Multi-phase feature fusion-based intestinal obstruction etiology identification model construction method and multi-phase feature fusion-based intestinal obstruction etiology identification model construction system

Abstract

The invention discloses a method and a system for constructing an intestinal obstruction pathogenesis identification model based on multi-phase feature fusion, which are suitable for clinical diagnosis scenes with complex intestinal obstruction types and various image expressions. The method comprises the following steps: acquiring a multi-stage CT image of a patient, performing image preprocessing and structural partitioning, extracting phase-to-phase structural consistency, identifying a pseudo-difference area and a non-focus interference area, and suppressing the characteristics of the pseudo-difference area and the non-focus interference area; and meanwhile, in combination with disease cause region difference enhancement, structure expectation template comparison and dynamic tolerance analysis, intelligent discrimination of multiple disease causes such as adhesion type, tumor type, telescoping type and the like is realized. The system has the advantages of being high in model interpretability, sensitive to negative signs and high in artifact interference resistance, and the accuracy and clinical applicability of automatic identification of ileus causes are improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical care information, and specifically relates to a method and system for constructing an intestinal obstruction etiology identification model based on multi-phase feature fusion. Background Art

[0002] Intestinal obstruction is a common clinical acute abdomen with complex and diverse etiologies, including postoperative adhesions, tumors, intussusception, volvulus, mesenteric vascular disorders, inflammatory adhesions, and functional disorders. Different types of intestinal obstruction exhibit highly heterogeneous CT imaging manifestations, ranging from typical enhancing masses or occlusions to mere abnormalities in intestinal morphology and arrangement. In some cases, no lesions are evident on the images. Therefore, relying solely on imaging analysis of a single time point and feature alone is insufficient to fully reflect the true cause and mechanism of intestinal obstruction.

[0003] Currently, image-assisted diagnosis relies primarily on manual interpretation by doctors combined with observation of single-phase CT images. Although some AI-assisted diagnostic methods have been applied to intestinal disease imaging analysis in recent years, most only use single-phase images for classification and identification, lacking an understanding of the etiology and the development of comprehensive structural-functional analysis, resulting in the following major problems:

[0004] 1. The coverage of etiologies is limited, making it difficult to handle cases of “structural abnormalities but no typical lesions”;

[0005] Traditional AI models mostly classify diseases by identifying the presence of enhancing masses or occlusions in the intestine. These models can only identify "positive enhancement" etiologies, while having low accuracy for lesions without clear enhancement, such as adhesions and postoperative intussusception reduction, or even completely missing them. Furthermore, traditional systems fail to utilize the blood supply information reflected at different time points in multi-phase enhanced images, making it impossible to assess the dynamic etiological characteristics revealed by enhancement differences.

[0006] 2. Lack of negative logic modeling capabilities, unable to simulate doctors' diagnostic approach of "suspecting what they don't see." In clinical practice, doctors often infer the cause of the disease by "not finding a certain structure that should be there," such as interruption of continuous intestinal structure, loss of intestinal wall enhancement, abnormal intestinal emptying pattern, etc. This judgment based on missing features and structural disorder is an important logical basis for doctors to judge adhesive and functional intestinal obstruction. Traditional AI methods generally can only extract existing dominant features and cannot identify structural changes that "should be present but are not." Therefore, they lack effective judgment capabilities in such "negative-dominated" cases.

[0007] 3. Susceptible to interference from intestinal peristalsis, flow of contents and contrast agents, resulting in misjudgment;

[0008] As a hollow organ, the intestine undergoes significant non-pathological changes in morphology, position, brightness, and other aspects of its appearance during different phases, influenced by respiration and digestive tract peristalsis. Furthermore, the flow of contrast agents within the intestinal lumen and vascular system can cause variations in enhancement distribution. Traditional models, lacking dynamic behavior modeling and interference rejection mechanisms, can easily misinterpret these non-lesion-related differences as causal changes, leading to overfitting and generalization failure. This misinterpretation is particularly pronounced in patients with mild or postoperative disease.

