Nasointestinal tube positioning examination method and system based on abdominal X-ray plain film and medium

Abstract

The invention discloses a nasointestinal tube positioning examination method and system based on an abdominal X-ray plain film and a medium, and relates to the technical field of medical image processing.The method comprises the steps that an abdominal X-ray plain film image nasointestinal tube positioning model is established, and an image of an abdominal X-ray plain film of a target patient is obtained through an HIS system; an AI intelligent examination engine is developed to recognize, analyze and interpret images and clinical information of X-ray plain films of the abdomen of a target patient, a virtual human anatomical model is generated based on a CT scanning template, a result is output, and when it is detected that the catheter position is abnormal, a real-time early warning mechanism immediately triggers a grading response process, and an HIS system pops up a positioning result prompt. According to the method, an image and text double-path fusion model is constructed, nasointestinal tube positioning is remarkably optimized, RoBERTa-wwm-ext is combined with BiLSTM-CRF, position and morphological features are extracted from a report, implicit inference and cross-modal attention dynamic weighted image-text features are supported, the comprehensive confidence is composed of image matching, text semantics and abnormal modes, and the misjudgment risk is reduced through multi-source complementation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, more particularly to a nasojejunal tube positioning review method and system based on abdominal X-ray plain film and a medium. BACKGROUND

[0002] As a critical access for nutrition support in critical patients, the tip of the nasojejunal tube needs to be accurately placed in the duodenum or jejunum to reduce the risk of reflux and aspiration. Currently, abdominal X-ray plain film examination is recognized as the "gold standard" and final confirmation method for nasojejunal tube positioning. However, its clinical application still faces the following problems:

[0003] The spatial relationship between the catheter tip and the anatomical landmarks needs to be analyzed in multiple dimensions (such as oblique film angle), and non-radiology professionals are prone to misjudgment due to lack of experience.

[0004] From image acquisition, report generation to manual review, the average time consumption is more than 30 minutes, especially in the ICU and other scenes that require emergency intervention, which may delay treatment.

[0005] Radiology professionals often need to handle a large number of routine examinations, making it difficult to respond to critical patient emergency needs in real time.

[0006] Therefore, it is necessary to propose a nasojejunal tube positioning review method based on abdominal X-ray plain film to solve the above problems. SUMMARY

[0007] (I) Technical problems solved

[0008] The purpose of the present application is to solve the problems of existing nasojejunal tube positioning methods, such as misjudgment due to lack of experience, long process and delay of treatment, and difficulty for full-time professionals to respond to critical emergencies. The present application provides a nasojejunal tube positioning review method based on abdominal X-ray plain film, a system and a medium.

[0009] (II) Technical solutions

[0010] In order to achieve the above purpose, the present application specifically adopts the following technical solutions:

[0011] A nasojejunal tube positioning review method based on abdominal X-ray plain film, comprising the following steps:

[0012] Step 1: Establish an abdominal X-ray plain film image nasojejunal tube positioning model and complete the corresponding coding. The model is a double-path feature coding model.

[0013] Step 2: Obtain the image and clinical report information of the target patient's abdominal X-ray plain film through the HIS system.

[0014] Step three: develop an AI intelligent review engine to identify and analyze the image and clinical information of the target patient's abdominal X-ray film. Based on the CT scan template, a virtual human anatomy model is generated, the two-dimensional X-ray film is projected and converted into a three-dimensional coordinate system, the catheter path is reconstructed, and the results are output. Set multiple confidence threshold values, provide graded warnings for different clinical scenarios, and make dynamic threshold judgments. Collect typical misplacement cases (such as catheter coiling in the stomach and tip entering the chest cavity), and use generative adversarial networks (GAN) to generate simulated data to enhance model robustness and build an abnormal pattern library.

[0015] Step four: When the catheter position is detected to be abnormal, the real-time warning mechanism triggers the graded response process immediately, and the HIS system pops up a positioning result reminder to guide clinical work.

[0016] Further, the dual-path feature encoding model includes an image feature extraction module, a text semantic analysis module, and a feature fusion layer.

[0017] The image feature extraction module uses an improved U-Net network structure, expands the receptive field through dilated convolution, and focuses on the relative positions of the catheter tip and the lower edge of the diaphragm and the right edge of the spine through the attention gate mechanism.

[0018] The text semantic analysis module extracts the position description of the nasoenteric tube from the radiology report text and generates a text vector containing position coordinates and shape description.

[0019] Feature fusion layer: dynamically weight image features and text features through cross-modal attention mechanism, and output comprehensive confidence score through gate fusion strategy.

[0020] Further, the improved U-Net network structure includes a deep encoder, which is an extension of the traditional U-Net 4-layer encoder to 6 layers, each containing 2 convolution blocks. The deep structure extracts multi-scale features from local texture to global structure, for example: the first layer focuses on the edge details of the catheter, and the sixth layer can identify the overall profile of the diaphragm.

[0021] Dilated convolution is introduced in the third and fifth layers of the encoder through "dilation rate" to expand the receptive field. The jump connection is optimized to retain the up-sampling jump connection of U-Net, but the connection method is adjusted. After the fourth layer of the decoder, residual connection is added to directly inject the middle layer features of the encoder into the corresponding position of the decoder, enhancing the continuity judgment of the curved catheter.

