A method for non-invasive dynamic monitoring of intestinal mucosal ulcer

Abstract

The application discloses a kind of methods for intestinal mucosa ulcer noninvasive dynamic monitoring, belong to medical image processing technical field.The method includes: training intestinal segment segmentation model and intestinal lesion segmentation model, ulcer classification model, intestinal visualization model;Abdominal pelvic CT / MR image is obtained, and clinical interested intestinal segment and lesion area are automatically segmented using intestinal segment segmentation model and lesion segmentation model, corresponding mask is generated;After mask and abdominal pelvic CT / MR image are fused, input ulcer classification model, output the probability value P of intestinal segment existing ulcer, compare with preset threshold T to determine whether there is ulcer, and generate classification probability heat map H;The point of maximum probability value is positioned ulcer on classification probability heat map H;According to the positioning result, shooting visual angle and focal point are determined, and virtual endoscope image showing ulcer area is generated through visualization model.The application realizes the noninvasive automatic diagnosis positioning and visualization of ulcer, and can be further compared by multi-phase image to realize dynamic monitoring.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a method for non-invasive dynamic monitoring of intestinal mucosal ulcers. Background Technology

[0002] In clinical practice, the presence of intestinal mucosal ulcers is a crucial dividing line for diagnostic and treatment decisions. The appearance of ulcers usually indicates that the lesion has broken through the mucosal layer and involved the submucosa, marking the progression of the disease from "functional or mild inflammation" to "organic damage," thus requiring more aggressive and precise intervention strategies. Simultaneously, continuous monitoring of the ulcers is necessary during treatment to dynamically assess their progress or healing, shifting the focus of assessment from symptom improvement to objective healing indicators observed endoscopically.

[0003] Currently, the assessment of intestinal mucosal ulcers usually relies on colonoscopy, but this method is invasive and requires insufflation and anesthesia, making it difficult to perform on patients with cardiopulmonary failure, intestinal obstruction, or intestinal perforation. With the development of medical imaging technologies such as CT (computed tomography) and MRI (magnetic resonance imaging), new possibilities have been provided for non-invasive assessment of intestinal diseases.

[0004] Existing literature research indicates that CT / MR-based diagnostic methods for intestinal diseases mainly include the following categories:

[0005] One method is based on manual assessment, in which radiologists visually assess whether there are ulcers in the intestines. This method relies entirely on manual interpretation, which is time-consuming, labor-intensive, and greatly affected by subjective factors, resulting in limited consistency among different physicians.

[0006] The second method is based on region classification, which involves manually delineating the intestinal segment of interest on CT images and then inputting it into a classification model for intestinal disease classification. Although this type of method incorporates machine learning models, it still relies on manually delineating the intestinal segment of interest, lacks automated intestinal segment extraction capabilities, and its model focuses on the diagnosis of Crohn's disease and inflammation, without addressing ulcer identification. Whether the model can learn cross-modal associations between CT / MR images and mucosal ulcers remains unclear.

[0007] Thirdly, visualization-based methods utilize CT / MR data to reconstruct virtual endoscopic images to visually display the internal structure of the intestine. However, these methods only provide random visualizations of the intestine and cannot automatically locate the specific position of the ulcer for targeted visualization. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method for non-invasive dynamic monitoring of intestinal mucosal ulcers, enabling automatic intestinal segment extraction, ulcer diagnosis, precise localization, and visualization.

[0009] To achieve the above-mentioned technical objectives, the present invention is implemented through the following technical solution:

[0010] A method for non-invasive dynamic monitoring of intestinal mucosal ulcers includes the following steps:

[0011] S100, the intestinal segmentation model and the intestinal lesion segmentation model, are trained on a training set containing intestinal segment mask annotations and lesion mask annotations. Two independent deep learning segmentation networks are trained to segment intestinal segments of clinical interest and intestinal lesion regions, respectively.

[0012] S200. Train an ulcer classification model by collecting a dataset of CT / MR images with known labels, including intestinal segment masks, lesion area masks, and corresponding endoscopic diagnostic results. Train a deep learning classification network based on the dataset. The classification network outputs a probability value P of the presence of an ulcer in the intestinal segment, where P ranges from [0,1]. The ulcer classification model can generate a classification probability heatmap H based on the class activation mapping method, where H(x,y,z) represents the probability response value at coordinates (x,y,z) in the heatmap.