[0009] 4. Unable to simulate the doctor's structural chain inference logic and ignore the "patency" judgment path;

[0010] Doctors often diagnose intestinal obstruction not solely based on the location of the cause, but rather on a continuous logical judgment based on multiple structural information, including proximal dilatation, distal collapse, and the presence of a fluid-air level in the middle. Traditional models generally use a holistic image input and unified classification output approach. This lacks segmented modeling and inference mechanisms for intestinal structural sequences, making it impossible to replicate the doctor's reasoning process for the "patency chain."

[0011] 5. The model results lack interpretability, making it difficult for doctors to trust and widely use them;

[0012] Most existing AI-assisted systems are "black box models," outputting only a disease category or probability value. They lack visualization and interactive mechanisms for the model's judgment areas and rationale. Especially in the field of acute abdomen, physicians place high demands on the traceability and interpretability of results. Without a clear explanation for why the model made a certain judgment, these models are unlikely to be adopted by clinicians or used independently for decision support.

[0013] In summary, current technologies for intestinal obstruction image recognition suffer from issues such as incomplete etiology identification, insufficient control of non-lesion interference, lack of negative logic, lack of structural chain judgment, and high model uninterpretability. Therefore, there is an urgent need for an intelligent etiology determination method and system that integrates multi-phase structural changes, possesses negative reasoning capabilities, supports dynamic tolerance identification, and provides interpretable results to better serve the auxiliary diagnosis of intestinal obstruction, a clinically complex acute abdomen. Summary of the Invention

[0014] The purpose of the present invention is: the present invention aims to provide a method and system for constructing an intestinal obstruction etiology identification model based on multi-phase feature fusion. The present invention improves the ability to identify multiple types of causes of intestinal obstruction, has the ability to judge negative signs, suppress false differences and perform dynamic tolerance analysis, and significantly enhances the accuracy, robustness and clinical interpretability of the model.

[0015] The technical solution adopted in the present invention is as follows:

[0016] The method for constructing a model for identifying the cause of intestinal obstruction based on multi-phase feature fusion includes the following steps:

[0017] S1. Acquire multiple phase CT images of the target patient, including at least plain scan, arterial, and venous phase images, and perform image preprocessing;

[0018] S2, dividing each phase image into a plurality of local image regions of fixed size, and extracting structural features of the local image regions;

[0019] S3. Calculate the structural consistency of each local image region between different phases, identify areas with structural differences but no lesions, and mark them as pseudo-difference areas;

[0020] S4, extracting the intestinal structure area and generating an intestinal mask, comparing the pseudo-difference area with the intestinal mask, and screening out the non-lesion interference area;

[0021] S5. During the deep learning model training process, the intermediate features of non-lesion interference areas are suppressed to guide the model to avoid focusing on pseudo-difference areas;

[0022] S6. Based on the feature differences between images of different phases, the structural change regions related to the etiology are extracted, and the model feature map is enhanced;

[0023] S7. Construct an expected template of the intestinal structure and compare it with the current image to see if the expected structure is present. If an "abnormal missing" situation occurs, assist in determining the cause of the negative sign.

[0024] S8. Perform dynamic tolerance analysis on the changes in the morphology, brightness, volume, and position of the intestinal cavity area in multiple phase images to identify areas of physiological dynamic changes and correct false difference judgments;

[0025] S9. Send the fused image features to the model classification module to output the cause type judgment result and positioning information of intestinal obstruction.

[0026] Preferably, the image preprocessing includes image reconstruction, size standardization, layer thickness unification and grayscale normalization operations.

[0027] Preferably, the calculation of the structural consistency is based on the grayscale distribution, edge direction, texture characteristics or stability evaluation of the structural contour of the local image region.

[0028] Preferably, the non-lesion interference area in the pseudo-difference area includes: an area with uneven distribution of contrast agent, an area with changes in intestinal contents, an image artifact area or an area with non-intestinal structure in the abdominal cavity.

[0029] Preferably, feature suppression of non-lesion interference areas is achieved by controlling the activation response intensity of the model's intermediate convolutional feature map, which is used to weaken the influence of pseudo-difference areas on model judgment.

[0030] Preferably, the extraction of the cause-related structural change region is obtained by calculating the residual between the feature maps of different phases, and constructing the cause attention map to weight the model features.

[0031] Preferably, the structure expectation template is constructed based on the arrangement direction, morphological continuity, intestinal wall enhancement distribution and periintestinal fat tissue status of the normal intestinal cavity structure.

