CT image feature extraction method for risk stratification of gastrointestinal stromal tumor

Abstract

The invention relates to the technical field of image data processing, in particular to a CT image feature extraction method for gastrointestinal stromal tumor risk stratification, which comprises the following steps of: firstly, acquiring an abdominal CT plane scanning period image and a venous period image of a gastrointestinal stromal tumor patient, and performing coarse segmentation by utilizing gray scale difference of the two periods to obtain a tumor suspected area; a suspected content area is identified by analyzing gray change characteristics of the edge of the area, whether the suspected content area is a real gastrointestinal tract content (such as a contrast agent or bubbles) is further verified, the confirmed content is removed, a pure gastrointestinal stromal tumor area is finally obtained, and radiomics characteristics for risk stratification are extracted from the pure gastrointestinal stromal tumor area. According to the scheme, interference of gastrointestinal tract contents on tumor boundary judgment and internal texture analysis is effectively eliminated, so that the segmentation result is more accurate, high-quality data is provided for image omics feature extraction, the accuracy and reliability of risk layering are enhanced, and preoperative noninvasive accurate evaluation is realized.

Description

A CT image feature extraction method for risk stratification of gastrointestinal stromal tumors Technical Field

[0001] This application belongs to the field of image data processing technology, specifically relating to a method for extracting CT image features for risk stratification of gastrointestinal stromal tumors. Background Technology

[0002] Gastrointestinal stromal tumors (GISTs) exhibit diverse biological behaviors; some tumors are indolent, while others are highly aggressive. Therefore, precise risk stratification is crucial for developing individualized treatment and follow-up plans. Currently, computed tomography (CT) imaging is commonly used for assisted diagnosis and assessment. However, GISTs often suffer from inaccurate tumor segmentation on CT images due to blurred boundaries and confusion with gastrointestinal contents, which in turn affects the reliability of subsequent radiomics feature extraction and risk stratification.

[0003] In existing technologies, methods for tumor segmentation based directly on grayscale or texture are easily affected by intracavitary contrast agents, bubbles, and other contents, resulting in extracted features that cannot truly reflect the biological characteristics of the tumor and affecting the accuracy of risk stratification models. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a CT image feature extraction method for risk stratification of gastrointestinal stromal tumors (GISTs). The method includes: acquiring abdominal computed tomography (CT) images of a patient with a GIST, the CT images including at least plain and venous phase images; coarsely segmenting the CT images based on the grayscale differences between the plain and venous phase images to obtain suspected GIST regions; analyzing the edge grayscale variation characteristics of the suspected GIST regions and identifying suspected contents regions located within the suspected GIST regions; verifying whether the suspected contents regions are actual gastrointestinal contents, and removing regions verified as containing contents from the suspected GIST regions to obtain pure GIST regions; and extracting radiomics features for risk stratification from the pure GIST regions.

[0005] In one embodiment, the step of coarsely segmenting the computed tomography image based on the grayscale difference between the plain scan image and the venous scan image to obtain a suspected region of gastrointestinal stromal tumor includes: registering and aligning the venous scan image with the plain scan image, and using the grayscale difference between the venous scan and plain scan images as the enhancement value of the target pixel; acquiring a set of target pixels with enhancement values ​​higher than a first threshold to form an initial enhancement region; selecting seed points within the initial enhancement region and growing the seed points based on a region growing algorithm to obtain an initial suspected region; and performing a morphological dilation operation on the initial suspected region to obtain the suspected region of gastrointestinal stromal tumor.

[0006] In one embodiment, analyzing the edge grayscale change characteristics of the suspected gastrointestinal stromal tumor (GIST) region and identifying the suspected contents region located within the suspected GIST region includes: selecting multiple sampling points on the edge of the suspected GIST region, and for a target sampling point among the multiple sampling points, obtaining a profile line perpendicular to the edge direction of the suspected GIST region and extending into and out of the suspected GIST region, respectively; obtaining a fitting curve of the pixel grayscale values ​​on the target profile line, and calculating the first-order derivative curve and the second-order derivative curve of the fitting curve; based on the first-order derivative curve... Based on the second derivative curve, the hard edge coefficient of the target sampling point is obtained; based on the hard edge coefficient, hard boundary segments on the edge of the suspected gastrointestinal stromal tumor region are identified, and the difference in gray-scale mean between the inner and outer sides of the hard boundary segments is analyzed to screen out high-density suspected boundaries and low-density suspected boundaries; region growth is performed from the high-density suspected boundaries and the low-density suspected boundaries into the region of the suspected gastrointestinal stromal tumor region to obtain high-density suspected contents region and low-density suspected contents region, and the high-density suspected contents region and the low-density suspected contents region are merged into the suspected contents region.

[0007] In one embodiment, obtaining the hard edge coefficient of the target sampling point based on the first-order derivative curve and the second-order derivative curve includes: obtaining a first mean of the gray values ​​of the first-order derivative curve; obtaining a second mean and a first variance of the absolute gray values ​​of the second-order derivative curve; and obtaining the hard edge coefficient of the target sampling point according to the first mean, the second mean, and the first variance.

[0008] In one embodiment, verifying whether the suspected contents region is genuine gastrointestinal contents includes: identifying liquid surface regions within gastrointestinal segments in the computed tomography image; measuring the distance between the hard boundary of the suspected contents region and the nearest liquid surface region to obtain a hard boundary liquid surface coefficient; obtaining an adhesion coefficient based on the hard boundary liquid surface coefficient; and determining whether the suspected contents region is genuine gastrointestinal contents based on the adhesion coefficient.

[0009] In one embodiment, measuring the distance between the hard boundary of the suspected contents region and the nearest liquid surface region to obtain a hard boundary liquid surface coefficient includes: obtaining the average distance between the hard boundary and the nearest liquid surface region; obtaining the closest distance between the hard boundary and the nearest liquid surface region; and obtaining the hard boundary liquid surface coefficient based on the average distance and the closest distance.

[0010] In one embodiment, obtaining the adhesion coefficient based on the hard boundary liquid level coefficient includes: obtaining a third mean of the distances from each pixel of the hard boundary to the nearest gastrointestinal tract wall surface; obtaining a second variance of the distances from each pixel of the hard boundary to the nearest gastrointestinal tract wall surface; and obtaining the adhesion coefficient based on the hard boundary liquid level coefficient, the third mean, and the second variance.

