Deep learning based colorectal cancer mr image assisted diagnosis system
By using a correction module to identify and correct the inner and outer edges of lesion areas in colorectal cancer MRI images, the training accuracy problem caused by manual annotation errors is solved, resulting in more accurate lesion identification and model training, and improving the reliability of diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHONGSHAN HOSPITAL FUDAN UNIV
- Filing Date
- 2026-05-07
- Publication Date
- 2026-07-31
AI Technical Summary
In existing technologies, human annotation errors are caused by factors such as overlap of lesion edges with non-tumor tissue signals, blurred boundaries, and organ movement, which affect the training accuracy and diagnostic reliability of deep learning models.
The acquisition module obtains MRI images of colorectal cancer patients and manually labeled original lesion areas. The correction module determines the thickness distribution and morphological similarity of the inner and outer edges, identifies areas with abnormal thickness, and selects gradient peak points in the area to be corrected to determine the real inner and outer edge points, constructing real lesion areas as deep learning labels.
It improves the accuracy of lesion identification, avoids errors caused by manual annotation, ensures that the auxiliary diagnostic model can be trained based on accurate lesion features, and improves the reliability of model identification.
Smart Images
Figure CN122492610A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and more specifically to a deep learning-based MRI image-assisted diagnostic system for colorectal cancer. Background Technology
[0002] Magnetic resonance imaging (MRI) has become an important diagnostic tool for colorectal cancer due to its ability to clearly present lesion details. Deep learning-based MRI-assisted diagnostic systems for colorectal cancer are also increasingly being used. In practical colorectal cancer diagnosis, it is usually necessary to manually delineate the lesion area on the MRI images of colorectal cancer patients beforehand, serving as the real-world label for training the network in the aforementioned deep learning-based MRI-assisted diagnostic system for colorectal cancer.
[0003] In related technologies, the subtraction method is commonly used to manually annotate lesion areas in MRI images of colorectal cancer patients. Specifically, the annotator delineates the outer and inner edges of the tumor in the MRI image of the colorectal cancer patient, and obtains the actual thickness and shape of the lesion through the difference operation between the two edges.
[0004] However, in actual annotation, rectal tumor lesions grow invasively along the intestinal wall. The inner edge of the lesion needs to be distinguished from non-tumor tissues such as residual fluid and feces in the intestinal lumen. These non-tumor tissues have similar signals to tumors, making them difficult to identify accurately. The outer edge of the lesion needs to be clearly defined in relation to surrounding blood vessels, lymph nodes, rectal mesofascia, and adjacent organs. However, due to signal overlap and organ movement, manual delineation is prone to omissions or over-drawing of boundaries, resulting in errors in the training labels used for auxiliary diagnostic models. This reduces the training accuracy of deep learning models and affects the accuracy of lesion identification and the reliability of auxiliary diagnostic models. Summary of the Invention
[0005] This invention provides a deep learning-based MRI image-assisted diagnostic system for colorectal cancer, which solves the problem in the prior art where inaccurate localization of inner and outer edges is caused by factors such as overlap of signals between lesion edges and non-tumor tissues, blurred boundaries, and organ movement, making manual annotation prone to omissions and over-drawing errors, thus affecting the accuracy of network training and the reliability of model diagnosis.
[0006] The deep learning-based MRI-assisted diagnostic system for colorectal cancer of the present invention adopts the following technical solution: One embodiment of the present invention provides a deep learning-based MRI image-assisted diagnostic system for colorectal cancer, the system comprising the following modules: The acquisition module is used to acquire MRI images of colorectal cancer patients and the original lesion areas marked manually using the subtraction method. The correction module is used to determine the thickness distribution between the outer and inner edges of the original lesion area and identify areas with abnormal thickness, determine the morphological similarity between the outer and inner edges of the areas with abnormal thickness, and identify areas with abnormal thickness whose morphological similarity does not meet the preset conditions as areas to be corrected. The correction module is further configured to select a first outer edge sampling point in the outer edge line of the region to be corrected and determine its corresponding inner edge candidate range, and determine the real inner edge point in the first gradient peak point based on the number of first gradient peak points, gradient value, distance from the first outer edge sampling point and curvature of the line connecting the two nearest inner edge points in the inner edge candidate range; The correction module is further configured to select an inner edge sampling point in the outer edge line of the region to be corrected and determine its corresponding outer edge candidate range, and determine the true outer edge point in the second gradient peak point based on the number of second gradient peak points in the outer edge candidate range and the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point; The identification module is used to determine the actual lesion area based on the actual inner edge point and the actual outer edge point, and to use the actual lesion area as a deep learning label to construct an auxiliary diagnostic model, so as to identify colorectal cancer lesions through the auxiliary diagnostic model.