[0032] Preferably, when the target area has no typical positive features in the multi-phase images, but shows absence or disorder when compared with the structural template, and at the same time does not belong to the physiological dynamic change area, the model can be judged as adhesion type or post-intussusception reduction type intestinal obstruction.

[0033] Preferably, the dynamic tolerance analysis is based on any one or more of the following: the degree of deformation of the intestinal contour, the range of volume change, the center position offset, and the amplitude of local brightness change.

[0034] Preferably, a system for identifying the cause of intestinal obstruction based on multi-phase feature fusion, the system being used to implement any of the above-mentioned methods for constructing a model for identifying the cause of intestinal obstruction based on multi-phase feature fusion, comprises:

[0035] An image acquisition module, used for acquiring multi-phase CT images;

[0036] Image processing module, used for image preprocessing, structure extraction and region segmentation;

[0037] Difference analysis module, used to identify pseudo-difference areas, dynamically changing areas, and abnormal missing areas;

[0038] Feature guidance module, used to suppress features in pseudo-difference regions and enhance causal regions;

[0039] The model inference module is used to fuse multi-phase image features and output the cause category and location results of intestinal obstruction.

[0040] The method and system for constructing an intestinal obstruction etiology identification model based on multi-phase feature fusion provided by this invention provide a more clinically practical, robust, and interpretable solution to the imaging challenge of acute abdomen diagnosis of intestinal obstruction, a common clinical condition with complex etiology and variable signs. Its beneficial effects are mainly reflected in the following aspects:

[0041] 1. The causes of intestinal obstruction include adhesions, tumors, intussusception, intestinal torsion, mesenteric vascular disorders, postoperative changes and other types. Different causes have significantly different manifestations on CT, and some even have no typical lesions visible. The present invention supports multi-phase comparison of etiology characteristics, and uses the enhancement differences between the arterial phase and the venous phase to judge abnormal blood supply or masses; it introduces structural template comparison and abnormality missing recognition mechanisms to identify special types such as adhesion type and intussusception reduction type that "have no clear lesions but structural disorder." Compared with the traditional method of only identifying "visible lesions", the present invention can cover more types of causes and adapt to real clinical scenarios.

[0042] 2. Clinically, the diagnosis of a significant portion of intestinal obstruction cases relies on a comprehensive reasoning approach based on the absence of enhancing lesions but an irrational structure. Traditional AI models are prone to missing these cases. This invention incorporates the logic of "should-be-present structure but not present" to construct a template for the expected intestinal arrangement, morphology, and enhancement. This approach combines the inference of "absence of positive signals + abnormal negative arrangement" in the image to identify adhesive and functional intestinal obstructions. This approach simulates the physician's negative thinking process, addressing the blind spot of traditional models that only recognize positive signals.

[0043] 3. Intestinal content migration, peristaltic behavior, and uneven contrast agent distribution can easily cause interphase structural differences, which traditional models often misclassify as lesions. This method identifies pseudo-difference regions through multi-phase structural consistency analysis; introduces a dynamic tolerance mechanism to identify non-pathological changes caused by intestinal peristalsis or gas movement; and actively suppresses the characteristic influence of such regions during training and inference. This significantly reduces the model's misclassification rate and is particularly suitable for analyzing early-stage, mild intestinal obstruction and postoperative intestinal fluctuations.

[0044] Clinicians assess intestinal obstruction not only by the cause but also by its functional consequences, such as proximal dilatation, distal collapse, and fluid-air levels. This invention utilizes structural sequence modeling to integrate the phase states of the proximal and distal intestinal lumen, simulating a "continuous structural reasoning chain" to analyze whether there is significant blockage or functional disruption. This approach, more closely resembling a physician's reading of radiographs, moves the model beyond a mere "image classifier" to a "patency logical inference tool."