[0022] The attention gate mechanism includes spatial attention focusing. After the fourth layer of the decoder, an attention gate module is inserted, which receives high-level feature maps (such as the diaphragm region) from the encoder and low-level feature maps (such as the catheter tip) from the decoder. The following steps are taken to generate a gate signal, including:

[0023] Global average pooling is performed on the two feature maps to generate a global description vector;

[0024] The interaction weights of the two are learned through a fully connected layer (FC);

[0025] After activation by the Sigmoid function, a gating coefficient of 0-1 is obtained, which is used to suppress regions unrelated to intestinal gas artifacts and highlight the key landmarks of the diaphragmatic inferior border.

[0026] Further, the text semantic analysis module selects the RoBERTa-wwm-ext model, freezes the bottom layer parameters and only trains the top classification head, to improve the recognition ability of Chinese medical terms; In addition to the standard classification task, the text generation task is additionally trained, so that the model is better at inferring key information from incomplete reports;

[0027] The text semantic analysis module also includes BiLSTM sequence modeling, which sends the 768-dimensional hidden state output by BERT into a bidirectional LSTM to ensure that it can utilize both the anatomical structure description before "catheter tip located" and the functional description after it;

[0028] The text semantic analysis module also includes a CRF conditional random field decoding unit; linear chain CRF is used to define label transition probabilities; both position and shape information are recognized simultaneously, and the optimal combination is selected through joint probability maximization.

[0029] Further, when the cross-modal attention mechanism dynamically weights the image features and text features, it first performs double-channel feature representation; the image channel extracts the visual features of the catheter tip region (ROI) from the segmentation map and generates a fixed-length vector through region pooling; the text channel performs maximum pooling operation on the text vector to retain the core semantic information of the key position description; dynamic weight distribution is achieved by comparing the relevance of image ROI and text keywords through an attention network.

[0030] The gating fusion strategy includes a dynamic mixing rule and confidence calculation. The dynamic mixing rule dynamically adjusts the fusion ratio according to the reliability of the two modalities.

[0031] The confidence calculation method is: the final score is the weighted sum of the image matching score, the text semantic matching score, and the abnormal pattern matching score; the image matching score.

[0032] Further, the virtual human anatomy model is constructed as follows:

[0033] CT reference model generation: Extract anonymized chest CT scan data from the hospital PACS system and convert it to a three-dimensional mesh model using voxelization;

[0034] Anatomical landmark registration: mapping the key structures of bones and diaphragm in CT model to the standard human coordinate system through registration algorithm;

[0035] Personalized calibration: affine transformation of virtual model using anatomical landmarks in X-ray for current patient image, compensating for individual body shape differences.

[0036] Further, the catheter path reconstruction method is as follows:

[0037] Multi-view projection matching: X-ray is regarded as the orthogonal projection view of the virtual model, and the coordinates of the catheter tip in three-dimensional space are optimized through bundle adjustment;

[0038] Bézier curve fitting: sampling the catheter pixel points in the segmentation map, using cubic Bézier curve interpolation to generate a smooth path, and calculating the curvature radius to judge the risk of coiling;

[0039] Depth verification mechanism: if the position difference of the same pixel point in the front and back two X-rays exceeds 3mm, it is marked as suspicious artifact;

[0040] The output results include the three-dimensional coordinates of the catheter tip, the path curvature report and the matching degree of anatomical landmarks.

[0041] Further, the real-time warning mechanism includes HIS pop-up warning, dynamic intervention suggestion generation and result output;

[0042] HIS pop-up warning is that when the confidence is lower than the preset threshold, an emergency alarm message is pushed to HIS through HL7 v2.x interface, including: sound and light alarm linkage, triggering bed head monitor vibration + LED red light flashing in high-risk departments such as ICU; mobile terminal push: sending a text and image message containing treatment suggestions to the responsible nurse through the hospital's internal APP;

[0043] Dynamic intervention suggestion generation includes knowledge graph reasoning, dose calculation and contraindication check;

[0044] Knowledge graph reasoning: building a treatment path network based on Neo4j graph database;

[0045] Dose calculation: if the depth needs to be adjusted, the safe range is automatically calculated;

[0046] Contraindication check: if the patient has "duodenal ulcer", the recommended advancement operation is prohibited;

[0047] The result output is to form a standardized intervention instruction set, including operation priority, specific treatment steps and time window limit information.