[0007] For example, the correction module achieves the determination of the thickness distribution between the outer and inner edges of the original lesion area and the identification of thickness abnormal areas by performing the following method: selecting multiple second outer edge sampling points of equal arc length along the circumferential direction on the outer edge line of the original lesion area; for each second outer edge sampling point, calculating the distance between the second outer edge sampling point and its nearest inner edge point along the normal direction, and recording it as the local thickness between the second outer edge sampling point and the corresponding nearest inner edge point; calculating the average value of the local thickness between all the second outer edge sampling points and the corresponding nearest inner edge point, recording it as the global average thickness, and determining a normal thickness range based on the global average thickness and a preset adjustment ratio; identifying the second outer edge sampling point and the nearest inner edge point corresponding to the local thickness exceeding the normal thickness range as abnormal points; connecting consecutively adjacent second outer edge sampling points among the abnormal points to obtain an abnormal outer edge line; connecting consecutively adjacent nearest inner edge points among the abnormal points to obtain an abnormal inner edge line; and identifying the area formed by the abnormal outer edge line and the abnormal inner edge line as the thickness abnormal area.
[0008] For example, the correction module achieves the determination of the morphological similarity between the outer and inner edges of the thickness anomaly region by executing the following method, and determines the thickness anomaly region whose morphological similarity does not meet the preset condition as the region to be corrected, including: for each thickness anomaly region, calculating the curvature of the abnormal outer edge line and the abnormal inner edge line respectively, determining the ratio of the smaller value to the larger value of the curvature, and recording it as curvature similarity; determining the similarity between the abnormal outer edge line and the abnormal inner edge line based on the dynamic time warping algorithm, and recording it as curve similarity; determining the morphological similarity between the abnormal outer edge line and the abnormal inner edge line based on the curvature similarity and the curve similarity, and determining the thickness anomaly region whose morphological similarity is less than a first preset threshold as the region to be corrected.
[0009] For example, the correction module implements the following method to select a first outer edge sampling point in the outer edge line of the region to be corrected and determine its corresponding inner edge candidate range, including: selecting multiple first outer edge sampling points of equal arc length along the circumferential direction of the outer edge line of the region to be corrected; for each first outer edge sampling point, determining the tangent of the outer edge line of the region to be corrected at the first outer edge sampling point, determining the normal of the first outer edge sampling point based on the tangent at the first outer edge sampling point, and obtaining the intersection point of the normal of the first outer edge sampling point and the inner edge line of the region to be corrected; taking the intersection point of the normal of the first outer edge sampling point and the inner edge line of the region to be corrected as the center, extending along the normal of the first outer edge sampling point by a preset number of pixels to form the inner edge candidate range corresponding to the first outer edge sampling point.
[0010] For example, the correction module implements the following method to determine the true inner edge point among the first gradient peak points based on the number of first gradient peak points in the inner edge candidate range, gradient value, distance to the first outer edge sampling point, and curvature of the line connecting the two nearest inner edge points, including: obtaining the gradient value of each pixel in the inner edge candidate range, and recording the pixel with the largest gradient value as the first gradient peak point; when there is only one first gradient peak point, recording the first gradient peak point as the true inner edge point corresponding to the first outer edge sampling point; when there is more than one first gradient peak point, determining the probability of the true inner edge point based on the gradient value of the first gradient peak point, the distance to the first outer edge sampling point, and the curvature of the line connecting the two nearest inner edge points, and determining the first gradient peak point with the largest probability of the true inner edge point as the true inner edge point.