[0045] 5. Ileus is an acute abdominal condition, requiring rapid, accurate, and clear diagnosis. Traditional AI models, with their black-box output, struggle to gain physician trust. The system outputs include etiology type, location segment, structural defect indication, residual plot display, and pseudo-difference suppression regions. This visual interface allows physicians to understand why the model made a specific decision and what was overlooked, providing high clinical interpretability. This facilitates rapid clinical decision-making, improving physician acceptance and practical application. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the method flow of the present invention;

[0047] Figure 2 Schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] See also Figure 1 and 2 The present invention relates to a method for constructing an intestinal obstruction etiology identification model based on multi-phase feature fusion, comprising the following steps:

[0050] S1. Acquire multiple phase CT images of the target patient, including at least plain scan, arterial, and venous phase images, and perform image preprocessing; the image preprocessing includes image reconstruction, size standardization, layer thickness unification, and grayscale normalization operations.

[0051] In clinical practice, imaging diagnosis of patients with intestinal obstruction usually relies on enhanced abdominal CT scans to exclude blood supply disorders, tumor lesions, or determine the degree of obstruction. This step is performed by acquiring images in the following ways:

[0052] Use CT equipment with multi-phase scanning capabilities (such as 64-slice or 128-slice CT);

[0053] Three-phase image acquisition was performed according to the routine clinical scanning protocol:

[0054] Plain scan phase: used to observe intestinal gas distribution, fluid level, and intestinal expansion / collapse;

[0055] Arterial phase (15 to 25 seconds after contrast agent injection): used to observe whether the intestinal wall is enhanced and determine whether blood supply is blocked;

[0056] Venous phase (60-70 seconds after injection): helps identify tumors or vascular retention, and shows indirect signs such as intestinal wall edema and lymph nodes;

[0057] The three phase images of the same patient should maintain the same body position, and the scanning range should be from below the diaphragm to the upper edge of the pubic bone, covering all small intestines and part of the colon.

[0058] After the scan is completed, the three phase images are exported uniformly into DICOM format and the phase labels are recorded.

[0059] S2, dividing each phase image into a plurality of local image regions of fixed size, and extracting structural features of the local image regions;

[0060] To more precisely identify lesion features and non-lesion changes, the following preprocessing and structural analysis were performed on the three-phase images:

[0061] Reconstruct the three phase images to make the slice thickness consistent (1 mm is recommended), unify them into a fixed size (e.g., 512 × 512), and perform grayscale normalization.

[0062] The image is divided into fixed-size image blocks (e.g., 16 × 16 pixels) to ensure that different parts of the intestine are covered;

[0063] For each image block, the following structural information is extracted:

[0064] Grayscale statistics (to determine liquid / gas distribution);

[0065] Edge contour direction (used to identify the folding and arrangement direction of the intestinal wall);

[0066] Local texture (used to determine intestinal wall thickening, mass, and intussusception);

[0067] By comparing the features of the image blocks across the three phases, the degree of structural change is analyzed and a structural consistency score is output. This step corresponds to the doctor's thought process of observing the intestinal tract across multiple phases to determine whether it is continuous, whether there is abnormal enhancement, and whether there are morphological changes.

[0068] S3. Calculate the structural consistency of each local image region across different phases, identify areas with significant structural differences but no obvious lesions, and mark them as pseudo-difference regions. The calculation of structural consistency is based on the grayscale distribution, edge orientation, texture characteristics, or stability assessment of the structural contours of the local image region. Non-lesion interference areas within the pseudo-difference region include: areas of uneven contrast agent distribution, areas of intestinal content changes, areas of image artifacts, or areas of non-intestinal abdominal cavity structures.

[0069] Based on the structural consistency score obtained in step 2, a threshold is set to screen out regions with "abnormally dramatic structural changes across multiple phases." These regions are primarily caused by: intestinal content movement; variable enhancement agent distribution; imaging errors caused by intestinal peristalsis; instrument artifacts; or localized ambiguity. These non-lesional abnormalities are labeled "pseudo-difference regions" to prevent the model from misidentifying them as causal features. A pseudo-difference mask is generated to guide the model to avoid these regions during training.

[0070] S4. Extract the intestinal structure area and generate an intestinal mask. Compare the pseudo-difference area with the intestinal mask to screen out the non-lesion interference area. Feature suppression of the non-lesion interference area is achieved by controlling the activation response intensity of the intermediate convolutional feature map of the model to weaken the influence of the pseudo-difference area on the model judgment.