[0013] S300. Train an intestinal visualization model and generate virtual endoscopic images based on CT / MR voxel data. The visualization model is trained using predetermined tissue boundaries in CT / MR images as control conditions and real endoscopic images of the corresponding parts as supervision signals.

[0014] S400. Acquire CT / MR images of the abdominal and pelvic region, and automatically segment at least one intestinal segment of clinical interest using the trained intestinal segmentation model to generate an intestinal segment mask; automatically segment the lesion region in the intestinal segment using the trained lesion segmentation model to generate a lesion mask.

[0015] S500: The intestinal segment mask and lesion mask are stitched and fused with the abdominal and pelvic CT / MR images in the channel dimension to form a multi-channel input tensor, which is input to the ulcer classification model trained in S200. The model outputs a probability value P of the presence of an ulcer in the intestinal segment. The probability value P is compared with a preset threshold T. If P ≥ T, it is determined that an ulcer exists; if P < T, it is determined that an ulcer does not exist. At the same time, the ulcer classification model also generates a classification probability heatmap H corresponding to the classification decision.

[0016] S600. On the classification probability heatmap H, take the point with the highest probability value as the target location L:

[0017] ;

[0018] in, This represents the spatial domain of the classification probability heatmap H. The voxel coordinates in the heatmap are represented by arg max, which represents the coordinates that maximize H(x,y,z).

[0019] S700: When S500 determines that an ulcer exists, the shooting angle and focus are determined according to the target position L obtained in S600. The voxel data of the abdominal and pelvic CT / MR image is input into the visualization model trained in S300. The predetermined tissue boundary in the abdominal and pelvic CT / MR image is used as the control condition. A virtual endoscopic image showing the ulcer area is generated through three-dimensional voxel rendering. When S500 determines that no ulcer exists, a position is randomly selected on the center line of the intestinal segment for conventional virtual endoscopic roaming display.

[0020] The S100 intestinal segmentation model is used to segment at least one intestinal segment from the terminal ileum, ileocecal junction, ascending colon, transverse colon, descending colon, sigmoid colon, and rectum in the abdominal and pelvic CT / MR image; the lesion segmentation model is used to segment the lesion region from the abdominal and pelvic CT / MR image or the segmented intestinal segment.

[0021] The execution order of intestinal segmentation and lesion segmentation in S100 can be interchanged; when lesion segmentation is performed first, the lesion area is directly segmented on the abdominal and pelvic CT / MR image; when intestinal segmentation is performed first, the lesion area is segmented within the already segmented intestinal segment.

[0022] The ulcer classification model in S200 generates a classification probability heatmap H based on the class activation mapping method.

[0023] In step S500, the preset threshold T is set to 0.5 by default and can be adjusted according to clinical screening needs; when the threshold T decreases, the sensitivity increases. Elevated and specific Decrease; as the threshold T increases, sensitivity decreases. Reduced and specific Increase;

[0024] Wherein, the sensitivity and specificity Defined according to the following formulas respectively:

[0025] ;

[0026] ;

[0027] In the formula, The number of correctly predicted ulcer samples, This represents the number of ulcer samples that were missed during diagnosis. The number of normal samples that are correctly predicted. This represents the number of normal samples that were misdiagnosed.

[0028] The method also includes a dynamic monitoring phase:

[0029] S800: Acquire CT / MR images of the subject's abdominal and pelvic region at at least two different time points, and execute S400 to S700 respectively to compare the lesion mask, the probability value P, and the target location L at different time points to generate a dynamic monitoring report reflecting the progression or healing of the lesion.

[0030] S900. Compare the ulcer classification results, classification probability values ​​P, and classification probability heatmaps H at different time points;

[0031] S1000. Compare the target ulcer location L at different time points to determine whether lesion metastasis has occurred.

[0032] The comparison of lesion masks in S800 includes a quantitative comparison of the number of lesions, lesion volume, and / or lesion topology.