[0011] For example, the correction module implements the determination of the true inner edge point probability based on the gradient value of the first gradient peak point, the distance to the first outer edge sampling point, and the curvature of the line connecting the two nearest inner edge points by performing the following method: selecting multiple second outer edge sampling points of equal arc length along the circumferential direction on the outer edge line of the original lesion region; calculating the average distance of each second outer edge sampling point to its nearest inner edge point along the normal direction, denoted as the global average thickness; calculating the absolute value of the difference between the distance between the first gradient peak point and the first outer edge sampling point and the global average thickness, denoted as the thickness offset factor; connecting the first gradient peak point and its two nearest inner edge points and calculating the curvature of the connecting line, denoted as the morphological curvature factor; calculating the product of the thickness offset factor and the morphological curvature factor, denoted as the thickness morphological comprehensive factor of the first gradient peak point; and denoting the ratio of the gradient value of the first gradient peak point to the thickness morphological comprehensive factor as the true inner edge point probability.
[0012] For example, the correction module selects inner edge sampling points and determines their corresponding outer edge candidate ranges within the outer edge line of the region to be corrected by executing the following method: selecting multiple inner edge sampling points of equal arc length along the circumferential direction of the inner edge line of the region to be corrected; for each inner edge sampling point, determining the tangent of the inner edge line of the region to be corrected at the inner edge sampling point, determining the normal of the inner edge sampling point based on the tangent at the inner edge sampling point, and obtaining the intersection point of the normal of the inner edge sampling point and the outer edge line of the region to be corrected; extending a preset number of pixels along the normal of the inner edge sampling point as the center, with the intersection point of the normal of the inner edge sampling point and the outer edge line of the region to be corrected as the center, to form the outer edge candidate range corresponding to the inner edge sampling point.
[0013] For example, the correction module determines the true outer edge point from the second gradient peak points based on the number of second gradient peak points in the outer edge candidate range and the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point by executing the following method, including: obtaining the gradient value of each pixel in the outer edge candidate range, and recording the pixel with the largest gradient value as the second gradient peak point; when there is only one second gradient peak point, recording the second gradient peak point as the true outer edge point corresponding to the inner edge sampling point; when there is more than one second gradient peak point, determining the probability of the true outer edge point based on the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point, and determining the second gradient peak point with the largest probability of the true outer edge point as the true outer edge point.
[0014] For example, the correction module implements the determination of the true outer edge point probability based on the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point by performing the following method, including: for the second gradient peak point, determining the adjacent sampling points on both sides of it within the outer edge candidate range along the normal direction of the inner edge sampling point; determining the gradient direction angle value of the adjacent sampling points and the normal direction angle value of the inner edge sampling point, and calculating the cosine value of the gradient direction angle value and the normal direction angle value; calculating the product of the cosine value of the gradient direction angle value and the normal direction angle value of the adjacent sampling points on both sides, and recording it as the true outer edge point probability.
[0015] For example, the identification module achieves the determination of the true lesion region based on the true inner edge points and the true outer edge points by performing the following method: performing curve smoothing processing on all the corrected true inner edge points to form a continuous corrected inner edge line, and splicing the corrected inner edge line with the normal inner edge line of the original lesion region, excluding the segment corresponding to the region to be corrected, to obtain a complete inner edge curve; performing curve smoothing processing on all the corrected true outer edge points to form a continuous corrected outer edge line, and splicing the corrected outer edge line with the normal outer edge line of the original lesion region, excluding the segment corresponding to the region to be corrected, to obtain a complete outer edge curve; calculating the difference region between the complete outer edge curve and the complete inner edge curve in the MRI image of the colorectal cancer patient to obtain the true lesion region after eliminating manual annotation errors.
