing the single triage time from minutes to seconds, effectively meeting the timeliness requirements of high-concurrency emergency scenarios. At the same time, it has extremely high compatibility with thick-slice CT scan data commonly used in emergency departments, significantly improving the robustness of the model in actual clinical scenarios.

[0020] 2. By employing a pyramid-shaped multi-scale local magnification mechanism, features are extracted and integrated across three levels: global view, regional view, and local fine view. This enables the network to observe lesions that account for a very small proportion, such as tiny rupture points, focal inflammation, and early obstruction, in greater detail. This significantly enhances the sensitivity to early critical signs and effectively reduces the rate of missed diagnoses.

[0021] 3. Through a cross-modal attention fusion mechanism driven by medical history and imaging, deep semantic fusion of clinical medical history text information and imaging features is achieved. This enables the system to simulate the diagnostic thinking of clinicians who "ask for medical history first and then look at the images." It also targets and enhances anatomical areas with high semantic weight in the medical history, significantly improving the comprehensive judgment ability and triage accuracy of cases with atypical imaging manifestations but typical clinical symptoms.

[0022] 4. By employing a hierarchical strategy of strongly supervised spatial localization pre-training and / or weakly supervised global semantic contrast pre-training, the cross-modal fusion module not only accurately learns the localization ability of "entity word → image spatial region" but also supplements the implicit understanding ability of complex clinical semantics. The two complement each other and jointly support the model's excellent performance in the early warning and triage task of acute abdominal pain. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and their descriptions and are used to explain this specification, but do not constitute an undue limitation of this specification.

[0024] Figure 1 This is a flowchart illustrating an emergency early warning and triage method in one specific embodiment of this specification.

[0025] Figure 2 This is a simplified block diagram illustrating an exemplary computer program product that can be used to perform one or more tasks described herein. Detailed Implementation

[0026] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0027] The emergency triage method 100 provided in the embodiments of this application includes: acquiring patient image data 110; acquiring patient clinical text data 120; extracting features from the image data to generate image features 130; performing semantic parsing on the clinical text data to generate a text semantic feature vector 140; having a cross-modal attention fusion module guide the image features through cross-modal attention based on the text semantic feature vector to generate a comprehensive feature vector 150; and performing risk assessment for at least one preset emergency type based on the comprehensive feature vector, and outputting triage results 160.

[0028] This method directly extracts features from two-dimensional slice sequences, avoiding the high computational complexity of traditional three-dimensional voxel analysis. This reduces the triage time from minutes to seconds, meeting the timeliness requirements of emergency scenarios. Simultaneously, by employing a cross-modal attention fusion mechanism to deeply integrate clinical text information with image features, it achieves semantically guided region focusing, enhancing the comprehensive judgment ability for cases with atypical imaging presentations.

[0029] In one embodiment, acquiring patient image data specifically includes receiving emergency abdominal CT scan data and performing preprocessing such as window width and level adjustment, grayscale normalization, and size standardization. In other embodiments, the image data may also be of other modalities, such as image data from other body parts, and is not limited here.

[0030] Acquiring patients' clinical text data can include receiving patient clinical information input by doctors, such as chief complaints, symptom descriptions, past medical history, and physical examination results. Regarding the input methods for clinical text data, in addition to manual input by doctors as shown in the example, automatic capture and parsing of structured electronic medical records, or automatic text-to-speech technology after voice input, can also be used to further improve the convenience of clinical use.

[0031] Feature extraction from multiple two-dimensional slices in the image data includes: performing at least two levels of magnification processing on the two-dimensional slices in the image data to generate a multi-scale image slice sequence; and performing feature extraction and fusion on the multi-scale image slice sequence to obtain multi-scale image features, which are used as the image features. In one embodiment, this multi-scale processing mechanism, through pyramid-style local magnification, enables the network to observe image features at three levels: global view, regional view, and local fine view. This significantly enhances the detection sensitivity for lesions with extremely small proportions, such as tiny rupture points and focal inflammation, and effectively reduces the missed diagnosis rate of early critical signs.