[0033] The intestinal segment of clinical interest and the lesion are located in the digestive system, including the small intestine, large intestine and stomach; the method can be used for monitoring intestinal ulcers, intestinal erosions or intestinal inflammation.

[0034] The beneficial effects of this invention are:

[0035] Compared with existing technologies, this invention introduces a cascaded architecture of deep learning segmentation and classification models. This establishes a cross-modal association between lesions and mucosal ulcers on CT / MR images, eliminating reliance on manual intestinal segment delineation and experience-based interpretation, significantly improving diagnostic efficiency and repeatability. The introduction of classification probability heatmaps not only enhances the interpretability of the deep learning model's decision-making process but also provides an effective technical approach for accurately locating the core region of ulcers, overcoming the shortcomings of existing visualization methods that cannot automatically locate ulcers and can only display them randomly. Based on this, virtual endoscopy rendering technology based on precise location can transform abstract tomographic image data into intuitive images of intraluminal anatomical structures, effectively assisting doctors in visual confirmation and comprehensive judgment. Simultaneously, the dynamic monitoring function of this invention supports automated longitudinal comparative analysis of multi-temporal images, capturing changes in the number, volume, and spatial location of lesions, providing clinicians with objective and quantitative imaging evidence for evaluating treatment effectiveness. The verification results show that the present invention has better sensitivity and specificity than traditional manual visual assessment methods in ulcer classification tasks, and can be flexibly adapted to the needs of different clinical application scenarios by adjusting the classification threshold, thus having good clinical promotion value. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments are briefly introduced below.

[0037] Figure 1 This is a flowchart illustrating the overall process of a method for non-invasive dynamic monitoring of intestinal mucosal ulcers based on CT / MR images, as provided in an embodiment of the present invention.

[0038] Figure 2 This is a schematic diagram of the effect of S100 intestinal segmentation and lesion segmentation in an embodiment of the present invention, where the left side is the ground truth (GT) and the right side is the model prediction result.

[0039] Figure 3 This is a schematic diagram of S500 ulcer classification and classification probability heatmap generation in an embodiment of the present invention. The left column is the original CT image, and the right column is the corresponding classification probability heatmap. The red dot indicates the position with the highest probability in the heatmap.

[0040] Figure 4 The diagrams for S600 and S700 of this invention, which are based on classification probability heatmaps for ulcer localization and virtual endoscopy visualization, respectively illustrate the diagnostic process for healthy intestinal segments and the diagnostic process for ulcerated intestinal segments.

[0041] Figure 5 This is a flowchart illustrating the dynamic monitoring steps S800 to S1000 in an embodiment of the present invention, showing the logical relationship between the three sub-steps: lesion segmentation comparison (S800), ulcer classification comparison (S900), and ulcer location comparison (S1000). Detailed Implementation

[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0043] Example 1

[0044] This embodiment uses non-invasive monitoring of ulcers in the colon, rectum, and terminal ileum based on abdominal and pelvic CTE (CT Enterography) images as an example for detailed explanation. However, those skilled in the art should understand that the present invention is also applicable to the diagnosis of diseases in other parts of the digestive system, such as the small intestine and stomach, as well as the monitoring of similar diseases such as intestinal inflammation and intestinal erosion, and is also applicable to MR images.

[0045] like Figure 1 As shown, the method in this embodiment includes a model training phase (S100 to S300) and an inference application phase (S400 to S700).

[0046] S100: Training intestinal segmentation model and intestinal lesion segmentation model

[0047] This step includes training a multi-class intestinal segmentation model and an intestinal lesion segmentation model.

[0048] Two independent deep learning segmentation networks are trained using a training set containing intestinal segment mask annotations and lesion mask annotations. In this embodiment, the U-Net or nnUNet architecture is preferred.

[0049] The first segmentation network (intestinal segmentation model) was used to segment seven clinically relevant intestinal segments from CT images: terminal ileum, ileocecal junction, ascending colon, transverse colon, descending colon, sigmoid colon, and rectum. This model is a multi-class semantic segmentation model.

[0050] The second segmentation network (lesion segmentation model) is used to segment intestinal lesion regions from CT images. This model is a binary semantic segmentation model.