[0048] A system of a nasogastric tube positioning review method based on abdominal X-ray plain film, the system comprises:

[0049] Data acquisition and transmission module: including image acquisition device interface for connecting existing X-ray machine or PACS system of hospital, supporting DICOM standard, automatically obtaining abdominal X-ray plain film image data of target patient;Including HIS / EMR system integration interface, realizing seamless docking with hospital information system through HL7 FHIR or HL7 v2.x medical information exchange standard, automatically extracting patient clinical information such as medical history and catheter record;

[0050] Data processing and analysis module: including double-path feature coding model server, built-in improved U-Net network structure and BERT-BiLSTM-CRF hybrid model, responsible for processing image data segmentation, identification and report text analysis, output structured position information and confidence score;Including three-dimensional space modeling engine: generating virtual human anatomy model based on CT scanning template, and converting two-dimensional X-ray film projection into catheter path reconstruction in three-dimensional coordinate system by using computer vision algorithm;

[0051] Decision support and feedback module includes AI intelligent review engine, integrates dynamic threshold judgment logic and abnormal pattern library, comprehensively evaluates catheter position, generates positioning report and graded early warning suggestion;Including real-time early warning system, when detecting catheter position abnormality, reminding medical staff through HIS system pop-up window alarm, sound and light alarm and various ways;Including knowledge graph query interface, providing visual interface for doctors to query ESPEN guideline recommended treatment scheme, assisting clinical decision-making;

[0052] User interaction and management module: including operation terminal, operation terminal includes bedside mobile terminal and central control room large screen, displays three-dimensional reconstruction result, early warning information and treatment suggestion;Including permission management system, ensuring that users of different levels access corresponding function areas, protecting patient privacy and data security.

[0053] A computer readable storage medium, the computer storage medium stores a computer program, the computer program includes program instructions, the program instructions are executed by the processor when the processor executes the above-mentioned method of any one of the method.

[0054] (Three) beneficial effects

[0055] The beneficial effects of the present application are as follows:

[0056] 1、The present application significantly optimizes the positioning of the nasointestinal tube by constructing an image-text dual-path fusion model, and improves the U-Net to expand the encoder to 6 layers, each layer containing 2 convolution blocks, and uses the empty convolution to make the receptive field reach 13 pixels, accurately identifies the distant landmark of the lower edge of the diaphragm; the decoder introduces attention gate, generates coefficients to suppress intestinal gas artifacts, highlights key landmarks, uses RoBERTa-wwm-ext combined with BiLSTM-CRF to extract position and shape features from the report, supports implicit inference, dynamically weights the cross-modal attention graph-text features, and the comprehensive confidence is composed of image matching, text semantics and abnormal patterns, and multiple sources complement each other to reduce the risk of misjudgment.

[0057] 2、The present application generates a three-dimensional grid model based on CT scan data, and maps the skeleton and diaphragm structure to the standard coordinate system through the ICP algorithm, ensuring the consistency of the model with the real anatomical structure; the position of the thoracic vertebrae and the angle of the xiphoid in the patient's X-ray film are used as landmarks for affine transformation, compensating for individual body size differences; the catheter tip coordinates are optimized by multi-view projection matching combined with the beam method, reducing the three-dimensional positioning error; the catheter pixel points in the segmented image are used to calculate the radius of curvature to judge the risk of entanglement; if the same pixel point has a displacement of more than 3mm in the front and rear X-ray films, it is marked as a suspicious artifact to avoid misjudgment.

[0058] 3、The present application realizes the automatic closed-loop management of the diagnosis and treatment process through real-time hierarchical early warning and intelligent intervention, when the confidence is lower than the threshold, triggers the HIS pop-up window warning, sound and light alarm and mobile terminal push, ensures that high-risk departments receive the alarm within 1 second, and improves the response timeliness; based on the Neo4j knowledge graph, a treatment path network is constructed, and a standardized instruction set is generated combined with the dose calculation formula, the operation priority is clear, and the disposal steps are refined. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The present application is based on the abdominal X-ray plain film nasointestinal tube positioning review method. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0061] Embodiment 1

[0062] Please refer to FIG. 1, a nasointestinal tube positioning review method based on abdominal X-ray plain film, comprising the following steps:

[0063] Step one: Establish an abdominal X-ray plain film image nasojejunal tube positioning model, complete the corresponding coding, and the model is a double-path feature coding model;

[0064] Step two: Obtain the image and clinical report information of the target patient's abdominal X-ray plain film through the HIS system;

[0065] Step three: Develop an AI intelligent review engine to identify and analyze the image and clinical information of the target patient's abdominal X-ray plain film. Based on the CT scan template, a virtual human anatomy model is generated, the two-dimensional X-ray film is projected and converted into a three-dimensional coordinate system, the catheter path is reconstructed, and the results are output. Set multiple confidence threshold values (such as 85%, 95%), provide graded warnings for different clinical scenarios (such as ICU intensive care, general ward), and perform dynamic threshold determination. Record typical misplacement cases (such as catheter coiled in the stomach, sharp end entering the chest), generate simulated data through generative adversarial networks (GAN) to strengthen the robustness of the model, and build an abnormal pattern library;

[0066] Step four: When the catheter position is detected to be abnormal, the real-time warning mechanism triggers the graded response process immediately, and the HIS system pops up a positioning result reminder to guide clinical work.

[0067] The double-path feature coding model includes an image feature extraction module, a text semantic analysis module, and a feature fusion layer;

[0068] The image feature extraction module uses an improved U-Net network structure, which expands the receptive field through dilated convolution and focuses on the relative positions of the catheter tip, diaphragm lower edge, and spinal right edge key anatomical landmarks through the attention gate mechanism;

[0069] The text semantic analysis module extracts the nasojejunal tube position description from the radiology report text and generates a text vector containing position coordinates and shape description;

[0070] The feature fusion layer dynamically weights the image features and text features through the cross-modal attention mechanism (Cross-Modal Attention) and outputs the comprehensive confidence score through the gate fusion strategy.