[0032] Specifically, in one embodiment, multi-scale depth analysis is performed on the preprocessed two-dimensional CT slices. First, each two-dimensional slice is scaled to a standard input size of 448×448 pixels. Then, each slice undergoes multi-scale processing, the specific process of which is as follows: The first level is the global view, which takes the entire CT slice as input and extracts global anatomical structure information and overall spatial distribution features.

[0033] The second level is the regional field of view, which divides the CT slice into 2×2 sub-region blocks. Each sub-region is magnified to the same size as the original input to extract mesoscale regional anatomical details.

[0034] The third level is the local fine field of view, which further divides the slice into more fine 4×4 sub-region blocks. Each sub-region is further magnified to extract local fine lesion features.

[0035] Regarding the number of layers in the pyramid-shaped multi-scale slices, in the context of primary hospitals with limited computing resources, it can be simplified to two layers: the original image and a generated 2x magnified local slice, sacrificing some sensitivity in detecting small lesions for faster inference speed. In scenarios with extremely high requirements for detecting small lesions, such as aortic rupture screening, it can be expanded to four layers: the original image and generated 2x, 4x, and 8x magnified local slices, further enhancing the ability to extract fine local features.

[0036] Furthermore, feature extraction and fusion are performed on the multi-scale image slice sequence, including: extracting features from slices of different scales in the multi-scale image slice sequence to obtain visual features at the corresponding scales. In one embodiment, slices at each scale are fed into a visual encoder, such as a lightweight visual encoder like EfficientViT or MobileViT, to capture pathological features at different scales.

[0037] Regarding the selection of visual encoders, in addition to the lightweight EfficientViT scheme used in Example 1, large-scale pre-trained visual encoders such as InternViT and SigLIP can be used in high-precision schemes to improve feature representation capabilities and triage accuracy; in medical-specific schemes, domain-specific encoders such as BiomedCLIP, which are pre-trained with large-scale medical image data, can be used to enhance model performance by leveraging domain knowledge.

[0038] The visual features at different scales are integrated across scales to generate multi-scale fused features. Specifically, a combination of channel concatenation and attention weighting is used for multi-scale feature fusion: first, feature maps at each scale are concatenated along the channel dimension; then, a lightweight channel attention module automatically learns the importance weights of features at different scales, highlighting scale information with high contribution, thus forming a multi-scale fused feature map.

[0039] The image features are generated by establishing contextual associations for the multi-scale fused features along the slice sequence dimension. Specifically, the multi-scale fused feature map is processed through a sliding window attention mechanism to output an enhanced feature representation that incorporates multi-scale contextual information. This scheme, through cross-scale integration and temporal fusion of slice sequences, preserves the fine features of multiple scales while establishing spatial continuity information between adjacent slices, forming a unified feature representation that includes both global context and maintains local finesse.

[0040] In one embodiment, contextual associations are established for the multi-scale fused features along the slice sequence dimension, specifically employing a sliding window attention mechanism. This mechanism captures spatial continuity information between adjacent slices while maintaining computational efficiency. This mechanism avoids the high computational complexity of traditional global attention and effectively captures the temporal dependencies of the slice sequences while ensuring inference speed.

[0041] Furthermore, the clinical text data undergoes semantic parsing to generate text semantic feature vectors. This includes: identifying medical elements in the clinical text data and extracting at least one of anatomical location elements, organ name elements, symptom description elements, and medical history information elements; and mapping the extracted elements to text semantic feature vectors using a pre-trained language model in the medical field. The pre-trained language model can be based on the Transformer architecture. This approach accurately extracts key semantic elements through medical entity recognition and then uses a pre-trained language model to map discrete medical entities into high-dimensional continuous feature vectors containing rich clinical prior knowledge, providing a semantically guided foundation for subsequent cross-modal fusion.

[0042] First, the system performs structured semantic extraction on the clinical medical history input by the doctor. It receives textual information including the patient's chief complaint, symptom characteristics, past medical history, and physical examination results. Using a medical domain dictionary, it performs word segmentation and medical entity recognition to accurately extract key elements such as anatomical location, organ names, symptom descriptions, and medical history information. Then, using the ClinicalBERT pre-trained language model, these discrete medical entities are mapped into high-dimensional continuous textual semantic feature vectors.