[0071] To further locate the intestinal cavity and intestinal wall areas and extract the "anatomical areas that may actually produce pathogenic changes," the following steps are performed:

[0072] Identify low-density areas in the intestinal lumen based on a range set based on the Hounsfield Unit value (HU value) (e.g., -150 to +50);

[0073] Morphological processing was used to remove mislabeling of non-intestinal organs (e.g., stomach, bladder);

[0074] Combined with connected domain analysis to extract the complete intestinal path;

[0075] An intestinal structure mask is formed and overlapped with the pseudo-difference area to extract the pseudo-difference area outside the intestinal cavity, such as interference sources such as contrast agent overflow and free gas in the abdominal cavity, to form a "non-lesion interference area".

[0076] This step simulates the doctor's operation process of paying attention to "whether there are pathological signals inside and outside the intestinal cavity" and "whether there are meaningless artifacts."

[0077] S5. During the deep learning model training process, the intermediate features of non-lesion interference areas are suppressed to guide the model to avoid focusing on pseudo-difference areas;

[0078] During the deep neural network training phase, a mask of non-lesion interference regions is used to guide the model to reduce its focus on features in these regions. The specific approach is:

[0079] In the intermediate convolutional feature map, the activation intensity of the non-lesion interference area is reduced (e.g., multiplied by a scaling factor of 0.3 to 0.5);

[0080] The gradients generated in non-lesion interference areas are controlled to avoid overfitting the model to these areas.

[0081] This mechanism helps improve the model's ability to focus on the characteristics of real causes (such as obstruction points, tumors, and intussusceptions) and reduce misjudgments.

[0082] S6. Based on the feature differences between images of different phases, the etiology-related structural change regions are extracted, and the model feature map is enhanced; the etiology-related structural change regions are extracted by calculating the residuals between the feature maps of different phases, and an etiology attention map is constructed to weight the model features.

[0083] Compare the depth feature maps of the three phases and calculate the convolution feature difference (i.e. residual feature map). The residual map represents:

[0084] Whether certain structures (eg, bowel wall, tumor, mesentery) enhance during the arterial or venous phase;

[0085] Whether the intussusception area shows a "target sign" morphology;

[0086] Whether there is intestinal wall edema, changes in gas content, etc.

[0087] By using the residual map as an attention map to enhance the main model's feature map perception, the possible location of the cause of the disease is highlighted, and the model is assisted in completing accurate identification.

[0088] S7. Construct a structural expectation template of the intestinal cavity structure and compare the current image to see whether the expected structural expression exists. When "abnormal missing" occurs, assist in determining the cause of the negative sign. The structural expectation template is constructed based on the arrangement direction, morphological continuity, intestinal wall enhancement distribution and periintestinal fat tissue status of the normal intestinal cavity structure.

[0089] In clinical practice, doctors often use their experience to judge "structures that should be there but are not there" as the basis for diagnosis, for example:

[0090] Normal intestinal tract should be arranged continuously and smooth;

[0091] After intussusception is reduced, there may be no visible lesion but "intermittent imaging traces" at its location;

[0092] Adhesion-type obstruction is common in cases where there is no obvious lesion but the intestinal arrangement is abnormal.

[0093] This step establishes an expected intestinal structure template (e.g., continuity, symmetry, and intestinal wall enhancement distribution) and compares the current image with the expected template. If a structural feature that should be present is missing (e.g., a sudden collapse of a section of intestine without lesion), it is considered "abnormally absent" and can be used to identify adhesion-type and postoperative recovery-type obstruction.

[0094] S8. Perform dynamic tolerance analysis on the changes in the morphology, brightness, volume, and position of the intestinal cavity area in multiple phase images to identify areas of physiological dynamic changes and correct false difference judgments;

[0095] Considering the dynamic nature of the intestine, the following may occur in different phases: gas movement; volume change; and peristalsis leading to shape distortion.

[0096] This step establishes a reasonable range of physiological dynamic changes and evaluates: whether the position of the intestinal tube moves naturally; whether the changes in intestinal cavity volume are continuous; and whether the distribution of contents transitions naturally.

[0097] If a pseudo-difference area coincides with the "dynamic reasonable range", it is regarded as a "normal creeping change" and the area is removed from the pseudo-difference mask to avoid model misjudgment.