[0051] The performance of the segmentation model is evaluated using the Dice similarity coefficient:

[0052] ;

[0053] in, The segmentation region predicted by the model. The segmented regions are the actual labeled regions. This represents the number of voxels at the intersection of the predicted region and the ground truth labeled region. and These represent the total number of voxels in the predicted region and the ground truth labeled region, respectively. The value of DSC ranges from [0,1], with a value closer to 1 indicating a better match between the segmentation result and the ground truth label. After training, the model parameters of the two segmentation networks are fixed for use in the subsequent inference stage S400.

[0054] S200: Training an ulcer classification model

[0055] This step is used to build a deep learning classifier that can distinguish whether an intestinal segment has ulcers.

[0056] A large dataset of CTE images with known labels was collected. Specifically, a dataset containing CT images of intestinal segments was collected, with labels including: intestinal segment masks, lesion area masks, and corresponding endoscopic diagnostic results. Based on this data, a 3D deep learning classification network, such as 3D-ResNet or 3D-DenseNet, was trained. The goal of the training was to enable the network to learn a nonlinear mapping from the voxel feature space of the CT images to the ulcer label space.

[0057] The classification model outputs a probability value P, which ranges from [0,1] and represents the confidence level that the input intestinal segment has an ulcer. The closer P is to 1, the higher the model believes that the intestinal segment is likely to have an ulcer, and the closer P is to 0, the higher the model believes that the intestinal segment is normal.

[0058] Simultaneously, the ulcer classification model can generate a classification probability heatmap H. H(x,y,z) represents the probability response value at coordinates (x,y,z) in the heatmap, which reflects the contribution of the corresponding voxel to the classification decision. In this embodiment, the class activation mapping method is preferably used to generate the classification probability heatmap.

[0059] S300: Training a Gut Visualization Model

[0060] This step is used to build a generative model that can generate virtual endoscopic images based on CT voxel data.

[0061] The model employs a conditional generative adversarial network or diffusion model architecture, using pre-extracted tissue boundaries (such as intestinal wall edges) from CT images as control conditions and real endoscopic images of the corresponding sites as supervisory signals for training. After training, the generator part of the model is fixed for subsequent visualization steps S700.

[0062] S400: Fully Automated Intestinal Segment Extraction and Lesion Division

[0063] After obtaining new CT images of the abdominal and pelvic region to be tested, they are input into the pre-trained intestinal segmentation model in S100. The model automatically segments the intestine into seven segments: terminal ileum, ileocecal junction, ascending colon, transverse colon, descending colon, sigmoid colon, and rectum, generating binary masks (intestinal segment masks) for each segment. For example... Figure 2 The two images above, taking case crohnCTE10029_6_SE000006 as an example, show the ground truth (GT) annotations for intestinal segmentation on the left and the model prediction results for intestinal segmentation on the right. The model can accurately segment each intestinal segment.

[0064] Simultaneously, the CT image is input into the lesion segmentation model trained in S100, which automatically segments the lesion region and generates a lesion mask. For example... Figure 2 The following two images show the case crohnCTE30054_058924 as an example. The left side shows the ground truth (GT) of the lesion segmentation, and the right side shows the model prediction result of the lesion segmentation. The model can accurately segment the lesion area.

[0065] In this step, the execution order of intestinal segment segmentation and lesion segmentation can be reversed. If lesion segmentation is performed first, the lesion region is segmented directly on the original abdominal and pelvic CT image; if intestinal segment segmentation is performed first, the lesion region is segmented within the range of the already segmented intestinal segment. Finally, both the intestinal segment mask and the lesion mask need to be obtained for use in subsequent steps.

[0066] S500: Ulcer classification and heat map generation

[0067] The original CT image, the intestinal segment mask and the lesion mask generated in S400 are fused, and input into the ulcer classification model trained in S200.