[0071] The improved U-Net network structure includes a deep encoder, which is an extension of the traditional U-Net's 4-layer encoder to a 6-layer encoder, each layer containing 2 convolution blocks. The deep structure extracts multi-scale features from local textures to global structures, such as focusing on the catheter edge details in the first layer and identifying the diaphragm overall profile in the sixth layer.

[0072] The empty convolution is introduced in the 3rd and 5th layers of the encoder, expanding the receptive field through the "dilation rate"; for example, the 3rd layer uses a 3x3 convolution kernel with a dilation rate of 2, at which time the actual receptive field is 3x3 + 2x(3x3) = 9 pixels (covering a larger range around the catheter tip); the 5th layer has a dilation rate of 4, and the receptive field is further expanded to 13 pixels, ensuring that the diaphragm lower edge located at the T8-T9 level (about 50-80 mm from the catheter tip) can be captured; optimization is performed through a skip connection, retaining the up-sampling and down-sampling skip connections of the U-Net, but adjusting the connection method, adding a residual connection after the 4th layer of the decoder, directly injecting the middle layer features of the encoder into the corresponding position of the decoder, and enhancing the continuity judgment of the curved catheter;

[0073] The attention gate mechanism includes spatial attention focusing, which is inserting an attention gate module after the 4th layer of the decoder, the attention gate module receiving high-level feature maps of the encoder (such as the diaphragm area) and low-level feature maps of the decoder (such as the catheter tip), generating a gating signal through the following steps, including:

[0074] Global average pooling is performed on the two feature maps to generate a global description vector;

[0075] The interaction weight between the two is learned through a fully connected layer (FC);

[0076] After being activated by the Sigmoid function, a gating coefficient of 0-1 is obtained, which is used to suppress the area unrelated to the intestinal gas artifact and highlight the key markers of the diaphragm lower edge;

[0077] The output result is a segmentation map with a resolution of 128x128, including a green mark for the catheter tip (pixel coordinates such as [64, 102]), a red cross mark for the diaphragm lower edge, and a yellow dashed line frame for the right edge of the spine.

[0078] The text semantic analysis module selects the RoBERTa-wwm-ext model (based on BERT, more suitable for medical field short text), freezes the bottom layer parameters and only trains the top classification head, to improve the recognition ability of Chinese medical terms (such as "pylorus" and "jejunum" professional vocabulary); In addition to the standard classification task, the text generation task (such as automatically completing the missing information: "the catheter tip is located in the descending part of the duodenum") is additionally trained, so that the model is better at inferring key information from incomplete reports;

[0079] The text semantic analysis module also includes BiLSTM sequence modeling, which sends the 768-dimensional hidden state output by BERT into a bidirectional LSTM to ensure that it can utilize both the anatomical structure description (such as "gastric bubble visible") before the "catheter tip is located in" and the functional description (such as "suggestion to maintain the current depth") after it; for example, when the "gastric bubble" keyword is detected, the BiLSTM will associate it with the "pylorus" or "duodenum" location description that appears after it;

[0080] The text semantic analysis module further comprises a CRF conditional random field decoding unit; a linear chain CRF is used to define label transition probability; for example, if the previous word is “located”, the next word is more likely to be “duodenum” than “gastric antrum”, and if the keyword “coiled” appears, the morphological abnormality is automatically marked and the position confidence is reduced; both position (such as “duodenal descending part”) and shape (such as “no coiling”) information are supported, and the optimal combination is selected through joint probability maximization.

[0081] When the cross-modal attention mechanism dynamically weights the image features and the text features, first, a double-channel feature representation is performed; the image channel extracts the visual features (such as shape and contrast) of the catheter tip region (ROI) from the segmentation map and generates a fixed-length vector through region pooling; the text channel performs a max-pooling operation on the text vector to retain the core semantic information of the key position description; dynamic weight distribution is performed through an attention network to compare the relevance of the image ROI and the text keywords; for example, if the text mentions “duodenum”, the catheter tip near the region in the image will be given a higher weight; conversely, if the text emphasizes “gastric bubble”, the catheter segment in the stomach region of the image will be focused on;

[0082] The gating fusion strategy comprises a dynamic mixing rule and a confidence calculation, the dynamic mixing rule dynamically adjusts the fusion proportion according to the reliability of the two modalities; for example, when the image clarity is high (signal-to-noise ratio > 30 dB) and the text description is clear, a 7:3 image-dominant fusion is adopted; when there is long-term interference causing image blur, the text weight is increased to 6:4.

[0083] The confidence calculation method is: the final score is the weighted sum of the image matching score (such as “the tip is consistent with the duodenum position” gets 85 points), the text semantic matching score (such as “the end is located in the jejunum” gets 90 points) and the abnormal mode matching score (such as no “misplaced into the chest cavity” keyword is detected, which gets 100 points); the image matching score.