[0043] For example, the system receives text information including the patient's chief complaint (e.g., "migratory right lower quadrant pain for 6 hours"), symptom characteristics (e.g., "knife-like pain", "accompanied by nausea and vomiting"), and past medical history (e.g., "history of gallstones"). First, it uses a medical domain dictionary for word segmentation and medical entity recognition (NER) to accurately extract key elements such as anatomical location (right lower quadrant), organ names (appendix, gallbladder, pancreas, etc.), symptom descriptions (nature of pain, accompanying symptoms, etc.), and medical history information. Then, using pre-trained medical domain language models such as ClinicalBERT and PubMedBERT, these discrete medical entities are mapped into high-dimensional continuous "textual semantic feature vectors," which contain rich prior clinical knowledge.

[0044] Furthermore, the cross-modal attention fusion module guides the image features through cross-modal attention based on the text semantic feature vector to generate a comprehensive feature vector. Specifically, this includes: calculating the attention weight of the text semantic feature vector to each spatial location in the image features, where the attention weight reflects the correlation between the text semantic feature vector and each spatial region in the image features; and weighting and aggregating the image features based on the attention weight to generate the comprehensive feature vector.

[0045] In one embodiment, the text semantic feature matrix (composed of multiple text semantic vectors) is mapped to a query matrix via linear projection, and the image feature map is mapped to a key matrix and a value matrix. The dot product of the query matrix and the transpose of the key matrix is ​​calculated and normalized to obtain the attention weight matrix of the text semantic feature matrix to each spatial location in the image feature map. The attention weight matrix reflects the correlation between the text semantic feature matrix and each spatial region in the image feature map. The value matrix is ​​weighted and aggregated according to the attention weight matrix to generate the comprehensive feature vector. This cross-attention mechanism enables the precise guidance of medical history semantics to image features, allowing anatomical regions with higher semantic weights in the medical history to receive stronger responses in the image features, while suppressing feature interference from regions unrelated to the medical history, simulating the diagnostic thinking of clinicians who "ask for medical history first, then look at the images."

[0046] In one embodiment, the cross-modal attention fusion module uses a multi-layer cross-attention mechanism to complete cross-modal feature fusion. The specific steps are as follows: First, it receives two inputs: a text semantic feature matrix parsed from clinical text and an image feature map.

[0047] Then, the text semantic feature matrix is ​​mapped to a query matrix and the image feature map is mapped to a key matrix and a value matrix, respectively, through three independent learnable linear projection layers. The query matrix encodes what semantics need to be focused on, while the key matrix and value matrix encode the features of each spatial location in the image.

[0048] Next, the attention scores of text semantics to each spatial location of the image are calculated according to the scaling dot product attention formula.

[0049] in, For querying the matrix, The key matrix, For value matrices, This is the transpose of the key matrix. The feature dimension of the key matrix is ​​used for scaling to prevent the inner product from becoming too large and causing gradient vanishing. The attention weight matrix obtained after softmax normalization reflects the correlation between each semantic element in the medical history and each spatial region of the image. Specifically, the dot product of the query matrix and the transpose of the key matrix is ​​calculated and then scaled by the square root of the feature dimension of the key matrix to prevent the inner product from becoming too large and causing gradient vanishing.

[0050] Subsequently, the value matrix is ​​weighted and aggregated using the obtained attention weight matrix, enabling anatomical regions with higher semantic weights in the medical history to receive stronger responses in image features, while suppressing feature interference from regions unrelated to the medical history, thus achieving semantically targeted feature enhancement. For example, when the medical history indicates "migratory right lower quadrant abdominal pain," the system automatically assigns higher weights to the region corresponding to the appendix, strengthening the feature expression of that region. The above cross-attention structure is stacked in four layers to iteratively optimize cross-modal alignment. In each layer, layer normalization is first performed on the input features, followed by cross-attention calculation; then, the attention output is fused with the original input features through residual connections, progressively deepening the guiding effect of medical history semantics on image features while maintaining the stability of the training process. Finally, after multi-layer cross-attention fusion, a fixed-dimensional comprehensive diagnostic feature vector is obtained. This vector simultaneously encodes the visual appearance features of the image and the prior semantic guidance of the clinical medical history.