[0068] The model outputs the probability value P that an ulcer exists in this intestinal segment. The probability value P is compared with the preset threshold T, and the judgment criteria are as follows:

[0069] P≥T, it is determined that an ulcer exists; P<T, it is determined that no ulcer exists;

[0070] In this embodiment, the threshold T is set to 0.5 by default, that is, when P≥0.5, it is determined that there is an ulcer, and when P < 0.5, it is determined that there is no ulcer. It should be noted that the value of the threshold T can be adjusted according to the clinical application scenario: in the screening scenario, the threshold T can be lowered to improve sensitivity and reduce the missed diagnosis rate; in the diagnosis confirmation scenario, the threshold T can be raised to improve specificity and reduce the misdiagnosis rate. The adjustment strategy of the threshold T is detailed in the effect verification part.

[0071] At the same time, the model generates a classification probability heat map H. As shown in Figure 3 , taking the case crohnCTE20396_1290953_20190426_5_3 as an example, Figure 3 the left side is the original CT image, and the right side is the corresponding classification probability heat map. The red dot indicates the position with the maximum probability in the heat map, which has a good spatial correspondence with the actual occurrence position of the ulcer. The heat map H has the same spatial dimensions as the original CT image, and the value of H(x,y,z) indicates the contribution degree of the voxel at the coordinate (x,y,z) to the final classification decision.

[0072] S600: Ulcer localization

[0073] On the classification probability heat map H generated in S500, the point with the maximum probability value is taken as the target position L, that is, the target position L is determined according to the following formula:

[0074] ;

[0075] wherein, represents the spatial domain of the classification probability heat map H, The vector represents the voxel coordinates in the heatmap, and arg max represents the coordinates that maximize H(x,y,z). In other words, the voxel coordinates with the strongest response in the heatmap are determined as the most likely location of the ulcer.

[0076] Subsequently, a path point closest to the target position L is selected on the pre-calculated centerline of the intestinal segment as the origin point, and an optimal shooting angle is determined based on the direction vector from the origin point to the target position L.

[0077] S700: Visual Display

[0078] like Figure 4 As shown, branching is performed based on the classification results of S500.

[0079] Figure 4 The above shows the diagnostic process for a healthy intestinal segment: input abdominal and pelvic CT images → extract intestinal mask and intestinal segment lesion mask (no lesions detected) → classification model determines it as normal → randomly select intestinal segment locations for visualization display, resulting in a visualized image of a healthy intestine. Figure 4 The following shows the diagnostic process for ulcerated intestinal segments: Input abdominal and pelvic CT images → Extract intestinal masks and intestinal segment lesion masks → Classification model determines it as an ulcer → Locate the ulcer location → Visualize the location results to obtain a visualized image of the ulcerated intestine.

[0080] Specifically, if an ulcer is determined to exist (P ≥ T), the original CT voxel data, combined with the target location L, shooting angle, and focus information obtained by S600, is input into the visualization model trained by S300. Using the pre-extracted intestinal wall edge from the CT image as a control condition, a virtual endoscopic image focused at the target location L is generated through 3D voxel rendering. This virtual endoscopic image can intuitively display the morphology, size, and extent of the ulcer on the intestinal wall.

[0081] If it is determined that there is no ulcer (P < T), a location is randomly selected on the center line of the intestinal segment for a routine virtual endoscopic tour.

[0082] Effect verification

[0083] This embodiment was validated on an independent test set.

[0084] (a) Effect of intestinal segmentation

[0085] In the intestinal segmentation task, 160 sets of CTE data were used for training, and 40 sets of CTE data were used for testing. The quantitative and visualization results of intestinal segmentation are as follows: Figure 2The two images above, using case crohnCTE10029_6_SE000006 as an example, show the ground truth (GT) on the left and the model prediction results on the right. The model can automatically segment seven intestinal segments of clinical interest. The average Dice similarity coefficient for the segmented intestinal segments reaches 0.93.

[0086] (II) Effect of intestinal lesion segmentation

[0087] In the intestinal lesion segmentation task, 61 sets of CTE data were used for training, and 12 sets of CTE data were used for testing. The quantitative and visualization results of lesion segmentation are as follows: Figure 2 The two images below, using case crohnCTE30054_058924 as an example, show the ground truth (GT) on the left and the model prediction results on the right. The model can automatically segment the intestinal lesion area. The average Dice similarity coefficient of the lesion segmentation reaches 0.82.