[0016] The beneficial effects of the technical solution of the present invention are as follows: In the deep learning-based MRI image-assisted diagnostic system for colorectal cancer provided by this invention, the thickness distribution between the outer and inner edges of the manually annotated original lesion area is first determined by the aforementioned correction module. Based on this thickness distribution, abnormal thickness regions within the original lesion area are identified. Then, the morphological similarity between the inner and outer edges is used to filter out regions to be corrected from the abnormal thickness regions. This can improve the accuracy of locating errors caused by manual delineation and avoid misjudging real lesions as regions to be corrected. Outer edge sampling points are selected within the outer edge of the region to be corrected. Based on the number of first gradient peak points, gradient values, distances to the outer edge sampling points, and curvature of the lines connecting the corresponding inner edge candidate ranges, real inner edge points are filtered from the first gradient peak points. This can eliminate signal interference from non-tumor tissues such as residual fluid and feces in the intestinal lumen, improving the accuracy of lesion inner edge identification. Inner edge sampling points are selected within the inner edge of the region to be corrected, based on their corresponding outer edge sampling points... The number of second gradient peak points in the candidate range and the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point are used to screen true outer edge points from the second gradient peak points. This can filter out signal interference from surrounding blood vessels and lymph nodes, obtain a boundary that highly matches the true morphology of the lesion, and improve the accuracy of lesion outer edge identification. The true lesion area is determined by the true inner and outer edge points obtained by the correction module through the above identification module, and the auxiliary diagnostic model is trained with the true lesion area as the label. This solves the problem of training label distortion caused by manual annotation error, ensuring that the auxiliary diagnostic model can be trained based on accurate lesion features and avoiding model learning bias caused by label errors. Furthermore, the auxiliary diagnostic model built based on accurate true lesion area labels can learn colorectal cancer lesion features more accurately, thereby improving the reliability of the model's identification of colorectal cancer lesions and providing effective assistance for clinical diagnosis. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 The structural block diagram of the deep learning-based MRI image-assisted diagnostic system for colorectal cancer provided by this invention. Detailed Implementation
[0019] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the deep learning-based MRI image-assisted diagnostic system for colorectal cancer proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0021] The specific solution of the deep learning-based MRI image-assisted diagnostic system for colorectal cancer provided by the present invention will be described in detail below with reference to the accompanying drawings.
[0022] Please see Figure 1 The diagram illustrates a structural block diagram of a deep learning-based MRI image-assisted diagnostic system for colorectal cancer according to an embodiment of the present invention. The system includes the following modules: The acquisition module 110 is used to acquire MRI images of colorectal cancer patients and the original lesion areas marked manually using the subtraction method. The correction module 120 is used to determine the thickness distribution between the outer and inner edges of the original lesion area and identify areas with abnormal thickness, determine the morphological similarity between the outer and inner edges of areas with abnormal thickness, and identify areas with abnormal thickness whose morphological similarity does not meet the preset conditions as areas to be corrected. The correction module 120 is also used to select a first outer edge sampling point in the outer edge line of the region to be corrected and determine its corresponding inner edge candidate range. Based on the number of first gradient peak points, gradient value, distance from the first outer edge sampling point and curvature of the line connecting the nearest inner edge points on both sides in the inner edge candidate range, the actual inner edge point is determined in the first gradient peak points. The correction module 120 is also used to select inner edge sampling points in the outer edge line of the region to be corrected and determine the corresponding outer edge candidate range. Based on the number of second gradient peak points in the outer edge candidate range and the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point, the true outer edge point is determined in the second gradient peak points. The identification module 130 is used to determine the real lesion area based on the real inner edge point and the real outer edge point, and to use the real lesion area as a deep learning label to build an auxiliary diagnostic model, so as to identify colorectal cancer lesions through the auxiliary diagnostic model.
[0023] The functions of each module of the deep learning-based MRI image-assisted diagnostic system for colorectal cancer are described in detail below in one embodiment: The aforementioned acquisition module 110 is used to acquire MRI images of colorectal cancer patients and the original lesion areas marked manually using the subtraction method.
[0024] In this embodiment of the application, the MRI images of the colorectal cancer patients mentioned above are medical imaging data of the diseased area of the colorectal cancer patients acquired by magnetic resonance imaging (MRI) equipment; specifically, the image sequence mainly includes T2-weighted images (T2WI), diffusion-weighted images (DWI) and apparent diffusion coefficient maps (ADC), and must be quality assessed by a radiologist to exclude cases with significant motion artifacts or incomplete scans, so as to ensure the validity of the image data.