[0051] In scenarios with limited pre-training data, the spatial mask generated by the text-image spatial mapping unit can be converted into a spatial attention bias matrix of the same size as the image feature map. Target regions are assigned a positive bias, while non-target regions are assigned a zero or negative bias. When calculating attention weights in the cross-modal attention fusion unit, the spatial attention bias matrix is ​​explicitly superimposed on the attention score, forcing the model to focus on anatomical regions consistent with the medical history semantics during inference, serving as a fallback compensation mechanism when pre-training is insufficient.

[0052] In one embodiment, the cross-modal attention fusion module is trained through the following pre-training steps: First, using image data, anatomical elements, and their corresponding spatial masks as training samples, strong-supervised spatial localization pre-training is performed using spatial localization loss; and / or using the pairing relationship between image data and clinical text data as training samples, weak-supervised global semantic contrast pre-training is performed using contrast loss. In this hierarchical pre-training strategy, the strong-supervised stage enables the model to accurately learn the localization ability of "entity word → image spatial region," while the weak-supervised stage supplements the implicit understanding ability of complex clinical semantics. The two complement each other and jointly support the efficient learning of cross-modal fusion.

[0053] Furthermore, the spatial mask is generated as follows: for anatomical orientation features, the corresponding region of interest is determined using a pre-set anatomical partition lookup table to generate a first spatial mask; for organ name features, a segmentation mask for the corresponding organ is generated using a pre-trained organ segmentation sub-network, serving as a second spatial mask. This dual-path spatial mapping mechanism ensures both rapid localization of anatomical orientation entities and accurate segmentation of organ name entities, providing high-quality pixel-level annotations for strongly supervised pre-training.

[0054] To enable the cross-modal attention fusion module to fully learn cross-modal alignment capabilities, one embodiment of this application designs a hierarchical pre-training strategy, which is executed sequentially in the following three stages: Phase 1 is the pre-training of strongly supervised spatial localization. Training samples include CT slices, anatomical entity words, and spatial mask triples. The spatial mask is automatically generated using the aforementioned method, serving as precise pixel-level strongly supervised annotation. The training objective is: given the semantic embedding of an anatomical entity word as a query, the model is required to predict the corresponding anatomical region on the CT slice. The Dice loss between the prediction mask and the annotation mask, along with the spatial contrast alignment loss, are used as the joint optimization objective. This phase enables the model to accurately learn the ability to localize entity words to image spatial regions.

[0055] Phase two involves weakly supervised global semantic contrastive pre-training. For clinical semantic information that cannot be precisely spatialized, such as the nature of pain, onset time, and past medical history, a weakly supervised image-text contrastive learning approach is adopted. Complete CT slice sequences paired with complete clinical text are used as positive samples, and random mismatches are used as negative samples. The model learns the implicit semantic association between symptom descriptions and image-pathological patterns through joint optimization of image-text contrastive loss and image-text matching loss. This phase enables the model to spontaneously learn the weak alignment ability from clinical symptom semantics to image region features through a large amount of paired data, even without precise spatial annotation.

[0056] Phase 3 involves fine-tuning the downstream task. Based on the pre-trained weights from Phases 1 and 2, the entire model is fine-tuned end-to-end using labeled CT data of acute abdominal pain. The multi-task classification loss is used as the optimization objective to adapt the model to the final task of acute abdominal pain early warning and triage.

[0057] In one embodiment, the spatial mask is generated as follows: A list of medical entities output by a clinical text parsing unit is received, categorized into anatomical orientation and organ name classes, and then mapped to different spatial paths. For anatomical orientation entities, a rectangular region of interest is directly retrieved using a pre-defined abdominal anatomy partition lookup table, generating a binary spatial mask. For organ name entities, the current CT slice is fed into a built-in lightweight organ segmentation subnetwork, outputting a pixel-level segmentation mask of the target organ within that slice. The generated spatial mask primarily serves as strongly supervised annotation for the visual localization task during the pre-training phase.

[0058] Furthermore, risk assessments are performed for at least one preset emergency type, and triage results are output, including: outputting classification judgment results for the preset emergency type; generating risk level scores based on the classification judgment results; and outputting visualized triage information, which includes at least one of lesion area localization results, triage reports, and priority markers. This multi-level output mechanism provides both quantitative risk scoring and classification judgments, and generates intuitive visualized localization and structured triage reports, comprehensively meeting the dual needs of emergency physicians for rapid decision-making and clinical record keeping.