[0011] In one embodiment, determining whether the suspected contents region is real gastrointestinal contents based on the adhesion coefficient includes: determining that the suspected contents region is real gastrointestinal contents when the adhesion coefficient is greater than a preset adhesion threshold and the center of gravity of the suspected contents region conforms to the deposition law of gas rising or liquid sinking under gravity.

[0012] In one embodiment, the step of extracting radiomics features for risk stratification from the purified gastrointestinal stromal tumor region includes: extracting morphological features, first-order statistical features, and higher-order texture features from the purified gastrointestinal stromal tumor region; extracting dynamic enhancement features from the plain scan images and the venous scan images; and selecting a subset of key features associated with pathological risk levels from the morphological features, the first-order statistical features, the higher-order texture features, and the dynamic enhancement features, as the radiomics features for risk stratification.

[0013] In one embodiment, the method further includes: inputting the subset of key features into a machine learning classification model for training to construct a gastrointestinal stromal tumor risk stratification prediction model; and using the prediction model to perform risk stratification prediction of gastrointestinal stromal tumors in new computed tomography images.

[0014] The embodiments of this application have the following beneficial effects: First, abdominal CT plain scan and venous scan images of patients with gastrointestinal stromal tumors (GISTs) are acquired. Coarse segmentation is performed using the grayscale differences between the two phases to obtain suspected tumor regions. By analyzing the grayscale change characteristics of the region's edge, suspected contents are identified, and further verification is made to determine if they are genuine gastrointestinal contents (such as contrast agents or air bubbles). Confirmed contents are removed, ultimately yielding a pure GIST region. Radiomics features for risk stratification are then extracted from this purified region. This approach effectively eliminates the interference of gastrointestinal contents on tumor boundary determination and internal texture analysis, resulting in more accurate segmentation results and providing high-quality data for radiomics feature extraction. This enhances the accuracy and reliability of risk stratification, achieving precise preoperative non-invasive assessment.

[0015] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0016] To more clearly illustrate the implementation schemes of this application, the accompanying drawings used in the implementation schemes will be briefly introduced below. It should be understood that the accompanying drawings only show some implementation schemes of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from the accompanying drawings without creative effort.

[0017] Figure 1 is a flowchart illustrating a CT image feature extraction method for risk stratification of gastrointestinal stromal tumors according to an exemplary embodiment.

[0018] Figure 2 is a flowchart illustrating a method for coarsely segmenting computed tomography images based on the grayscale difference between plain and venous phase images to obtain suspected areas of gastrointestinal stromal tumors, according to an exemplary embodiment.

[0019] Figure 3 is a flowchart illustrating a method for analyzing the edge grayscale variation characteristics of a suspected gastrointestinal stromal tumor region and identifying a suspected content region located within the suspected gastrointestinal stromal tumor region, according to an exemplary embodiment.

[0020] Figure 4 is a flowchart illustrating a method for obtaining the hard edge coefficient of a target sampling point based on the first-order derivative curve and the second-order derivative curve, according to an exemplary embodiment.

[0021] Figure 5 is a flowchart illustrating a method for verifying whether a suspected contents region is genuine gastrointestinal contents according to an exemplary embodiment.

[0022] Figure 6 is a flowchart illustrating a method for measuring the distance between the hard boundary of a suspected content area and the nearest liquid surface area to obtain a hard boundary liquid surface coefficient according to an exemplary embodiment.

[0023] Figure 7 is a flowchart illustrating a method for obtaining the adhesion coefficient based on the hard boundary liquid surface coefficient according to an exemplary embodiment.

[0024] Figure 8 is a flowchart illustrating a method for determining whether a suspected contents region is a real gastrointestinal contents based on an adhesion coefficient, according to an exemplary embodiment.

[0025] Figure 9 is a flowchart illustrating a method for extracting radiomics features for risk stratification from a purified gastrointestinal stromal tumor region according to an exemplary embodiment.

[0026] Figure 10 is a flowchart illustrating yet another method for extracting CT image features for risk stratification of gastrointestinal stromal tumors according to an exemplary embodiment. Detailed Implementation

[0027] The specific embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this application.

[0028] It should be understood that the term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the following description.

[0029] It should be noted that the concepts of "first" and "second" mentioned in this application are only used to distinguish different devices, modules, or units, and are not used to limit the order of functions performed by these devices, modules, or units or their interdependencies. The modifications of "one" and "multiple" mentioned in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless explicitly stated in the context, they should be understood as "one or more". In the description of this application, unless otherwise stated, "multiple" means two or more, and other quantifiers are similar; "at least one item", "one item or multiple items", or similar expressions refer to any combination of these items, including any combination of single or multiple items.

[0030] Although operations or steps are described in a specific order in the accompanying drawings in the embodiments of this application, this should not be construed as requiring these operations or steps to be performed in the specific order or serial order shown, or requiring all of the shown operations or steps to be performed to obtain the desired result. In the embodiments of this application, these operations or steps may be performed serially; they may be performed in parallel; or a portion of these operations or steps may be performed.

[0031] The names of the messages or information exchanged between multiple devices in the embodiments of this application are for illustrative purposes only and are not intended to limit the scope of these messages or information. It is understood that before using the technical solutions of the various embodiments of this application, the type, scope of use, and usage scenarios of the personal information involved in this application should be disclosed to the user and their authorization obtained in accordance with relevant laws and regulations through appropriate means.

[0032] First, let me explain the application scenario of this application. Gastrointestinal stromal tumors (GISTs) have a very unique characteristic: not all GISTs recur or metastasize. Some GISTs are very "mild," and are essentially cured after resection; while others are highly invasive and prone to recurrence and metastasis. Therefore, it is unscientific to adopt a "one-size-fits-all" treatment and follow-up approach for all GIST patients. Thus, risk stratification of GISTs is necessary. CT image feature extraction involves quantifying and extracting various information that can describe the characteristics of the tumor from the patient's abdominal CT plain scan and enhanced scan images. These extracted CT features (especially radiomics features) are then correlated with the final pathological risk stratification results.

[0033] To extract CT image features for risk stratification of gastrointestinal stromal tumors (GISTs), GIST regions were acquired from patients' gastrointestinal CT images. During the segmentation of GISTs, their boundaries were not clear due to their invasiveness. Furthermore, some intraluminal or mixed-type GISTs may have gastrointestinal contents (such as high-density contrast agents or low-density air bubbles) attached to their surface, which could cause confusion with the internal texture of the GIST and thus seriously interfere with the determination of the true boundaries of the tumor.