[0088] (III) Effect of ulcer classification

[0089] In the ulcer classification task, five-fold cross-validation was performed based on CTE image data from 2588 patients (a total of 16583 intestinal segments). Sensitivity and specificity were used as evaluation metrics.

[0090] Sensitivity (Se), or true positive rate, reflects the model's ability to correctly detect ulcers, and is defined as:

[0091] ;

[0092] Where TP (True Positive) represents the number of correctly predicted ulcer samples, and FN (False Negative) represents the number of missed ulcer samples. The higher the sensitivity, the lower the missed diagnosis rate.

[0093] Specificity (Sp), or true negative rate, reflects the model's ability to correctly exclude non-ulcer cases, and is defined as:

[0094] ;

[0095] Here, TN (True Negative) represents the number of correctly predicted normal samples, and FP (False Positive) represents the number of misdiagnosed normal samples. The higher the specificity, the lower the misdiagnosis rate.

[0096] Experimental results show that this invention, in classifying the presence or absence of intestinal ulcers at the patient level, achieved an average sensitivity of 75.4% and an average specificity of 75.5% with five-fold cross-validation. The classification heatmap visualization results are as follows: Figure 3As shown, taking case crohnCTE20396_1290953_20190426_5_3 as an example, the red dot indicates the location with the highest probability in the heat map, which basically matches the actual location of the ulcer under endoscopy.

[0097] In contrast, the traditional manual visual assessment method reported in the existing literature, which relied on the majority consensus of three radiologists to visually assess the presence or absence of intestinal ulcers on MRE images of 77 patients, had the following results: sensitivity 66.7% and specificity 69.2%.

[0098] A comparison of the two methods shows that the present invention is superior to the traditional manual visual evaluation method in terms of both sensitivity and specificity, improving by approximately 8.7 percentage points and 6.3 percentage points, respectively.

[0099] Furthermore, in clinical applications of intestinal ulcer screening, it is usually necessary to prioritize improving sensitivity to avoid missed diagnoses of ulcers. Flexible control of model performance can be achieved by adjusting the preset threshold T. The principle is as follows: when the threshold T is lowered, the criteria for determining a positive result become more lenient, more samples are predicted as positive, TP increases, FN decreases, and according to formula 4, sensitivity Se increases; however, FP also increases accordingly, and according to formula 5, specificity Sp decreases. Conversely, raising the threshold T decreases sensitivity and increases specificity.

[0100] In the five-fold cross-validation of this embodiment, when the average sensitivity is controlled above 85.5% by lowering the threshold T, the average specificity of the model is 59.2%. This indicates that the classification model of the present invention has good adjustability and can adapt to different clinical application needs by adjusting the threshold T.

[0101] Example 2

[0102] Based on Example 1, this embodiment further provides a dynamic monitoring function to non-invasively and continuously monitor the condition of intestinal lesions and the occurrence of intestinal ulcers, thereby assisting doctors in evaluating clinical efficacy.

[0103] like Figure 5 As shown, abdominal and pelvic CT images of the same patient at two different times (e.g., time t1 before treatment and time t2 after treatment) are acquired. The images at these two times are then input into steps S400 to S700 of Example 1 to obtain the lesion segmentation results, ulcer classification results (including classification probability value P and classification probability heatmap H) and ulcer target location L at time t1 and t2.

[0104] S800: Lesion Segmentation Comparison

[0105] Compare the lesion masks at time points t1 and t2. The comparison methods include, but are not limited to:

[0106] (1) Comparison of the number of lesions: Statistically compare the number of connected components of the lesions at two time points. The number of connected components can be obtained through a connected component labeling algorithm.

[0107] (2) Lesion volume comparison: The lesion volume is calculated based on the number of voxels in the lesion mask and the spatial resolution of the CT image. The lesion volume V is calculated according to the following formula:

[0108] ;

[0109] Where N is the number of voxels contained in the lesion mask (i.e., the number of voxels with a value of 1 in the mask), and v0 is the physical volume corresponding to each voxel. v0 is determined by the spatial resolution of the CT / MR image.