[0059] The multi-task classification module employs a multi-task learning architecture, outputting independent binary classification judgments for four categories of critical illnesses in parallel. Each classification head in the module shares the comprehensive feature vector extracted by the aforementioned cross-modal attention fusion module for independent learning, ensuring detection specificity while leveraging inter-task correlations to improve overall performance. The classification module outputs predicted probability values ​​for each disease, where the threshold can be personalized by clinical experts according to their needs.

[0060] The risk level assessment unit integrates multi-task classification results with the severity of imaging signs to automatically generate clinical triage recommendations: when the probability of any critical illness category exceeds the high threshold or a clear imaging critical sign is detected, it is marked as critical, and it is recommended to immediately activate the emergency green channel; when the probability of any critical illness category exceeds the medium threshold or multiple suspicious signs are present, it is marked as high-risk, and it is recommended to prioritize emergency treatment; when there are some suspicious signs but they do not reach the threshold, it is marked as medium-risk, and it is recommended to observe in the general emergency department; when no obvious critical signs are detected, it is marked as low-risk, and it is recommended to follow up in a routine outpatient clinic.

[0061] The risk assessment unit further generates intuitive visual outputs, including: a heat map location map overlaid on the original CT slices, marking suspicious lesion areas of interest to the model, with color intensity indicating the degree of attention; a triage report containing basic patient information, positive probabilities of various critical illnesses, triage level recommendations, descriptions of key suspicious areas, etc.; and timestamps and priority markers to facilitate quick identification and execution by emergency triage nurses.

[0062] The preset acute condition types include at least one of acute aortic aneurysm rupture, organ rupture and hemorrhage, acute necrotizing pancreatitis, and strangulated intestinal obstruction. Regarding the expansion of targeted disease types, by increasing training data and expanding the output dimensions of the classification head, the triage coverage can be extended to other subtypes of acute abdomen, including but not limited to: acute appendicitis perforation, gastrointestinal perforation with free gas in the abdominal cavity, acute mesenteric artery embolism, and ruptured ectopic pregnancy with hemorrhage.

[0063] This embodiment of the emergency triage method achieves automated early warning and triage for four types of critical and severe illnesses—acute aortic aneurysm rupture, organ rupture and hemorrhage, acute necrotizing pancreatitis, and strangulated intestinal obstruction—through a deep integration of rapid two-dimensional slice processing, pyramid-style multi-scale feature enhancement, and cross-modal fusion of medical history and imaging. The single triage time is reduced from minutes to seconds, the sensitivity to detecting minute lesions is significantly improved, and the triage accuracy meets clinical application requirements. These four types of critical and severe illnesses cover the most dangerous and rapidly progressing typical scenarios of acute abdominal conditions, specifically addressing the core pain points of emergency triage.

[0064] The present invention also provides a computer program product 500, including a computer program or instructions that, when executed by a processor, can implement the steps of the method described in any of the above embodiments. Figure 2 This is a block diagram illustrating an exemplary product 500 that can be configured to perform the tasks described herein. As shown, product 500 may include a processor (e.g., one or more processors) 502, which may be a central processing unit (CPU), a graphics processing unit (GPU), a microcontroller, a reduced instruction set computer (RISC) processor, an application-specific integrated circuit (ASIC), an application-specific instruction set processor (ASIP), a physical processing unit (PPU), a digital signal processor (DSP), a field-programmable gate array (FPGA), or any other circuitry or processor capable of performing the functions described herein. Product 500 may further include communication circuitry 504, memory 506, mass storage device 508, input device 510, and / or communication link 512 (e.g., a communication bus) on which one or more components shown in the figure may exchange information.