[0098] S9. The fused image features are fed into the model classification module, which outputs the cause and location information for the intestinal obstruction. If the target region lacks typical positive features in multiple image phases, but is missing or disorganized when compared to the structural template and does not fall within a region of physiological dynamic change, the model can identify it as an adhesion-type or post-intussusception-reduction-type intestinal obstruction.

[0099] After integrating the above structural features, difference information, and dynamic analysis, the final image representation is fed into the model’s backend classifier, which outputs the following results:

[0100] Causes of intestinal obstruction: such as adhesion type, tumor type, intussusception type, functional type;

[0101] Etiology location area: such as jejunum, ileum, ascending colon, etc.;

[0102] Judgment basis prompts: such as "no signs of enhancement, but structural loss, suspected adhesive obstruction."

[0103] The output information can be displayed through a graphical interface for radiologists or surgeons to refer to for decision-making, thereby improving diagnostic efficiency and accuracy.

[0104] The present invention also proposes a system for identifying the cause of intestinal obstruction based on multi-phase feature fusion, which is used to implement any of the above-mentioned methods for constructing a model for identifying the cause of intestinal obstruction based on multi-phase feature fusion, comprising:

[0105] The image acquisition module is used to acquire multi-phase CT images. It connects to the hospital's image storage system or CT scanning equipment to automatically import multi-phase CT images of a specified patient, including the plain scan phase, arterial phase, and venous phase. It retains the original DICOM format and phase labels of the images and provides an interface for recording acquisition information, including injection time, scan time, and scan parameters. This module is compatible with existing hospital systems and supports batch image import and historical case retrieval.

[0106] The image processing module is used for image preprocessing, structure extraction, and region segmentation. It preprocesses images from different phases, including reconstruction, resizing, layer thickness standardization, and grayscale normalization. It then divides the image into multiple fixed-size blocks and extracts structural features from each block, such as brightness distribution, edge orientation, texture pattern, and density variation. It then constructs a structural consistency matrix for identifying structurally unstable regions within the image. This module can adjust the block size and is decoupled from the model structure, ensuring universal applicability.

[0107] The difference analysis module is used to identify pseudo-difference areas, dynamically changing areas, and abnormal missing areas. Specifically, it includes:

[0108] Pseudo-difference identification unit: Score the structurally unstable areas to determine whether the differences are caused by non-pathological factors such as peristalsis, content migration, contrast agent artifacts, etc., and generate a "pseudo-difference mask".

[0109] Intestinal structure recognition unit: Utilizes HU value and connected domain recognition technology to automatically extract the intestinal area, exclude non-intestinal structures such as the stomach and bladder, generate an intestinal mask and compare it with the pseudo-difference mask to extract the "non-lesion interference area".

[0110] Dynamic change tolerance recognition unit: monitors the position, morphology, volume, and brightness changes of the intestinal area in multiple images, establishes a reasonable range of dynamic behavior, and marks physiological dynamic areas.

[0111] Structural abnormality and missing detection unit: compares the current image with the preset normal intestinal structure template to determine whether there is any structure that is not visible or is interrupted or collapsed, and marks the abnormal missing area as a basis for judging adhesion type or intussusception reduction type.

[0112] The feature guidance module suppresses features in pseudo-difference regions and enhances causal regions. It receives masks output by the image processing and difference analysis modules and suppresses features in pseudo-difference regions within the deep model to prevent misleading results. It also weights interphase residual features to guide the model's attention to causal regions, supporting joint mapping with causal focus regions for enhanced saliency. This module, without modifying the main model architecture, can be used as an auxiliary module during training or as an optional component during inference, maintaining high processing efficiency and adapting to lightweight networks.

[0113] The model inference module is used to fuse multi-phase image features and output the cause classification and location of intestinal obstruction. It receives the fused multi-phase image deep features and uses a pre-trained deep neural network (such as ResNet with attention mechanism) to complete the following tasks: intestinal obstruction cause classification, cause location location, optional negative inference suggestion output, and output standardized structured diagnosis results, including classification results, image location, and attention area.

[0114] The output and interaction module provides doctors with an image overlay display interface that displays: images of different phases, areas of high interest identified by the model, suppressed interference areas, and indications of abnormal missing areas. Results reports can be exported (PDF or embedded in an electronic medical record format); any area can be clicked to view the model's judgment logic (to aid interpretability).