[0110] ;

[0111] Δx, Δy, and Δz represent the voxel spacing in the x, y, and z directions of the image, respectively, and can be directly read from the header information of the DICOM (Digital Imaging and Communications in Medicine) file. The lesion volume V is measured in cubic millimeters (mm³) or cubic centimeters (cm³).

[0112] By comparing the lesion volume at times t1 and t2, the treatment effect can be quantitatively assessed: a decrease in volume indicates that the lesion has improved or healed, while an increase in volume indicates that the lesion has progressed or worsened.

[0113] (3) Comparison of lesion topology: Analyze the changes in shape, connectivity and distribution pattern of the lesion area, such as whether the lesion area is split from one connected domain into multiple connected domains, or whether multiple connected domains are merged into one connected domain.

[0114] At the same time, the lesion segmentation masks at two different time points can be overlaid to visually observe the changes in the lesion area through visual evaluation.

[0115] S900: Comparison of Ulcer Classifications

[0116] Compare the changes in the location distribution and response intensity of the ulcer classification results, classification probability values ​​P, and classification probability heatmap H at times t1 and t2.

[0117] For example, if the classification probability value P decreases from 0.9 at time t1 to 0.2 at time t2, it indicates that the ulcer is showing a clear healing trend; conversely, if P increases from 0.3 to 0.8, it suggests that a new ulcer may have developed or that an existing lesion may have worsened. The change in the response area of ​​the classification probability heatmap H from a large-scale concentration to a small-scale dispersion also indicates the regression of the lesion.

[0118] S1000: Comparison of ulcer locations.

[0119] By comparing the ulcer target location L located by the heat map at time t1 and t2, it can be determined whether the lesion location has shifted, disappeared, or been newly added.

[0120] The spatial distance ΔL between the ulcer locations at two different times is calculated using the following formula:

[0121] ;

[0122] Here, (x1, y1, z1) and (x2, y2, z2) represent the target locations of the ulcer at times t1 and t2, respectively. The value of ΔL reflects the degree of migration of the ulcer core location in three-dimensional space: ΔL = 0 indicates that the ulcer location has not moved, and the larger the ΔL, the more significant the location shift. If no ulcer is detected at time t2 (classified as normal), it indicates that the ulcer has healed.

[0123] Through comparative analysis of S800, S900, and S1000, the system can automatically generate a structured dynamic monitoring report, which quantitatively reflects key information such as changes in lesion volume (ΔV = V2 − V1), evolution of ulcer classification probability (ΔP = P2 − P1), and ulcer location migration (ΔL), providing objective and quantitative imaging evidence for clinicians to adjust medication regimens and assess treatment effectiveness.

[0124] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not describe all details exhaustively, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification.

Claims

1. A method for non-invasive dynamic monitoring of intestinal mucosal ulcers, characterized in that, Includes the following steps: S100, Training intestinal segmentation model and intestinal lesion segmentation model: Based on a training set containing intestinal segment mask annotations and lesion mask annotations, train two independent deep learning segmentation networks, which are used to segment intestinal segments of clinical interest and intestinal lesion regions, respectively. S200. Train an ulcer classification model by collecting a dataset of CT / MR images with known labels, including intestinal segment masks, lesion area masks, and corresponding endoscopic diagnostic results. Train a deep learning classification network based on the dataset. The classification network outputs a probability value P of the presence of an ulcer in the intestinal segment, where P ranges from [0,1]. The ulcer classification model can generate a classification probability heatmap H based on the class activation mapping method, where H(x,y,z) represents the probability response value at coordinates (x,y,z) in the heatmap. S300. Train an intestinal visualization model and generate virtual endoscopic images based on CT / MR voxel data. The visualization model is trained using predetermined tissue boundaries in CT / MR images as control conditions and real endoscopic images of the corresponding parts as supervision signals. S400. Acquire CT / MR images of the abdominal and pelvic region, and automatically segment at least one intestinal segment of clinical interest using the trained intestinal segmentation model to generate an intestinal segment mask; automatically segment the lesion region in the intestinal segment using the trained lesion segmentation model to generate a lesion mask. S500: The intestinal segment mask and lesion mask are stitched and fused with the abdominal and pelvic CT / MR images in the channel dimension to form a multi-channel input tensor, which is then input into the ulcer classification model trained in S200. The model outputs the probability value P of the presence of an ulcer in the intestinal segment. The probability value P is compared with a preset threshold T. If P ≥ T, it is determined that an ulcer exists; if P < T, it is determined that an ulcer does not exist. At the same time, the ulcer classification model also generates a classification probability heatmap H corresponding to the classification decision. S600. On the classification probability heatmap H, take the point with the highest probability value as the target location L: ； in, This represents the spatial domain of the classification probability heatmap H. The voxel coordinates in the heatmap are represented by argmax, which represents the coordinates that maximize H(x,y,z). S700: When S500 determines that an ulcer exists, the shooting angle and focus are determined according to the target position L obtained in S600. The voxel data of the abdominal and pelvic CT / MR image is input into the visualization model trained in S300. The predetermined tissue boundary in the abdominal and pelvic CT / MR image is used as the control condition. A virtual endoscopic image showing the ulcer area is generated through three-dimensional voxel rendering. When S500 determines that no ulcer exists, a position is randomly selected on the center line of the intestinal segment for conventional virtual endoscopic roaming display.