[0065] Communication circuitry 504 may be configured to transmit and receive information using one or more communication protocols (e.g., TCP / IP) and one or more communication networks, including local area networks (LANs), wide area networks (WANs), the Internet, and wireless data networks (e.g., Wi-Fi, 3G, 4G / LTE, or 5G networks). Memory 506 may include a storage medium (e.g., a non-transitory storage medium) configured to store machine-readable instructions that, when executed, cause processor 502 to perform one or more functions described herein. Examples of machine-readable media may include volatile or non-volatile memories, including but not limited to semiconductor memories (e.g., electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM)), flash memory, and / or the like. Mass storage device 508 may include one or more disks, such as one or more internal hard disks, one or more removable disks, one or more magneto-optical disks, one or more CD-ROM or DVD-ROM disks, etc., on which instructions and / or data may be stored to facilitate operation of processor 502. Input device 510 may include a keyboard, mouse, voice input device, touch input device (e.g., touch screen) and / or similar devices for receiving user input from device 500.

[0066] It should be noted that Product 500 can operate as a standalone device or be connected to other computing devices (e.g., networked or clustered) to perform the functions described herein. Furthermore, even in Figure 2 Only one instance of each component is shown in the figure. Those skilled in the art will understand that product 500 may include multiple instances of one or more components shown in the figure.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these 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 the present invention.

Claims

1. An emergency triage and early warning method, characterized in that, include: Acquire patient imaging data; Obtain the patient's clinical text data; Feature extraction is performed on the image data to generate image features; The clinical text data is semantically parsed to generate text semantic feature vectors; The cross-modal attention fusion module guides cross-modal attention on the image features based on the text semantic feature vector to generate a comprehensive feature vector. Based on the comprehensive feature vector, a risk assessment is performed for at least one preset emergency type, and triage results are output.

2. The method according to claim 1, characterized in that, Feature extraction is performed on the image data to generate image features, including: The 2D slices in the image data are magnified at least two levels to generate a multi-scale image slice sequence. Feature extraction and fusion are performed on the multi-scale image slice sequence to obtain multi-scale image features, which are used as the image features.

3. The method according to claim 2, characterized in that, Feature extraction and fusion of the multi-scale image slice sequence includes: Feature extraction is performed on slices of different scales in the multi-scale image slice sequence to obtain visual features of the corresponding scales; The visual features at different scales are integrated across scales to generate multi-scale fusion features; The image features are generated by establishing contextual associations for the multi-scale fusion features along the slice sequence dimension.

4. The method according to claim 1, characterized in that, The clinical text data is semantically parsed to generate text semantic feature vectors, including: The clinical text data is subjected to medical element identification, and at least one of the following elements is extracted: anatomical location elements, organ name elements, symptom description elements, and medical history information elements. The extracted elements are mapped into text semantic feature vectors using a pre-trained language model in the medical field.

5. The method according to claim 1, characterized in that, The cross-modal attention fusion module guides cross-modal attention on the image features based on the text semantic feature vector to generate a comprehensive feature vector, including: Calculate the attention weight of the text semantic feature vector to each spatial location in the image feature, whereby the attention weight reflects the degree of correlation between the text semantic feature vector and each spatial region in the image feature; The image features are weighted and aggregated according to the attention weights to generate the comprehensive feature vector.

6. The method according to claim 1, characterized in that, The cross-modal attention fusion module is trained through the following pre-training steps: Supervised pre-training is performed using image data, anatomical features, and their corresponding spatial masks as training samples, through spatial localization loss; and / or The pairing relationship between image data and clinical text data is used as training samples, and weakly supervised pre-training is performed using contrastive loss.

7. The method according to claim 6, characterized in that, The spatial mask is generated in the following manner: For anatomical orientation features, the corresponding region of interest is determined using a pre-defined anatomical partition lookup table, generating a first spatial mask; and / or For organ name elements, a segmentation mask for the corresponding organ is generated through an organ segmentation subnetwork, which serves as a second spatial mask.

8. The method according to claim 1, characterized in that, Perform risk assessment for at least one pre-defined emergency type and output triage results, including: Output the classification results for the preset emergency type; A risk level score is generated based on the classification results. Output visual triage information, which includes at least one of lesion area localization results, triage report and priority marker.

9. The method according to claim 1, characterized in that, The preset emergency types include at least one of acute aortic aneurysm rupture, organ rupture and hemorrhage, acute necrotizing pancreatitis, and strangulated intestinal obstruction.

10. A computer program product comprising instructions that, when executed on a computer, cause the computer to perform the method as claimed in any one of claims 1 to 9.

Images