[0025] In the embodiments of this application, the above-mentioned subtraction method is a method for manually annotating the original lesion area in colorectal cancer MRI images. Specifically, the annotator (such as an experienced abdominal radiologist) uses medical image annotation software to obtain the true shape and thickness of the lesion by step-by-step delineation and difference operation, based on the anatomical and pathological characteristics of tumor lesions that often grow invasively along the intestinal wall.
[0026] In one specific implementation of this application, the process of determining the original lesion area by subtraction can be achieved as follows: on the MRI image of a colorectal cancer patient, the outer edge of the tumor is delineated along the boundary between the tumor and the surrounding normal tissue to determine the maximum spatial range of the tumor; the inner edge of the tumor is delineated along the boundary between the tumor and the non-tumor tissue inside the intestinal lumen to distinguish the tumor parenchyma from the contents of the intestinal lumen; the difference between the outer edge delineated area and the inner edge delineated area is calculated using the difference operation function of the medical image annotation software ITK-SNAP to obtain the original lesion area reflecting the actual thickness, annular infiltration range and spatial morphology of the tumor.
[0027] The colorectal cancer MRI image-assisted diagnostic system provided in this application, after acquiring the MRI images of a colorectal cancer patient and the manually annotated original lesion area through the acquisition module, can locate the manually annotated error area to be corrected through the correction module, and correct the manually drawn lesion area annotation information in the area to be corrected to determine the true lesion area. The function of the correction module 120 is described in detail below: The aforementioned correction module 120 is used to determine the thickness distribution between the outer and inner edges of the original lesion area and identify areas with abnormal thickness, determine the morphological similarity between the outer and inner edges of areas with abnormal thickness, and identify areas with abnormal thickness whose morphological similarity does not meet the preset conditions as areas to be corrected.
[0028] In this embodiment of the application, the original lesion area refers to the lesion area marked by the subtraction method in the acquisition module; its outer edge is the boundary line between the tumor and the surrounding normal tissue in the area, which is manually delineated; its inner edge is the boundary line between the tumor and the non-tumor tissue inside the intestinal lumen in the area, which is manually delineated.
[0029] The wall thickness of the colon and rectum varies in different anatomical locations, but for the same segment of the intestinal wall, the distance between its inner and outer edges usually exhibits a continuous and smooth trend. However, since the intestinal wall is distributed along the circumferential direction, when manually delineating the inner and outer edges of a lesion using the subtraction method, if there are annotation deviations such as non-parallelism, misalignment, or local overlap of the inner and outer edges, the continuous and smooth characteristic of the thickness of the same segment of the intestinal wall will be disrupted, leading to abnormal increases in local thickness (such as overdrawing the outer edge and omitting the inner edge) or abnormal decreases in thickness (such as omitting the outer edge and overdrawing the inner edge). Therefore, the embodiments of this application can preliminarily determine the area of manual annotation error by identifying abrupt changes or abnormal fluctuations in thickness between the outer and outer edges within the original lesion area.
[0030] In this embodiment of the application, the aforementioned thickness distribution refers to the change in local thickness between the outer and inner edges of the original lesion area in the circumferential direction, reflecting the spacing characteristics of the outer and inner edges of the original lesion area at different circumferential positions, and is the basis for judging whether there is over-drawing or omission in manual annotation; the aforementioned abnormal thickness area is the manual annotation error area initially determined by identifying the abrupt change or abnormal fluctuation in thickness between the outer and inner edges of the original lesion area.