[0084] This embodiment significantly improves the accuracy and efficiency of nasogastric tube positioning by constructing an image and text dual-path fusion model. The improved U-Net network expands the encoder to 6 layers, each containing 2 convolution blocks, and combines with the dilated convolution (such as the 3x3 convolution kernel with an expansion rate of 4 in the 5th layer) to expand the receptive field to 13 pixels, accurately identifying the distant anatomical landmarks of the diaphragm below the T8-T9 level, and the catheter tip positioning accuracy is high; the decoder introduces an attention gate mechanism, generates 0-1 gating coefficients through global average pooling and fully connected layers, suppresses irrelevant areas of intestinal gas artifacts, highlights key landmarks such as the diaphragm and right edge of the spine, and improves the annotation accuracy of the segmentation map; the RoBERTa-wwm-ext pre-training model is used in combination with the BiLSTM-CRF architecture to extract location descriptions (such as "duodenum descending part") and morphological features (such as "no coil") from unstructured reports, supporting implicit information inference (such as automatic completion of missing descriptions), and the text analysis accuracy is high; through the cross-modal attention mechanism, the image and text features are dynamically weighted, and the comprehensive confidence score is composed of three parts (image matching 85 points + text semantics 90 points + abnormal mode 100 points), realizing the complementarity of multiple sources of evidence and reducing the risk of misjudgment of a single modality.

[0085] Embodiment 2

[0086] This embodiment is further optimized based on Embodiment 1, specifically, the virtual human anatomy model is constructed as follows:

[0087] CT reference model generation: Extract anonymized chest CT scan data (DICOM format) from the hospital PACS system, and convert it to a three-dimensional mesh model (resolution 5mm x 5mm x 5mm) using voxelization;

[0088] Anatomical landmark registration: Map the key structures of bones and diaphragm in the CT model to the standard human coordinate system (such as the VRML coordinate system specified by DICOM) through registration algorithms (such as the ICP iterative closest point algorithm);

[0089] Personalized calibration: For the current patient image, use the anatomical landmarks in the X-ray film (such as the position of the thoracic vertebrae and the angle of the xiphoid process) to perform affine transformation on the virtual model, compensating for individual body shape differences.

[0090] Catheter path reconstruction is as follows:

[0091] Multi-view projection matching: Treat the X-ray film as an orthogonal projection view of the virtual model, and optimize the coordinates of the catheter tip in three-dimensional space through bundle adjustment. For example, the frontal film determines the horizontal position; the lateral film determines the vertical position; the oblique film assists in verifying spatial consistency.

[0092] Bézier curve fitting: sampling catheter pixels in segmentation map, generating smooth path using cubic Bézier curve interpolation, calculating curvature radius to judge coiling risk.

[0093] Deep verification mechanism: if the same pixel point has a position difference of more than 3 mm in the front and back two X-ray films, it is marked as a suspicious artifact (such as intestinal gas interference);

[0094] The output results include the three-dimensional coordinates of the catheter tip (such as [120, -80, 30] mm), the path curvature report (such as "curvature radius 5 cm, no significant coiling"), and the matching degree of anatomical landmarks (such as "deviation ± 2 mm from the lower edge of T8 vertebra").

[0095] The abnormal pattern library architecture includes a GAN adversarial training framework, an abnormal pattern injection strategy, and a continuous learning mechanism.

[0096] The GAN adversarial training framework includes a generator network using the StyleGAN2 architecture, which inputs real misplacement case image slices + text description and outputs high-resolution synthetic images and fake reports. For example, input: "catheter tip located in right main bronchus" (with lung texture noise); output, simulated X-ray film (with bronchial inflation artifact) + fake report "end located in right lower lobe"; it also includes a discriminator network, a dual-channel discriminator that evaluates image authenticity (convolutional neural network) and text rationality (BERT classifier);

[0097] The abnormal pattern injection strategy includes physical constraint injection, which forces anatomical contradictions (such as "catheter passing through aortic arch") into the generation process, forcing the model to learn the feature distribution of abnormal patterns; it also attacks adversarial samples, regularly injecting adversarial samples (such as slightly perturbed X-ray brightness) generated by the FGSM algorithm into the training set, simulating misjudgment under extreme lighting conditions;

[0098] The continuous learning mechanism includes an incremental training pipeline that automatically downloads the latest misplacement cases (desensitized data after ethical review) every week and fine-tunes model parameters through transfer learning; it also includes a meta-learning framework that uses the MAML algorithm to pre-train a general abnormality detector, which only requires a small number of samples to adapt to specific department needs.

[0099] This embodiment solves the depth missing problem of traditional two-dimensional interpretation through virtual human body models and three-dimensional reconstruction technology.

[0100] A three-dimensional mesh model is generated based on CT scan data, and the bone and diaphragm structures are mapped to a standard coordinate system through an ICP algorithm to ensure consistency with the real anatomical structure. The position of the thoracic vertebrae and the angle of the xiphoid in the patient's X-ray are used as landmarks for affine transformation to compensate for individual body size differences. The three-dimensional positioning error is reduced by matching the catheter tip coordinates through multi-view projection matching (anterior / side / oblique view) combined with bundle adjustment optimization. The curvature radius is calculated by using the catheter pixel points in the segmentation image to judge the risk of coiling. If the same pixel point has a displacement of more than 3 mm in the front and rear X-ray, it is marked as a suspicious artifact (such as intestinal gas interference) to avoid misjudgment.