[0034] Therefore, there is an urgent need for a CT image feature extraction method for risk stratification of gastrointestinal stromal tumors to solve the above problems. The present application will be described below with reference to specific embodiments.

[0035] Figure 1 is a flowchart illustrating a CT image feature extraction method for risk stratification of gastrointestinal stromal tumors according to an exemplary embodiment. As shown in Figure 1, this application embodiment provides a CT image feature extraction method for risk stratification of gastrointestinal stromal tumors, which may include the following steps: In step S10, abdominal computed tomography (CT) images of a patient with gastrointestinal stromal tumors are acquired, wherein the CT images include at least plain scan images and venous scan images.

[0036] In this step, abdominal computed tomographic images of patients with gastrointestinal stromal tumors (GISTs) are acquired. These images include at least plain and venous phase images. For example, when acquiring CT image data for risk stratification of GISTs, a standardized multi-phase contrast-enhanced scanning protocol can be used for suspected or confirmed GIST patients. The patient is in a supine position, and the scanning range extends from the top of the diaphragm to the level of the pubic symphysis. First, a plain abdominal scan is performed, followed by a non-ionic iodine contrast agent injected via the antecubital vein at a fixed flow rate. Intelligent triggering technology automatically initiates the arterial phase scan after a preset threshold is reached in the abdominal aortic target area, and a portal venous phase scan is performed 60-70 seconds after injection. All scans employ thin-slice reconstruction technology to ensure a slice thickness of no more than 1.0 mm, and dual reconstruction is performed using both a standard body reconstruction algorithm and a specific soft tissue reconstruction algorithm. The patient population includes adult men and women of different ages, with clinical manifestations including varying degrees of abdominal discomfort, gastrointestinal bleeding, or asymptomatic cases discovered incidentally during physical examinations. The tumors are located in different anatomical sites such as the stomach and small intestine, and their size and growth patterns exhibit typical intraluminal, extraluminal, and mixed growth characteristics, thus constructing an imaging dataset that comprehensively reflects the heterogeneity of the disease.

[0037] In step S20, based on the grayscale difference between the plain scan image and the venous scan image, the computed tomography image is coarsely segmented to obtain the suspected area of ​​gastrointestinal stromal tumor.

[0038] In this step, based on the grayscale difference between the plain and venous phase images, the computed tomography images are coarsely segmented to obtain suspected gastrointestinal stromal tumor (GIST) regions. For example, the venous phase images can be first registered and aligned with the plain phase images, and the grayscale difference between the venous and plain phase images can be used as the enhancement value of the target pixel. Then, a set of target pixels with enhancement values ​​higher than a first threshold is obtained to form an initial enhancement region. Next, seed points are selected within the initial enhancement region, and these seed points are grown using a region growing algorithm to obtain the initial suspected region. Finally, morphological dilation is performed on the initial suspected region to obtain the suspected GIST region.

[0039] In step S30, the edge grayscale change characteristics of the suspected gastrointestinal stromal tumor region are analyzed, and the suspected contents region located within the suspected gastrointestinal stromal tumor region is identified.

[0040] In this step, the edge grayscale change characteristics of the suspected area of ​​gastrointestinal stromal tumor are analyzed, and the suspected contents area located within the suspected area of ​​gastrointestinal stromal tumor is identified. For example, multiple sampling points can be selected on the edge of the suspected gastrointestinal stromal tumor (GIST) area. For the target sampling point among these points, a profile line is obtained that extends perpendicularly to the edge of the suspected GIST area and into and out of the suspected GIST area. Then, the fitting curve of the gray value of the pixel on the target profile line is obtained, and the first and second derivative curves of the fitting curve are calculated. Based on the first and second derivative curves, the hard edge coefficient of the target sampling point is obtained. Then, based on the hard edge coefficient, the hard boundary segment on the edge of the suspected GIST area is identified, and the difference in the mean gray value between the inner and outer sides of the hard boundary segment is analyzed to screen out high-density and low-density suspected boundaries. Finally, region growth is performed from the high-density and low-density suspected boundaries into the suspected GIST area to obtain a high-density suspected content area and a low-density suspected content area. The high-density and low-density suspected content areas are then merged into a suspected content area.

[0041] In step S40, it is verified whether the suspected contents area is real gastrointestinal contents, and the area verified as containing contents is removed from the suspected gastrointestinal stromal tumor area to obtain a pure gastrointestinal stromal tumor area.

[0042] In this step, it is verified whether the suspected contents area is indeed gastrointestinal content, and the areas verified as containing contents are removed from the suspected gastrointestinal stromal tumor (GIST) area to obtain a pure GIST area. For example, the fluid surface area within the gastrointestinal segment in the computed tomography (CT) image can be identified first. Then, the distance between the hard boundary of the suspected contents area and the nearest fluid surface area is measured to obtain the hard boundary fluid surface coefficient. Based on the hard boundary fluid surface coefficient, the adhesion coefficient is obtained. Finally, based on the adhesion coefficient, it is determined whether the suspected contents area is indeed gastrointestinal content.

[0043] In step S50, radiomics features for risk stratification are extracted from the purified gastrointestinal stromal tumor region.

[0044] In this step, radiomics features for risk stratification are extracted from the purified gastrointestinal stromal tumor (GIST) region. For example, morphological features, first-order statistical features, and higher-order texture features can be extracted first from the purified GIST region. Then, dynamic enhancement features are extracted from plain and venous images. Finally, a subset of key features associated with pathological risk levels is selected from the morphological features, first-order statistical features, higher-order texture features, and dynamic enhancement features to serve as radiomics features for risk stratification.

[0045] This application first acquires abdominal CT images of patients with gastrointestinal stromal tumors (GISTs) during the plain and venous phases. Coarse segmentation is performed using the grayscale differences between the two phases to obtain suspected tumor regions. By analyzing the grayscale change characteristics of the region's edges, suspected contents are identified, and further verification is conducted to determine if they are genuine gastrointestinal contents (such as contrast agents or air bubbles). Confirmed contents are removed, ultimately yielding a pure GIST region. Radiomics features for risk stratification are then extracted from this purified region. This approach effectively eliminates the interference of gastrointestinal contents on tumor boundary determination and internal texture analysis, resulting in more accurate segmentation and providing high-quality data for radiomics feature extraction. This enhances the accuracy and reliability of risk stratification, enabling precise, non-invasive preoperative assessment.