[0031] For example, the above-mentioned determination of the thickness distribution between the outer and inner edges of the original lesion area and identification of thickness abnormality areas can be achieved as follows: Select multiple second outer edge sampling points of equal arc length along the circumferential direction on the outer edge line of the original lesion area; For each second outer edge sampling point, calculate the distance between the second outer edge sampling point and its nearest inner edge point along the normal direction, and record it as the local thickness between the second outer edge sampling point and the corresponding nearest inner edge point; Calculate the average value of the local thickness between all second outer edge sampling points and their corresponding nearest inner edge points, and record it as the global average thickness; Determine the normal thickness range based on the global average thickness and a preset adjustment ratio; Identify the second outer edge sampling points and the nearest inner edge point corresponding to the local thickness exceeding the normal thickness range as abnormal points; Connect consecutively adjacent second outer edge sampling points among the abnormal points to obtain the abnormal outer edge line; Connect consecutively adjacent nearest inner edge points among the abnormal points to obtain the abnormal inner edge line; Identify the area formed by the abnormal outer edge line and the abnormal inner edge line as the thickness abnormality area.
[0032] The process of identifying regions with abnormal thickness described above is explained in detail below in a specific embodiment: The MRI image of a colorectal cancer patient, with the original lesion region manually annotated, is received from the acquisition module, and the inner and outer edges of the original lesion region are extracted; N second outer edge sampling points are selected along the outer edge of the original lesion region according to the principle of equal arc length in the circumferential direction, and denoted as the set of second outer edge sampling points. The value of N must ensure that the sampling points are evenly distributed and cover the entire outer circumference to avoid errors in thickness calculation caused by missing sampling in local areas; for any second outer edge sampling point Determine the normal direction of the lesion at the outer edge of the original lesion area (perpendicular to the outer edge at the second outer edge sampling point). (The tangent direction at the location points towards the inner edge); calculate the sampling point at the second outer edge. Along its normal direction to the nearest inner edge point straight-line distance ,Will This is recorded as the second outer edge sampling point. Its nearest inner edge point The local thickness between the N second outer edge sampling points and their nearest inner edge points is determined using the method described above, and a thickness distribution set is constructed. Calculate the average value of all local thicknesses in the thickness distribution set, and denote it as the global average thickness. This serves as a benchmark for determining whether the thickness is abnormal; based on the global average thickness. The normal thickness range is determined by a preset adjustment ratio (set here to ±10%, to be fine-tuned based on clinical experience or statistical results) as [0.9]. 1.1 This normal thickness range is set based on the rectal anatomical feature that the distance between the inner and outer edges of the same segment of the intestinal wall changes continuously and smoothly, covering reasonable thickness fluctuations caused by normal physiology and actual pathological conditions; for the second outer edge sampling point Determine its relationship with the nearest inner edge point. Local thickness Does it meet the following criteria: If the above criteria are met, it indicates that the area is either too thick or too thin, and the second outer edge sampling point will be moved to the next location. and its corresponding nearest inner edge point Points identified as outliers are denoted as the outlier set. Connecting consecutive adjacent second outer edge sampling points forms an abnormal outer edge line, and connecting consecutive adjacent nearest inner edge points forms an abnormal inner edge line. The region formed by the abnormal outer edge line and the abnormal inner edge line is defined as the thickness anomaly region. For example, if there exists a set of abnormal points... , Then The resulting colorectal region was identified as an area of abnormal thickness.
[0033] Since deep tumor infiltration or local tissue compression can also lead to increased intestinal wall thickness, this embodiment of the application, after identifying the abnormal thickness region using the above method, needs to further distinguish whether the abnormal thickness region is caused by manual annotation error or by a real lesion. Specifically, the colorectum is a continuous and smooth tubular structure, and its peripheral outer edge morphology has strong geometric consistency. Therefore, the region to be corrected within the abnormal thickness region can be screened by combining the similarity characteristics of the outer and outer edge morphology within the abnormal thickness region.
[0034] For example, the above-mentioned determination of the morphological similarity of the outer and inner edges of the thickness anomaly region, and the determination of the thickness anomaly region whose morphological similarity does not meet the preset condition as the region to be corrected, can be achieved as follows: For each thickness anomaly region, the curvature of the abnormal outer edge line and the abnormal inner edge line are calculated respectively, and the ratio of the smaller value to the larger value of the curvature is determined and recorded as the curvature similarity; the similarity between the abnormal outer edge line and the abnormal inner edge line is determined based on the dynamic time warping algorithm and recorded as the curve similarity; the morphological similarity between the abnormal outer edge line and the abnormal inner edge line is determined based on the curvature similarity and the curve similarity, and the thickness anomaly region whose morphological similarity is less than the first preset threshold is determined as the region to be corrected.