[0115] Case 1: Identification of post-intussusception ileus

[0116] The patient was a male who presented with abdominal pain at the time of admission. He had a history of intussusception but no evidence of intestinal tumors. Contrast-enhanced CT scan showed no obvious enhancing lesions.

[0117] Image performance:

[0118] Plain scan: The small intestine in the mid-abdomen is slightly dilated;

[0119] Arterial phase: no enhancing mass was observed, and the intestinal wall was slightly thickened in some areas;

[0120] Venous phase: The local intestinal tract is discontinuous, the mesenteric vessels are curled, and the enhancement pattern is approximately normal.

[0121] System execution process:

[0122] S1–S2: Import the three-phase images, preprocess them and extract the structural features of the intestinal lumen area in blocks;

[0123] S3–S4: an area of structural difference is identified, but not accompanied by obvious enhancement, and the system is labeled as “abnormal absence”;

[0124] S5–S6: pseudo-difference regions are differences in contrast agent distribution, which have been suppressed by features;

[0125] S7: Structural template comparison revealed that this section of intestine should have overlapped, but currently only traces of abnormal peristalsis remain;

[0126] S8: Dynamic tolerance judgment eliminates differences caused by creep;

[0127] S9: The model output is "post-intussusception reduction type intestinal obstruction" and indicates that "the structure of the middle ileum segment is disordered, there is no tumor enhancement, and there are residual traces."

[0128] The system of the present invention can make judgments based on "missing features + structural reasoning" in the absence of positive reinforcement features.

[0129] Case 2: Adhesive intestinal obstruction

[0130] The patient is a female who had a cesarean section six months ago and had repeated abdominal distension and cessation of flatulence for the past day.

[0131] Image performance:

[0132] Plain scan: The proximal part of the small intestine is significantly dilated, and the distal part is collapsed;

[0133] Arterial phase: no enhancing tumor, no vascular occlusion;

[0134] Venous phase: There is no obvious edema of the intestinal wall and the periintestinal space is slightly narrowed.

[0135] System execution process:

[0136] S1–S2: complete three-stage image feature extraction;

[0137] S3: The model identified “no enhancement, no visible lesions”, excluding tumors;

[0138] S4–S5: Dynamic intestinal motility behaviors were identified as physiological and not misclassified;

[0139] S6–S7: Comparison of the model with the structural template revealed that there was no transitional intestinal segment after the proximal dilation segment, suggesting a “structural interruption”;

[0140] S8: Dynamic tolerance elimination and structural anomaly confirmation;

[0141] S9: Output "adhesive intestinal obstruction" and prompt "no positive signs, structural interruption, consider postoperative adhesion".

[0142] The system of the present invention supports negative etiology reasoning and solves the misclassification problem of "no visible lesions" by traditional models.

[0143] Case 3: Tumor-type intestinal obstruction with local enhancement artifact

[0144] The patient was a male who had abdominal distension for 3 days and difficulty defecating. Physical examination revealed tenderness in the mid-abdomen and no history of surgery.

[0145] Image performance:

[0146] Plain scan period: the distal jejunum is slightly dilated;

[0147] Arterial phase: uneven enhancement of the intestinal wall in the local small intestine segment;

[0148] Venous phase: delayed enhancement, mild periintestinal fatty infiltration.

[0149] System execution process:

[0150] S1–S2: import the image and divide the region;

[0151] S3: Identify an area of differential enhancement located in the intestinal lumen and present stably;

[0152] S4: The model distinguishes contrast agent artifacts from lesion residuals, and non-lesion areas are suppressed;

[0153] S5–S6: The residual map shows that the enhancement in this area persists and the activation is enhanced;

[0154] S7–S9: The structural template is intact and consistent with typical tumor manifestations; output "tumor-type intestinal obstruction" and mark "persistent enhancement changes + asymmetric collapse of the intestinal lumen."

[0155] The system of the present invention can distinguish between pseudo differences and real lesions and avoid false suppression of real enhanced areas.