[0046] Figure 2 is a flowchart illustrating an exemplary embodiment of a method for coarsely segmenting computed tomography (CT) images based on the grayscale difference between plain and venous phase images to obtain a suspected region of a gastrointestinal stromal tumor (GIST). As shown in Figure 2, the method for coarsely segmenting the CT images based on the grayscale difference between the plain and venous phase images to obtain a suspected GIST region may include the following steps: In step S201, the venous phase image is registered and aligned with the plain phase image, and the grayscale difference between the target pixel in the venous phase and the plain phase is used as the enhancement value of the target pixel.

[0047] In this step, the venous phase image and the plain scan image are registered and aligned, and the grayscale difference between the venous phase and the plain scan image of the target pixel is used as the enhancement value of the target pixel. For example, the GIST can be coarsely segmented first, and then interference items can be removed. GIST is a highly vascularized tumor, so GIST will have strong enhancement in CT contrast-enhanced scans. The plain scan image and the venous phase image of the patient's stomach CT are aligned; the grayscale value h1 of the target pixel in the plain scan image and the grayscale value h2 of the target pixel in the venous phase are obtained; the difference h1' = h2 - h1 between the grayscale value of the target pixel in the venous phase and the grayscale value in the plain scan image is used as the enhancement value of the target pixel.

[0048] In step S202, a set of target pixels with enhancement values ​​higher than the first threshold is obtained to form an initial enhancement region.

[0049] In this step, a set of target pixels with enhancement values ​​higher than a first threshold is obtained to form an initial enhancement region. For example, the first threshold can be obtained using Otsu thresholding.

[0050] In step S203, seed points are selected within the initial reinforcement region, and the seed points are grown based on the region growth algorithm to obtain the initial suspected region.

[0051] In this step, seed points are selected within the initial enhancement region, and these seed points are grown using a region growing algorithm to obtain the initial suspected region. For example, the pixel with the highest grayscale value in each enhancement region can be used as the seed point; region growing begins from the seed point; the growing conditions allow for a relatively wide range of grayscale values ​​(e.g., a difference of 20 from the seed point's grayscale value) to include tumor regions with uneven density and their edges; and the new pixel must be connected to the already grown region; the growing process continues until no more neighboring pixels meeting this lenient criterion can be found; thus, the initial suspected region is obtained.

[0052] In step S204, a morphological expansion operation is performed on the initial suspected region to obtain the suspected region of gastrointestinal stromal tumor.

[0053] In this step, a morphological dilation operation is performed on the initial suspected region to obtain the suspected region of gastrointestinal stromal tumor (GIST). For example, since the tumor may have an invasive and ambiguous boundary with normal tissue, the true boundary may be inside or outside the edge of the coarse segmentation; the morphological dilation operation is applied to the obtained initial suspected region to uniformly dilate the initial suspected region outward by a preset number of pixels (e.g., 5 pixels); thereby obtaining the suspected GIST region in the image.

[0054] This application fully utilizes the imaging characteristics of gastrointestinal stromal tumors, such as their rich blood supply and significant enhancement after contrast administration, to quickly locate suspicious areas. At the same time, by expanding to encompass the blurred boundaries, it significantly reduces the risk of missed segmentation due to unclear boundaries, laying a complete regional foundation for subsequent content removal and feature extraction.

[0055] Figure 3 is a flowchart illustrating a method for analyzing the edge grayscale change characteristics of a suspected gastrointestinal stromal tumor (GIST) region and identifying suspected contents within that region, according to an exemplary embodiment. As shown in Figure 3, the method for analyzing the edge grayscale change characteristics of the suspected GIST region and identifying suspected contents within that region may include the following steps: In step S301, multiple sampling points are selected on the edge of the suspected GIST region, and for a target sampling point among the multiple sampling points, a cross-sectional line is obtained that is perpendicular to the edge of the suspected GIST region and extends into and out of the suspected GIST region, respectively.

[0056] In this step, multiple sampling points are selected on the edge of the suspected gastrointestinal stromal tumor (GIST) region. For the target sampling point among these points, cross-sectional lines are obtained that are perpendicular to the edge of the suspected GIST region and extend into and out of the suspected GIST region. For example, the obtained suspected GIST region may contain some contents of the gastric cavity, so it is necessary to analyze the edge of the suspected region to extract the contents. The edge of the suspected GIST region in the image obtained by the above operation is obtained, and a target sampling point is set at every pixel interval on the edge. For each target sampling point, a straight line passing through its adjacent edge pixels is obtained as the sampling point line of the target sampling point, and a straight line passing through the target sampling point and perpendicular to the target sampling point line is obtained as the perpendicular line of the target sampling point. Line segments of a certain length (e.g., 5 mm) on the perpendicular line of the target sampling point, located inside and outside the suspected region, are obtained as the cross-sectional lines of the target sampling point.

[0057] In step S302, the fitting curve of the gray value of the pixel on the target profile line is obtained, and the first derivative curve and the second derivative curve of the fitting curve are calculated.

[0058] In this step, the fitted curves of the grayscale values ​​of pixels on the target profile line are obtained, and the first-order and second-order derivative curves of the fitted curves are calculated. For example, the real tumor boundary is biologically infiltrated, therefore the real tumor boundary exhibits a gradual change; while the boundary between the tumor and its contents is physically in contact, therefore the boundary exhibits an abrupt change. For each profile line, the pixels located on the profile line and their grayscale values ​​are obtained, and the fitted curve q1 of the grayscale values ​​on the profile line is obtained. The first-order derivative curve q2 and the second-order derivative curve q3 of the fitted curve q1 of the sampled point profile line are calculated respectively. The first-order derivative reflects the rate of change of the pixel grayscale values ​​in the fitted curve q1, while the second-order derivative is calculated based on the curvature of the fitted curve q1, reflecting the curvature characteristics.

[0059] In step S303, the hard edge coefficient of the target sampling point is obtained based on the first derivative curve and the second derivative curve.

[0060] In this step, the hard edge coefficient of the target sampling point is obtained based on the first-order derivative curve and the second-order derivative curve. For example, the first mean of the gray value of the first-order derivative curve can be obtained first, then the second mean and the first variance of the absolute gray value of the second-order derivative curve can be obtained, and then the hard edge coefficient of the target sampling point can be obtained based on the first mean, the second mean, and the first variance.