[0035] The process of determining the region to be corrected described above will be explained in detail below in a specific embodiment: Extracting the abnormal outer edge line of the thickness abnormal region. and abnormal inner edge line Calculate the outer edge of the anomaly respectively. curvature and abnormal inner edge line curvature The curvature similarity is calculated using the following formula: Curvature similarity = in, It is the smaller value between the curvature of the abnormal outer edge line and the curvature of the abnormal inner edge line of the thickness anomaly region. The value is the larger of the curvature of the abnormal outer edge and the curvature of the abnormal inner edge of the region with abnormal thickness. As a continuous and smooth tubular organ, the colorectal wall and the inner and outer edges of the normal intestinal wall or the intestinal wall of the real lesion area need to maintain a synchronous curvature trend due to the constraints of the anatomical structure. However, manual annotation errors can cause the curvature trends of the inner and outer edges to become disconnected. Therefore, the closer the curvature similarity calculated by the above formula is to 1, the greater the possibility that the corresponding region with abnormal thickness is a real lesion.
[0036] After determining the curvature similarity, the curve similarity of the outer and inner edges of the thickness anomaly region is determined using the Dynamic Time Warping (DTW) algorithm, and the above morphological similarity is determined using the following formula: in, The morphological similarity between the outer and inner edges of the thickness anomaly region; The curvature similarity of the inner and outer edges within the thickness anomaly region is used to reflect the degree of consistency between the bending trends of the inner and outer edges. The similarity of the curves of the inner and outer edges within the region of abnormal thickness is used to quantify the degree of matching between the overall morphology of the inner and outer edges. The formula above combines the local curvature trend of the inner and outer edges (the same segment of the intestinal wall will not have a gently curving inner edge and a steep outer edge) and the overall morphology (the inner and outer edges must extend synchronously along the circumferential direction of the intestinal wall without obvious discontinuity) to screen regions to be corrected from the region of abnormal thickness. Specifically, taking the first preset threshold of 0.9 as an example, if the morphological similarity is greater than 0.9, the corresponding region of abnormal thickness is determined to be a true lesion region and no correction is needed; if the morphological similarity is less than or equal to 0.9, the corresponding region of abnormal thickness is determined to be a region to be corrected.
[0037] Because the grayscale distribution of the colorectum differs significantly from that of surrounding tissues, the peak position of its grayscale gradient usually corresponds to the actual lesion boundary. Therefore, this embodiment can update the inner and outer boundaries of the region to be corrected based on the gradient peak. Specifically, the correction function of the above correction module is described as follows: The correction module 120 is also used to select a first outer edge sampling point in the outer edge line of the region to be corrected and determine its corresponding inner edge candidate range. Based on the number of first gradient peak points, gradient value, distance from the first outer edge sampling point and curvature of the line connecting the nearest inner edge points on both sides in the inner edge candidate range, the actual inner edge point is determined in the first gradient peak points.
[0038] In this embodiment of the application, the first outer edge sampling point is a sampling point selected in the circumferential direction of the outer edge line of the area to be corrected based on the principle of equal arc length spacing, for locating the inner edge boundary.
[0039] In this embodiment, the aforementioned inner edge candidate range is a pixel interval formed by extending a preset number of pixels along the normal of the first outer edge sampling point, centered on the intersection of the normal of the first outer edge sampling point and the original inner edge line. The function of this inner edge candidate range is to reduce the search range of the real inner edge point, eliminate interference from irrelevant areas such as intestinal residual fluid and background tissue, and improve the gradient peak recognition efficiency.