[0101] The output results include three-dimensional coordinates (such as [120, -80, 30] mm), anatomical landmark matching degree (such as T8 vertebral body deviation ± 2 mm), and curvature report, providing quantitative reference for different clinical scenarios (such as ICU intensive care).

[0102] Example 3

[0103] This embodiment is further optimized based on Example 2, specifically, the real-time warning mechanism includes HIS pop-up warning, dynamic intervention suggestion generation and result output;

[0104] The HIS pop-up warning is that when the confidence is lower than the preset threshold (such as 85%), an emergency alarm message (FIR message type) is pushed to the HIS through the HL7 v2.x interface, including: sound and light alarm linkage, for high-risk departments such as ICU, triggering bed monitor vibration + LED red light flashing (complying with EN ISO 13485 standard) simultaneously; mobile terminal push: sending a text and image message containing treatment suggestions to the responsible nurse through the hospital's internal APP;

[0105] Dynamic intervention suggestion generation includes knowledge graph reasoning, dose calculation and contraindication check;

[0106] Knowledge graph reasoning: build a treatment path network based on Neo4j graph database;

[0107] Dose calculation: if the depth needs to be adjusted, the safe range is automatically calculated (formula: ideal depth = height x 0.1 ± 5 cm);

[0108] Contraindication check: if the patient has "duodenal ulcer", the recommended advancement operation is prohibited;

[0109] The result output is to form a standardized intervention instruction set, including operation priority (high / medium / low), specific treatment steps (such as "use a 50ml empty needle to inject 20ml of normal saline to assist advancement") and time window limit (such as "must be completed within 30 minutes") information.

[0110] The embodiment realizes the automatic closed-loop management of the diagnosis and treatment process through real-time hierarchical early warning and intelligent intervention. When the confidence is lower than the threshold (such as 85%), a HIS pop-up window warning (including a FIR message type), an audible and visual alarm (meeting the EN ISO 13485 standard) and a mobile terminal push are triggered, so that the high-risk department (such as an ICU) receives the alarm within 1 second, and the response timeliness is improved; a treatment path network (such as “mistakenly entering the gastric antrum -> trying to advance” or “duodenal ulcer -> prohibiting advancement”) is constructed based on a Neo4j knowledge graph, and a standardized instruction set is generated in combination with a dose calculation formula (ideal depth = height x 0.1 + 5 cm), so that the operation priority is clear (high / medium / low) and the disposal steps are refined (such as “use 50 ml of air needle to inject physiological saline to assist in advancing”);

[0111] Potential medical errors are avoided through the contraindication checking module (such as detecting “duodenal ulcer” to disable the advancement operation); the dynamic intervention suggestion is accompanied by a time window limit (such as “must be completed within 30 minutes”), which promotes the standardization of clinical decision-making.

[0112] The application further discloses a computer readable storage medium, a computer storage medium stores a computer program, the computer program includes program instructions, the program instructions make the processor execute the method of any one of the above embodiments when the processor executes.

[0113] The application further discloses a system of a nasogastric tube positioning review method based on an abdominal X-ray film, and the system comprises:

[0114] The data acquisition and transmission module comprises an image acquisition device interface for connecting an existing X-ray machine or PACS system of a hospital, supports the DICOM standard, and automatically obtains abdominal X-ray film image data of a target patient; and comprises a HIS / EMR system integration interface, realizes seamless connection with a hospital information system through an HL7 FHIR or HL7 v2.x medical information exchange standard, and automatically extracts patient clinical information such as a medical history and a catheterization record.

[0115] The data processing and analysis module comprises a double-path feature coding model server, which is internally provided with an improved U-Net network structure and a BERT-BiLSTM-CRF hybrid model, is responsible for processing segmentation and identification of image data and analysis of report text, and outputs structured position information and confidence scores; and comprises a three-dimensional space modeling engine: a virtual human anatomy model is generated based on a CT scanning template, and a catheter path reconstruction in a three-dimensional coordinate system is converted from a two-dimensional X-ray film projection by using a computer vision algorithm.

[0116] The decision support and feedback module includes an AI intelligent review engine, integrates dynamic threshold judgment logic and an abnormal pattern library, comprehensively evaluates the catheter position, generates a positioning report and a graded early warning suggestion, includes a real-time early warning system, when detecting an abnormal catheter position, immediately reminds medical staff through a pop-up window warning of the HIS system, sound and light alarm in multiple ways, includes a knowledge graph query interface, provides a visual interface for doctors to query the ESPEN guideline recommended treatment plan, and assists clinical decision-making;

[0117] The user interaction and management module includes an operation terminal, the operation terminal includes a bedside mobile terminal and a central control room large screen, displays three-dimensional reconstruction results, early warning information and treatment suggestions, includes a permission management system, ensures that users of different levels access corresponding function areas, and protects patient privacy and data security.

[0118] In summary: the application deeply integrates image segmentation and text analysis, breaks through the limitation of a single mode, intelligently allocates the fusion ratio according to the image definition and text integrity, and takes into account the optimal solution in different scenarios. By detecting suspicious artifacts through pixel differences between front and back films, and combining a CRF decoding unit to correct semantic contradictions (such as the "coiled" keyword automatically reducing the weight).