[0061] In step S304, hard boundary segments on the edge of the suspected gastrointestinal stromal tumor region are identified based on the hard edge coefficient, and the difference in gray-scale mean between the inner and outer sides of the hard boundary segments is analyzed to screen out high-density suspected boundaries and low-density suspected boundaries.

[0062] In this step, hard boundary segments on the edge of suspected gastrointestinal stromal tumor regions are identified based on the hard edge coefficient, and the difference in gray-scale mean between the inner and outer sides of the hard boundary segments is analyzed to screen out high-density and low-density suspected boundaries. For example, for the hard edge coefficient YX of the sampling points, Otsu threshold segmentation can be used to obtain sampling points where YX is greater than a set threshold; based on the obtained set of sampling points, hard edge segments on the suspected region edge can be obtained; for the obtained hard edge segments, the perpendicular lines of the pixels in the hard edge segments can be obtained respectively, and the perpendicular lines of each pixel are taken to be located at the two adjacent pixels on the inside and outside of the hard edge segment, thereby obtaining the inner and outer pixels of the hard edge; the mean gray value h1 of the inner pixels and the mean gray value h2 of the outer pixels can be obtained respectively; and the mean difference h2' of the gray values ​​of the inner and outer pixels can be obtained; when the gray value of the inner side of the hard boundary is higher or lower than that of the outer body tissue, it indicates that the hard boundary may contain contents; double threshold segmentation is performed to segment out the parts where the mean difference h2' of the hard edge is extremely large and extremely small, which are used as high-density suspected boundaries and low-density suspected boundaries respectively; the remaining boundaries are normal tumor tissue boundaries.

[0063] In step S305, regional growth is performed from the high-density suspected boundary and the low-density suspected boundary into the region of the suspected gastrointestinal stromal tumor, respectively, to obtain a high-density suspected content region and a low-density suspected content region, and the high-density suspected content region and the low-density suspected content region are merged into the suspected content region.

[0064] In this step, region growth is performed from the high-density and low-density suspected boundaries into the suspected gastrointestinal stromal tumor region, respectively, to obtain high-density suspected contents regions and low-density suspected contents regions. These two regions are then merged into a single suspected contents region. For example, the entire suspected region is segmented using a dual-threshold method to obtain a high threshold y1 and a low threshold y2. For the obtained high- and low-density suspected boundaries, region growth is performed from the boundaries into the suspected region, setting thresholds y1 and y2 as growth conditions to obtain high- and low-density suspected regions. These high- and low-density suspected regions represent potential contents within the tumor and are therefore merged into the suspected contents region.

[0065] This application's embodiments use a combination of mathematical morphology and differential geometry to accurately locate the physical contact boundary between the tumor and its contents, effectively distinguishing between the tumor's own biological infiltration boundary and the mutation boundary formed by the attachment of its contents, thus providing a reliable candidate region for the subsequent precise removal of the contents.

[0066] Figure 4 is a flowchart illustrating a method for obtaining the hard edge coefficient of a target sampling point based on the first-order derivative curve and the second-order derivative curve, according to an exemplary embodiment. As shown in Figure 4, obtaining the hard edge coefficient of the target sampling point based on the first-order derivative curve and the second-order derivative curve may include the following steps: In step S3031, obtaining the first mean value of the grayscale value of the first-order derivative curve.

[0067] In this step, the first mean value q2' of the grayscale value of the first derivative curve is obtained.

[0068] In step S3032, the second mean and the first variance of the absolute gray values ​​of the second derivative function curve are obtained.

[0069] In this step, the second mean q3' and the first variance c1 of the absolute gray values ​​of the second derivative function curve are obtained.

[0070] In step S3033, the hard edge coefficient of the target sampling point is obtained based on the first mean, the second mean, and the first variance.

[0071] In this step, the hard edge coefficient YX of the target sampling point is obtained based on the first mean q2', the second mean q3', and the first variance c1. For example, the hard edge coefficient YX of the target sampling point can be obtained by the following formula: YX = norm( Formula 1, where, The value is a very small positive number, such as 0.001. The normalization is performed using Max-Min normalization to the range [0,1] unless otherwise specified.

[0072] for The larger the value, the greater the change in grayscale value of the pixels on the profile line of the target sampling point. Therefore, the target sampling point is more likely to be a sudden edge, and thus its hard edge coefficient is larger. The larger the value, the greater the curvature variance of the gray value fitting curve of the pixel on the profile line of the target sampling point, and the smaller the mean curvature. Therefore, it indicates that the change of the gray value fitting curve is more acute, and thus the hard edge coefficient of the target sampling point is larger.

[0073] This application's embodiments combine two dimensions: grayscale change rate and curvature sharpness, effectively distinguishing between steep boundaries formed by physical contact between tumors and their contents, and gradual boundaries formed by biological infiltration between tumors and normal tissue. Compared to single threshold or simple gradient detection, this coefficient is more robust to noise and can adaptively highlight truly hard boundaries that need to be removed, improving the accuracy of suspected contents identification.

[0074] Figure 5 is a flowchart illustrating a method for verifying whether a suspected content area is genuine gastrointestinal content according to an exemplary embodiment. As shown in Figure 5, verifying whether the suspected content area is genuine gastrointestinal content may include the following steps: In step S401, identifying liquid surface areas within gastrointestinal segments in the computed tomography image.

[0075] In this step, fluid-filled areas within gastrointestinal segments in computed tomography (CT) images are identified. For example, high-density suspected areas might be undiluted contrast agent adhering to the tumor surface, while low-density suspected areas might be air bubbles on the tumor surface. For the acquired CT images, the negative y-axis direction is taken as the direction of gravity based on the acquisition viewpoint. A pre-trained 3D U-Net neural network automatically and coarsely constructs the gastrointestinal segments where suspected GIST regions are located in the images. Within the constructed gastrointestinal segments, the algorithm searches for specific low-grayscale areas (e.g., 0-50). The neural network automatically detects these areas, which, when released in the direction of gravity, form a continuous, well-defined planar interface perpendicular to the direction of gravity, thus obtaining the fluid-filled areas within the current gastrointestinal segment. It should be noted that the 3D U-Net neural network and its training process are techniques well-known to those skilled in the art and will not be further limited or elaborated upon here.