[0156] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing a model for identifying the cause of intestinal obstruction based on multi-phase feature fusion, characterized in that: The following steps are involved: S1. Acquire multiple phase CT images of the target patient, including at least plain scan, arterial, and venous phase images, and perform image preprocessing; S2, dividing each phase image into a plurality of local image regions of fixed size, and extracting structural features of the local image regions; S3. Calculate the structural consistency of each local image region between different phases, identify areas with structural differences but no lesions, and mark them as pseudo-difference areas; S4, extracting the intestinal structure area and generating an intestinal mask, comparing the pseudo-difference area with the intestinal mask, and screening out the non-lesion interference area; S5. During the deep learning model training process, the intermediate features of non-lesion interference areas are suppressed to guide the model to avoid focusing on pseudo-difference areas; S6. Based on the feature differences between images of different phases, the structural change regions related to the etiology are extracted, and the model feature map is enhanced; S7. Construct an expected template of the intestinal structure and compare it with the current image to see if the expected structure is present. If an abnormality is missing, assist in determining the cause of the negative sign. S8. Perform dynamic tolerance analysis on the changes in the morphology, brightness, volume, and position of the intestinal cavity area in multiple phase images to identify areas of physiological dynamic changes and correct false difference judgments; S9. Send the fused image features to the model classification module to output the cause type judgment result and positioning information of intestinal obstruction.

2. The method for constructing a model for identifying the cause of intestinal obstruction based on multi-phase feature fusion according to claim 1, characterized in that: The image preprocessing includes image reconstruction, size standardization, layer thickness unification and grayscale normalization operations.

3. The method for constructing an intestinal obstruction etiology identification model based on multi-phase feature fusion according to claim 1, characterized in that: The calculation of the structural consistency is based on the stability evaluation of the grayscale distribution, edge direction, texture features or structural contour of the local image region.

4. The method for constructing a model for identifying the cause of intestinal obstruction based on multi-phase feature fusion according to claim 1, characterized in that: Non-lesion interference areas in the pseudo-difference area include: uneven distribution of contrast agent, changes in intestinal content, image artifacts or non-intestinal structure areas in the abdominal cavity.

5. The method for constructing an intestinal obstruction etiology identification model based on multi-phase feature fusion according to claim 1, characterized in that: Feature suppression of non-lesion interference areas is achieved by controlling the activation response intensity of the model's intermediate convolutional feature map, which is used to weaken the influence of pseudo-difference areas on model judgment.

6. The method for constructing an intestinal obstruction etiology identification model based on multi-phase feature fusion according to claim 1, characterized in that: The extraction of etiology-related structural change regions is obtained by calculating the residuals between feature maps of different phases, and a etiology attention map is constructed to weight the model features.

7. The method for constructing an intestinal obstruction etiology identification model based on multi-phase feature fusion according to claim 1, characterized in that: The structure expectation template is constructed based on the arrangement direction, morphological continuity, intestinal wall enhancement distribution and periintestinal fat tissue status of the normal intestinal cavity structure.

8. The method for constructing an intestinal obstruction etiology identification model based on multi-phase feature fusion according to claim 1, characterized in that: When the target area does not show typical positive features in multi-phase images, but shows absence or disorder compared with the structural template, and at the same time does not belong to the physiological dynamic change area, the model judges it as adhesion-type or post-intussusception reduction-type intestinal obstruction.

9. The method for constructing an intestinal obstruction etiology identification model based on multi-phase feature fusion according to claim 1, characterized in that: The basis for dynamic tolerance analysis includes any one or more of the following: the degree of deformation of the intestinal contour, the range of volume change, the center position offset, and the amplitude of local brightness change.

10. The intestinal obstruction etiology identification system based on multi-phase feature fusion is characterized by: The system is used to implement the method for constructing an intestinal obstruction etiology identification model based on multi-phase feature fusion as described in any one of claims 1 to 9, comprising: An image acquisition module, used for acquiring multi-phase CT images; Image processing module, used for image preprocessing, structure extraction and region segmentation; Difference analysis module, used to identify pseudo-difference areas, dynamically changing areas, and abnormal missing areas; Feature guidance module, used to suppress features in pseudo-difference regions and enhance causal regions; The model inference module is used to fuse multi-phase image features and output the cause category and location results of intestinal obstruction.

Images