[0119] Through an improved U-Net (6-layer encoder + hollow convolution) combined with an attention gate mechanism, the relative positions of the catheter tip and the lower edge of the diaphragm and the right edge of the spine are accurately captured, and the interference of intestinal gas artifacts is effectively suppressed; the cross-modal attention mechanism dynamically weights the image features (catheter shape) and text semantics (position description), the comprehensive confidence score is improved, and the misjudgment rate is significantly reduced.

[0120] Based on the CT reference model, a virtual human anatomy structure is constructed, the standardized mapping of bones and diaphragm markers is realized through an ICP registration algorithm, and then combined with X-ray film personalized affine transformation compensation for body shape differences, the accuracy of catheter path reconstruction is ensured; the Bezier curve fitting and curvature radius calculation can quantify the coiling risk, and assist in identifying abnormal catheter running (such as coiling in the stomach or entering the chest)

[0121] Set multiple confidence threshold values (85%, 95%), trigger graded early warning for different departments (ICU / ordinary ward), HIS system sound and light alarm and mobile terminal push treatment suggestion, shorten the clinical response time. The dynamic intervention suggestion module integrates knowledge graph reasoning (treatment path network), dose calculation (safety depth formula) and contraindication check (such as duodenal ulcer patients forbidden to use push), reduces the operation risk. Collect typical misplacement cases (such as coiling, chest misentry) and generate simulated data through GAN, construct an abnormal pattern library, and improve the adaptability of the model to complex scenes (such as intestinal gas interference).

[0122] The AI engine automatically associates the image and report information in the HIS system, replaces the traditional manual comparison process, releases the human cost of radiology and clinical medical staff, and outputs results including three-dimensional coordinates, curvature report and anatomical landmark matching degree structured data, supports the formation of a standardized intervention instruction set (including priority, treatment steps and time window limit), and promotes multi-department collaboration consistency.

[0123] The above is only a preferred embodiment of the present application, and is not used to limit the present application. The patent protection scope of the present application is subject to the claims. Any equivalent structural changes made by using the content of the specification and drawings of the present application shall also be included in the protection scope of the present application.

Claims

1. A method for nasoenteric tube localization and examination based on abdominal X-ray films, characterized in that, Includes the following steps: Step 1: Establish a nasoenteric canal localization model based on abdominal X-ray plain film image and complete the corresponding coding. The model is a dual-path feature coding model. Step 2: Obtain the abdominal X-ray images and clinical report information of the target patient through the HIS system; Step 3: Develop an AI-powered intelligent review engine to identify, analyze, and interpret the images and clinical information of the target patient's abdominal X-ray film. Based on the CT scan template, generate a virtual human anatomical model, convert the two-dimensional X-ray film projection into a three-dimensional coordinate system for catheter path reconstruction, and output the results. Set multi-level confidence thresholds to provide graded warnings for different clinical scenarios and perform dynamic threshold determination. Include typical misplacement cases, enhance the robustness of the model by generating simulated data through Generative Adversarial Network (GAN), and build an abnormal pattern library. Step 4: When an abnormal catheter position is detected, the real-time early warning mechanism immediately triggers the graded response process, and the HIS system pops up a positioning result reminder to guide clinical work.

2. The method for nasoenteric canal localization and examination based on abdominal X-ray film according to claim 1, characterized in that: The dual-path feature coding model includes an image feature extraction module, a text semantic parsing module, and a feature fusion layer; The image feature extraction module adopts an improved U-Net network structure, which expands the receptive field through dilated convolution and combines attention gating mechanism to focus on the relative positions of the duct tip with key anatomical landmarks such as the lower edge of the diaphragm and the right edge of the spine. The text semantic parsing module extracts the nasoenteric tube location description from the radiology report text in a structured manner, and generates a text vector containing location coordinates and morphological description; The feature fusion layer dynamically weights image features and text features through a cross-modal attention mechanism, and outputs a comprehensive confidence score through a gating fusion strategy.

3. The method for nasoenteric canal localization and examination based on abdominal X-ray film according to claim 2, characterized in that: The improved U-Net network structure includes a deep encoder, which expands the traditional U-Net's 4-layer encoder to 6 layers, with each layer containing 2 convolutional blocks. The deep structure extracts multi-scale features from local texture to global structure. Dilated convolution is introduced in layers 3 and 5 of the encoder to expand the receptive field through dilation rate; it is optimized by skip connections, retaining the upsampling and downsampling skip connections of U-Net, but adjusting the connection method, adding residual connections after layer 4 of the decoder, directly injecting the features of the middle layer of the encoder into the corresponding position of the decoder, and enhancing the judgment of the continuity of the curved duct. The attention gating mechanism includes spatial attention focusing, which involves inserting an attention gating module after the fourth layer of the decoder. This module receives high-level feature maps from the encoder and low-level feature maps from the decoder, and generates a gating signal through the following steps: Global average pooling is performed on the two feature maps to generate a global description vector; The interaction weights between the two are learned through a fully connected layer; After activation by the Sigmoid function, a gating coefficient of 0-1 is obtained, which is used to suppress regions unrelated to intestinal gas artifacts and highlight key landmarks at the lower edge of the diaphragm.