2. The method for non-invasive dynamic monitoring of intestinal mucosal ulcers according to claim 1, characterized in that, The S100 intestinal segmentation model is used to segment at least one intestinal segment from the terminal ileum, ileocecal junction, ascending colon, transverse colon, descending colon, sigmoid colon, and rectum in the abdominal and pelvic CT / MR image; the lesion segmentation model is used to segment the lesion region from the abdominal and pelvic CT / MR image or the segmented intestinal segment.

3. The method for non-invasive dynamic monitoring of intestinal mucosal ulcers according to claim 2, characterized in that, The execution order of intestinal segmentation and lesion segmentation in S100 can be interchanged; when lesion segmentation is performed first, the lesion area is directly segmented on the abdominal and pelvic CT / MR image. When the intestinal segmentation is performed first, the lesion area is segmented within the already segmented intestinal segment.

4. The method for non-invasive dynamic monitoring of intestinal mucosal ulcers according to claim 1, characterized in that, The ulcer classification model in S200 generates a classification probability heatmap H based on the class activation mapping method.

5. The method for non-invasive dynamic monitoring of intestinal mucosal ulcers according to claim 1, characterized in that, In step S500, the preset threshold T is set to 0.5 by default and can be adjusted according to clinical screening needs; when the threshold T decreases, the sensitivity increases. Elevated and specific Decrease; as the threshold T increases, sensitivity decreases. Reduced and specific Increase; Wherein, the sensitivity and specificity Defined according to the following formulas respectively: ； ； In the formula, The number of correctly predicted ulcer samples, This represents the number of ulcer samples that were missed during diagnosis. The number of normal samples that are correctly predicted. This represents the number of normal samples that were misdiagnosed.

6. The method for non-invasive dynamic monitoring of intestinal mucosal ulcers according to claim 1, characterized in that, It also includes a dynamic monitoring phase: S800: Acquire CT / MR images of the subject's abdominal and pelvic region at at least two different time points, and execute S400 to S700 respectively to compare the lesion mask, the probability value P, and the target location L at different time points to generate a dynamic monitoring report reflecting the progression or healing of the lesion. S900. Compare the ulcer classification results, classification probability values ​​P, and classification probability heatmaps H at different time points; S1000. Compare the target ulcer location L at different time points to determine whether lesion metastasis has occurred.

7. The method for non-invasive dynamic monitoring of intestinal mucosal ulcers according to claim 6, characterized in that, The comparison of lesion masks in S800 includes a quantitative comparison of the number of lesions, lesion volume, and / or lesion topology.

8. The method for non-invasive dynamic monitoring of intestinal mucosal ulcers according to any one of claims 1-7, characterized in that, The intestinal segment of clinical interest and the lesion are located in the digestive system, including the small intestine, large intestine and stomach; the method can be used for monitoring intestinal ulcers, intestinal erosions or intestinal inflammation.

Images