[0076] In step S402, the distance between the hard boundary of the suspected contents region and the nearest liquid surface region is measured to obtain the hard boundary liquid surface coefficient.

[0077] In this step, the distance between the hard boundary of the suspected contents area and the nearest liquid surface area is measured to obtain the hard boundary liquid surface coefficient. For example, the average distance between the hard boundary and the nearest liquid surface area can be obtained first, then the closest distance between the hard boundary and the nearest liquid surface area can be obtained, and then the hard boundary liquid surface coefficient can be obtained based on the average distance and the closest distance.

[0078] In step S403, the adhesion coefficient is obtained based on the hard boundary liquid level coefficient.

[0079] In this step, the adhesion coefficient is obtained based on the hard boundary liquid level coefficient. For example, the third mean of the distance from each pixel of the hard boundary to the nearest gastrointestinal tract wall surface can be obtained first, then the second variance of the distance from each pixel of the hard boundary to the nearest gastrointestinal tract wall surface can be obtained, and then the adhesion coefficient can be obtained based on the hard boundary liquid level coefficient, the third mean, and the second variance.

[0080] In step S404, based on the adhesion coefficient, it is determined whether the suspected contents area is actual gastrointestinal contents.

[0081] In this step, based on the adhesion coefficient, it is determined whether the suspected content area is actually gastrointestinal content. For example, if the adhesion coefficient is greater than a preset adhesion threshold and the center of gravity of the suspected content area conforms to the deposition law of gas rising or liquid sinking under gravity, the suspected content area can be determined to be actual gastrointestinal content.

[0082] The embodiments of this application transform the physical constraints (gravity settling, wall adhesion) of contents within the lumen into calculable imaging indicators, making the removal operation not only dependent on density differences but also more consistent with physiological and anatomical realities, significantly reducing the risk of misjudging real lesions such as tumor necrosis and hemorrhage as contents.

[0083] Figure 6 is a flowchart illustrating a method for measuring the distance between the hard boundary of a suspected content area and the nearest liquid surface area to obtain a hard boundary liquid surface coefficient according to an exemplary embodiment. As shown in Figure 6, measuring the distance between the hard boundary of the suspected content area and the nearest liquid surface area to obtain the hard boundary liquid surface coefficient may include the following steps: In step S4021, the average distance between the hard boundary and the nearest liquid surface area is obtained.

[0084] In this step, the average distance d between the hard boundary and the nearest liquid surface region is obtained. For example, for the obtained content region, its hard boundary is obtained; the liquid surface region closest to the hard boundary is obtained, and the distance d between the hard boundary and the nearest liquid surface region is obtained.

[0085] In step S4022, the closest distance between the hard boundary and the nearest liquid surface region is obtained.

[0086] In this step, the nearest distance d' between the hard boundary and the nearest liquid surface region is obtained. For example, the Hausdorff distance can be used to obtain the nearest distance d' between the hard boundary and the nearest liquid surface region.

[0087] In step S4023, the hard boundary liquid level coefficient is obtained based on the average distance and the nearest distance.

[0088] In this step, the hard boundary liquid level coefficient YM is obtained based on the average distance d and the nearest distance d'. For example, the hard boundary liquid level coefficient YM can be obtained by the following formula: YM = norm(1 / (d × d' + ... Formula 2, where, For a very small positive number, such as 0.001, norm is a normalization process.

[0089] When measuring the distance between the hard boundary of a suspected content area and the nearest liquid surface area, the average distance from all pixels on the hard boundary to the liquid surface area is first calculated, reflecting the overall proximity of the two. Simultaneously, the Hausdorff distance is used to calculate the closest distance between them, characterizing the degree of contact between the boundary and the liquid surface at their closest point. Multiplying the average distance and the closest distance, taking the reciprocal, and normalizing the result yields the hard boundary-liquid surface coefficient. A larger coefficient indicates that the hard boundary is not only close to the liquid surface overall but also has locally tightly fitted sections, and its shape more closely resembles the straight characteristics of the liquid surface boundary.

[0090] The embodiments of this application take into account both overall similarity and local minimum distance, and can sensitively capture the boundary of contents attached near the liquid surface, providing reliable input parameters for subsequent calculation of the adhesion coefficient.

[0091] Figure 7 is a flowchart illustrating a method for obtaining an adhesion coefficient based on a hard boundary liquid surface coefficient according to an exemplary embodiment. As shown in Figure 7, obtaining the adhesion coefficient based on the hard boundary liquid surface coefficient may include the following steps: In step S4031, obtaining the third average value of the distance from each pixel point of the hard boundary to the nearest gastrointestinal tract wall surface.

[0092] In this step, the third mean l' of the distances from each pixel of the hard boundary to the nearest gastrointestinal wall surface is obtained. For example, since tumors are solid tissue driven by cell proliferation, they are almost unaffected by immediate gravity; while liquid or gaseous contents, when at rest, are inevitably affected by gravity and physically constrained by the shape of the gastrointestinal lumen. Therefore, the gastrointestinal segment where the suspected GIST region is located is obtained; and based on the obtained gastrointestinal segment, the interior and exterior of the current segment are determined. Since the suspected contents are intestinal contents, they should be entirely within the gastrointestinal lumen. The number of pixels n1 inside the gastrointestinal segment and the number of pixels n2 outside the suspected content region are obtained. The three-dimensional mesh of the hard boundary of the suspected content region is obtained, and the mesh of the inner wall surface of the gastrointestinal segment where it is located is extracted. The shortest distance l between each pixel in the hard boundary of the suspected content region and the inner wall surface mesh is obtained; then, the third mean l' of the shortest distance l between each pixel in the hard boundary is obtained.

[0093] In step S4032, the second variance of the distance from each pixel of the hard boundary to the nearest gastrointestinal tract wall surface is obtained.

[0094] In this step, the second variance c2 of the distance l from each pixel of the hard boundary to the nearest gastrointestinal tract wall surface is obtained.

[0095] In step S4033, the adhesion coefficient is obtained based on the hard boundary liquid level coefficient, the third mean, and the second variance.

[0096] In this step, the adhesion coefficient FZ is obtained based on the hard boundary liquid level coefficient YM, the third mean l', and the second variance c2. For example, the adhesion coefficient FZ can be obtained by the following formula: FZ = norm(YM / (l'×c2+)) Formula 3, where, For a very small positive number, such as 0.001, norm is a normalization process.