4. The method for nasoenteric canal localization and examination based on abdominal X-ray film according to claim 2, characterized in that: The text semantic parsing module uses the RoBERTa-wwm-ext model, freezing the underlying parameters and training only the top-level classification head to improve the recognition ability of Chinese medical terms; in addition to the standard classification task, the text generation task is trained to make the model better at inferring key information from incomplete reports. The text semantic parsing module also includes BiLSTM sequence modeling, which feeds the 768-dimensional hidden state output by BERT into a bidirectional LSTM to ensure that the anatomical description preceding "the end of the duct is located" and the functional description following it can be utilized simultaneously. The text semantic parsing module also includes a CRF (Conditional Random Field) decoding unit; it uses a linear chain CRF to define the label transition probability; it supports the simultaneous recognition of both positional and morphological information, and selects the optimal combination by maximizing the joint probability.

5. The method for nasoenteric canal localization and examination based on abdominal X-ray film according to claim 2, characterized in that: When the cross-modal attention mechanism dynamically weights image features and text features, it first performs dual-channel feature representation; in the image channel, visual features of the duct tip region are extracted from the segmentation image and a fixed-length vector is generated through region pooling; in the text channel, max pooling is performed on the text vector to retain the core semantic information of key position descriptions; dynamic weight allocation compares the relevance between image ROI and text keywords through an attention network. The gating fusion strategy includes dynamic fusion rules and confidence calculation. The dynamic fusion rules dynamically adjust the fusion ratio based on the reliability of the two modalities. The confidence score is calculated as follows: the final score is a weighted sum of the image matching score, the text semantic matching score, and the abnormal pattern matching score.

6. The method for nasoenteric canal localization and examination based on abdominal X-ray film according to claim 1, characterized in that: The virtual human anatomy model construction method described in step three is as follows: CT baseline model generation: Anonymized chest CT scan data are extracted from the hospital PACS system and converted into a three-dimensional mesh model using voxelization. Anatomical landmark registration: Key skeletal and diaphragmatic structures in CT models are mapped to the standard human coordinate system using a registration algorithm; Personalized calibration: Based on the current patient images, affine transformations are performed on the virtual model using anatomical landmarks in the X-ray to compensate for individual body shape differences.

7. The method for nasoenteric canal localization and examination based on abdominal X-ray film according to claim 1, characterized in that: The catheter route reconstruction method described in step three is as follows: Multi-view projection matching: Treating X-ray images as orthogonal projection views of a virtual model, and optimizing the coordinates of the catheter tip in three-dimensional space through bundle adjustment; Bézier curve fitting: Sample the duct pixels in the segmentation image, use cubic Bézier curve interpolation to generate a smooth path, and calculate the radius of curvature to determine the risk of coiling. Depth verification mechanism: If the position of the same pixel differs by more than 3mm in two consecutive X-ray images, it is marked as a suspicious artifact; The output includes the three-dimensional coordinates of the catheter tip, a path curvature report, and anatomical landmark matching.

8. The method for nasoenteric canal localization and examination based on abdominal X-ray film according to claim 1, characterized in that: The real-time early warning mechanism described in step four includes HIS pop-up alerts, dynamic intervention suggestion generation, and result output; HIS pop-up alerts are sent to the HIS system when the confidence level is lower than a preset threshold. The emergency alert message includes: sound and light alarm linkage, which triggers the vibration of the bedside monitor and the flashing of the LED red light in high-risk ICU departments; and mobile push: sending a text and image message with treatment suggestions to the responsible nurse through the hospital's internal APP. Dynamic intervention recommendations are generated, including knowledge graph reasoning, dosage calculation, and contraindication checks. The output is a standardized set of intervention instructions, which includes operation priorities, specific treatment steps, and time window restrictions.

9. A system for nasoenteric tube localization and examination based on abdominal X-ray film according to any one of claims 1-8, characterized in that: The system includes: Data acquisition and transmission module: includes an interface for connecting to the hospital's existing X-ray machine or PACS system to automatically acquire abdominal X-ray images of the target patient; Data processing and analysis module: includes a dual-path feature encoding model server with a built-in improved U-Net network structure and BERT-BiLSTM-CRF hybrid model; includes a 3D spatial modeling engine that generates virtual human anatomy models based on CT scan templates and uses computer vision algorithms to convert 2D X-ray projections into duct path reconstruction in a 3D coordinate system; Decision support and feedback module: including AI intelligent review engine, integrating dynamic threshold judgment logic and abnormal pattern library, comprehensively evaluates catheter position, generates positioning report and graded early warning suggestions; including real-time early warning system, which immediately alerts medical staff through HIS system pop-up warning, sound and light alarm and other means when abnormal catheter position is detected; User interaction and management module: includes operation terminals, including bedside mobile terminals and a large screen in the central control room, which display 3D reconstruction results, early warning information and handling suggestions.

10. A computer-readable storage medium, characterized in that, The computer storage medium stores a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the method as described in any one of claims 1 to 8.

Images