[0097] The adhesion coefficient is obtained by first calculating the mean distance from each pixel on the hard boundary to the nearest gastrointestinal wall surface. A smaller mean value indicates that the contents region is generally closely attached to the inner wall of the lumen. Simultaneously, the variance of this distance is calculated; a smaller variance indicates that the distances between the points on the hard boundary and the inner wall are concentrated, showing a uniform attachment. The adhesion coefficient is obtained by combining the hard boundary liquid level coefficient with the aforementioned mean and variance distances and then normalizing the result. This coefficient comprehensively reflects the degree to which the suspected contents region conforms to the morphology of the gastrointestinal lumen in spatial location: a high liquid level coefficient indicates a close relationship with the liquid surface, while a small mean and small variance distance indicate that it is generally closely attached to the inner wall and has a consistent shape.

[0098] The multi-dimensional adhesion quantification method of this application embodiment makes the judgment of contents no longer limited to a single density or boundary feature, but evaluates it within the geometric constraints of the entire gastrointestinal lumen, which significantly improves the accuracy of verification.

[0099] Figure 8 is a flowchart illustrating a method for determining whether a suspected content area is genuine gastrointestinal content based on an adhesion coefficient, according to an exemplary embodiment. As shown in Figure 8, determining whether the suspected content area is genuine gastrointestinal content based on the adhesion coefficient may include the following steps: In step S4041, if the adhesion coefficient is greater than a preset adhesion threshold and the center of gravity of the suspected content area conforms to the deposition law of gas rising or liquid sinking under gravity, the suspected content area is determined to be genuine gastrointestinal content.

[0100] In this step, if the adhesion coefficient is greater than a preset adhesion threshold (e.g., 0.7) and the center of gravity of the suspected content area conforms to the deposition law of gas rising or liquid sinking under gravity, the suspected content area is determined to be actual gastrointestinal contents. For example, when the suspected area is contents within the gastrointestinal lumen, the contents will adhere to the inner wall of the lumen and thus conform to the shape of the lumen space. When determining whether a suspected content area is actual gastrointestinal contents, an adhesion coefficient threshold is set, and areas exceeding this threshold are considered to have high adhesion characteristics. Simultaneously, the center of gravity of this area is analyzed, combined with a preset gravity direction in the image, to determine whether it conforms to the deposition law of gas rising and liquid sinking under gravity. Only suspected content areas that simultaneously meet both conditions—an adhesion coefficient higher than the threshold and a center of gravity position conforming to the gravity deposition law—are ultimately confirmed as actual gastrointestinal contents. These content areas are then removed from the GIST suspected areas, thus obtaining the GIST region in the image.

[0101] This application combines geometric attachment features with physical sedimentation features, effectively eliminating false positive areas that are actually exophytic growths of tumors but are attached to the inner wall, as well as solid tissues that are abnormally dense but not constrained by gravity, thus ensuring the accuracy and physiological rationality of the removal operation.

[0102] Figure 9 is a flowchart illustrating a method for extracting radiomics features for risk stratification from a purified gastrointestinal stromal tumor region according to an exemplary embodiment. As shown in Figure 9, the extraction of radiomics features for risk stratification from the purified gastrointestinal stromal tumor region may include the following steps: in step S501, extracting morphological features, first-order statistical features, and higher-order texture features of the purified gastrointestinal stromal tumor region.

[0103] In step S502, dynamic enhancement features are extracted from the plain scan image and the venous scan image.

[0104] In step S503, a subset of key features associated with pathological risk levels is selected from the morphological features, the first-order statistical features, the higher-order texture features, and the dynamic enhancement features to serve as the radiomics features used for risk stratification.

[0105] For example, three types of radiomics features can be extracted from the purified gastrointestinal stromal tumor region after content removal: first, morphological features, including volume, surface area, sphericity, and surface irregularity, quantifying the overall geometric properties of the tumor; second, first-order statistical features, reflecting the mean, variance, skewness, and kurtosis of CT value distribution, characterizing the gray-level heterogeneity within the tumor; and third, higher-order texture features, based on algorithms such as gray-level co-occurrence matrix and run-length matrix, describing the complex patterns of pixel spatial arrangement, indirectly reflecting the heterogeneity of the tumor's microstructure. Simultaneously, dynamic enhancement features, including enhancement peak, enhancement rate, and clearance pattern, are extracted from plain and venous phase images to reflect the tumor's hemodynamic characteristics. Based on this, variance thresholding, correlation analysis, and LASSO regression are used to screen out a subset of key features most strongly correlated with pathological risk levels from the hundreds of features mentioned above.

[0106] The feature extraction and screening process in this application compresses high-dimensional image information into a compact set of features with biological interpretability and clinical predictive power, which serves as radiomics features for risk stratification, providing high-quality input for subsequent risk stratification analysis and modeling.

[0107] Figure 10 is a flowchart illustrating another CT image feature extraction method for risk stratification of gastrointestinal stromal tumors according to an exemplary embodiment. As shown in Figure 10, the method may further include the following steps: In step S60, the subset of key features is input into a machine learning classification model for training to construct a gastrointestinal stromal tumor risk stratification prediction model.

[0108] In step S70, the prediction model is used to perform risk stratification prediction for gastrointestinal stromal tumors in new computed tomography images.

[0109] For example, a selected subset of key features can be used as input to train a machine learning classification model, constructing a risk stratification prediction model for gastrointestinal stromal tumors (GISTs). During training, cross-validation is employed to optimize model hyperparameters, and classifiers such as support vector machines, random forests, or logistic regression are selected, labeled with pathological risk levels (e.g., NIH modified standards). The model's discriminative power, calibration, and clinical applicability are evaluated on an independent test set. Finally, the trained prediction model is used to automatically predict the risk stratification of GISTs in new computed tomography images.

[0110] This application's embodiments directly link preoperative non-invasive imaging features with postoperative pathological risks, enabling clinicians to obtain objective and quantitative risk assessment results before surgery, providing a basis for individualized treatment decisions. This avoids the limitations of traditional stratification methods that rely on limited indicators such as tumor size and mitotic figures, and truly achieves precise risk stratification based on the entire tumor imaging phenotype.

[0111] This application also provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement the steps of the CT image feature extraction method for risk stratification of gastrointestinal stromal tumors provided in this application.

[0112] In another exemplary embodiment, a computer program product is also provided, comprising a computer program executable by a programmable electronic device, the computer program having a code portion for performing the above-described CT image feature extraction method for risk stratification of gastrointestinal stromal tumors when executed by the programmable electronic device.

[0113] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

Claims

1. A method for extracting CT image features for risk stratification of gastrointestinal stromal tumors, characterized in that, The method includes: acquiring abdominal computed tomography (CT) images of a patient with gastrointestinal stromal tumor (GIST), the CT images including at least plain scan images and venous scan images; coarsely segmenting the CT images based on the grayscale differences between the plain scan images and the venous scan images to obtain suspected GIST regions; analyzing the edge grayscale change characteristics of the suspected GIST regions and identifying suspected contents regions located within the suspected GIST regions; verifying whether the suspected contents regions are actual gastrointestinal contents, and removing regions verified as containing contents from the suspected GIST regions to obtain pure GIST regions; and extracting radiomics features for risk stratification from the pure GIST regions.

2. The CT image feature extraction method for risk stratification of gastrointestinal stromal tumors according to claim 1, characterized in that, The step of coarsely segmenting the computed tomography (CT) image based on the grayscale difference between the plain scan image and the venous scan image to obtain a suspected region of gastrointestinal stromal tumor (GIST) includes: registering and aligning the venous scan image with the plain scan image, and using the grayscale difference between the venous scan and plain scan images as the enhancement value of the target pixel; acquiring a set of target pixels with enhancement values ​​higher than a first threshold to form an initial enhancement region; selecting seed points within the initial enhancement region and growing the seed points based on a region growing algorithm to obtain an initial suspected region; and performing a morphological dilation operation on the initial suspected region to obtain the suspected GIST region.

3. The CT image feature extraction method for risk stratification of gastrointestinal stromal tumors according to claim 1, characterized in that, The analysis of the edge grayscale change characteristics of the suspected gastrointestinal stromal tumor (GIST) region and the identification of suspected contents within the suspected GIST region includes: selecting multiple sampling points on the edge of the suspected GIST region; for a target sampling point among the multiple sampling points, obtaining a profile line perpendicular to the edge of the suspected GIST region and extending into and out of the suspected GIST region, respectively; obtaining a fitting curve of the pixel grayscale values ​​on the target profile line; and calculating the first-order derivative curve and the second-order derivative curve of the fitting curve; based on the first-order derivative curve and the second-order derivative curve... The derivative function curve is used to obtain the hard edge coefficient of the target sampling point; based on the hard edge coefficient, hard boundary segments on the edge of the suspected gastrointestinal stromal tumor region are identified, and the difference in gray-scale mean between the inner and outer sides of the hard boundary segments is analyzed to screen out high-density suspected boundaries and low-density suspected boundaries; region growth is performed from the high-density suspected boundaries and the low-density suspected boundaries into the region of the suspected gastrointestinal stromal tumor region to obtain high-density suspected contents region and low-density suspected contents region, and the high-density suspected contents region and the low-density suspected contents region are merged into the suspected contents region.

4. The CT image feature extraction method for risk stratification of gastrointestinal stromal tumors according to claim 3, characterized in that, The step of obtaining the hard edge coefficient of the target sampling point based on the first derivative curve and the second derivative curve includes: obtaining the first mean of the gray value of the first derivative curve; obtaining the second mean and the first variance of the absolute gray value of the second derivative curve; and obtaining the hard edge coefficient of the target sampling point according to the first mean, the second mean and the first variance.

5. The CT image feature extraction method for risk stratification of gastrointestinal stromal tumors according to claim 1, characterized in that, The verification of whether the suspected contents region is real gastrointestinal contents includes: identifying the liquid surface region within the gastrointestinal segment in the computed tomography image; measuring the distance between the hard boundary of the suspected contents region and the nearest liquid surface region to obtain the hard boundary liquid surface coefficient; obtaining the adhesion coefficient based on the hard boundary liquid surface coefficient; and determining whether the suspected contents region is real gastrointestinal contents based on the adhesion coefficient.

6. The CT image feature extraction method for risk stratification of gastrointestinal stromal tumors according to claim 5, characterized in that, Measuring the distance between the hard boundary of the suspected contents region and the nearest liquid surface region to obtain the hard boundary liquid surface coefficient includes: obtaining the average distance between the hard boundary and the nearest liquid surface region; obtaining the closest distance between the hard boundary and the nearest liquid surface region; and obtaining the hard boundary liquid surface coefficient based on the average distance and the closest distance.

7. The CT image feature extraction method for risk stratification of gastrointestinal stromal tumors according to claim 5, characterized in that, The step of obtaining the adhesion coefficient based on the hard boundary liquid level coefficient includes: obtaining the third mean of the distance from each pixel of the hard boundary to the nearest gastrointestinal tract wall surface; obtaining the second variance of the distance from each pixel of the hard boundary to the nearest gastrointestinal tract wall surface; and obtaining the adhesion coefficient according to the hard boundary liquid level coefficient, the third mean, and the second variance.

8. The CT image feature extraction method for risk stratification of gastrointestinal stromal tumors according to claim 5, characterized in that, The step of determining whether the suspected contents region is a real gastrointestinal contents based on the adhesion coefficient includes: determining that the suspected contents region is a real gastrointestinal contents when the adhesion coefficient is greater than a preset adhesion threshold and the center of gravity of the suspected contents region conforms to the deposition law of gas rising or liquid sinking under the action of gravity.

9. The method for extracting CT image features for risk stratification of gastrointestinal stromal tumors according to claim 1, characterized in that, The step of extracting radiomics features for risk stratification from the purified gastrointestinal stromal tumor region includes: extracting morphological features, first-order statistical features, and higher-order texture features from the purified gastrointestinal stromal tumor region; extracting dynamic enhancement features from the plain scan images and the venous scan images; and selecting a subset of key features associated with pathological risk levels from the morphological features, the first-order statistical features, the higher-order texture features, and the dynamic enhancement features, as the radiomics features for risk stratification.

10. The method for extracting CT image features for risk stratification of gastrointestinal stromal tumors according to claim 9, characterized in that, The method further includes: inputting the subset of key features into a machine learning classification model for training to construct a risk stratification prediction model for gastrointestinal stromal tumors; and using the prediction model to perform risk stratification prediction of gastrointestinal stromal tumors in new computed tomography images.

Images