[0040] For example, the above-mentioned selection of a first outer edge sampling point in the outer edge line of the region to be corrected and determination of its corresponding inner edge candidate range can be achieved as follows: Select multiple first outer edge sampling points with equal arc lengths along the circumferential direction of the outer edge line of the region to be corrected; For each first outer edge sampling point, determine the tangent of the outer edge line of the region to be corrected at the first outer edge sampling point, determine the normal of the first outer edge sampling point based on the tangent at the first outer edge sampling point, and obtain the intersection point of the normal of the first outer edge sampling point and the inner edge line of the region to be corrected; Taking the intersection point of the normal of the first outer edge sampling point and the inner edge line of the region to be corrected as the center, extend along the normal of the first outer edge sampling point by a preset number of pixels to form the inner edge candidate range corresponding to the first outer edge sampling point.
[0041] In this embodiment, the first gradient peak point is the pixel with the largest gray-scale gradient value within the inner edge candidate range. Since the gray-scale difference between colorectal tumor tissue and non-tumor tissue in the intestinal lumen is significant, the first gradient peak point theoretically corresponds to the true inner edge boundary. However, due to local folding and artifacts, there may be multiple first gradient peak points, and further screening is needed to determine the true inner edge point.
[0042] In this embodiment, the aforementioned true inner edge points are the pixels in the region to be corrected that best match the true boundary between the colorectal tumor and the non-tumor tissue inside the intestinal lumen. They are the basis for correcting inner edge annotation errors and constructing accurate inner edge curves.
[0043] For example, the above method of determining the true inner edge point from the first gradient peak points based on the number of first gradient peak points in the inner edge candidate range, the gradient value, the distance to the first outer edge sampling point, and the curvature of the line connecting the two nearest inner edge points can be implemented as follows: obtain the gradient value of each pixel in the inner edge candidate range, and record the pixel with the largest gradient value as the first gradient peak point; when there is only one first gradient peak point, record the first gradient peak point as the true inner edge point corresponding to the first outer edge sampling point; when there is more than one first gradient peak point, determine the probability of the true inner edge point based on the gradient value of the first gradient peak point, the distance to the first outer edge sampling point, and the curvature of the line connecting the two nearest inner edge points, and determine the first gradient peak point with the largest probability of the true inner edge point as the true inner edge point.
[0044] When there is more than one first gradient peak point, the above method of determining the probability of the true inner edge point based on the gradient value of the first gradient peak point, the distance to the first outer edge sampling point, and the curvature of the line connecting the two nearest inner edge points can be implemented as follows: Select multiple second outer edge sampling points of equal arc length along the circumferential direction on the outer edge line of the original lesion area; calculate the average distance between each second outer edge sampling point and its nearest inner edge point along the normal direction, and record it as the global average thickness; calculate the absolute value of the difference between the distance between the first gradient peak point and the first outer edge sampling point and the global average thickness, and record it as the thickness offset factor; connect the first gradient peak point and its two nearest inner edge points and calculate the curvature of the connecting line, and record it as the morphological curvature factor; calculate the product of the thickness offset factor and the morphological curvature factor, and record it as the thickness morphological comprehensive factor of the first gradient peak point; record the ratio of the gradient value of the first gradient peak point to the thickness morphological comprehensive factor as the probability of the true inner edge point.
[0045] The process of determining the true inner edge point described above will be explained in detail below in a specific embodiment: For any region to be corrected, its abnormal outer edge line is extracted. and abnormal inner edge line Following the principle of equal arc length along the outer edge of the anomaly in the area to be corrected Multiple first outer edge sampling points are selected in the circumferential direction to construct a first outer edge sampling point set. The first outer edge sampling point selected should be evenly distributed along the outer edge of the abnormality. Above, and the normal direction of each sampling point must point to the inner edge, providing a directional reference for subsequently determining the candidate range of the inner edge; for any first outer edge sampling point Determine the outer edge of the abnormality First outer edge sampling point The tangent at the point is used as the first outer edge sampling point, with the direction perpendicular to the tangent and pointing towards the inner edge. The normal direction; determine the first outer edge sampling point. The intersection of the normal and the manually labeled original inner edge line ; with this intersection point Centered on the image, extend M adjacent pixels along the normal direction (the range can be determined based on the image noise level) to obtain the aforementioned candidate inner edge range. ; Calculate the inner edge candidate range The grayscale gradient values of each pixel within the set are used to obtain the grayscale gradient value set. And select the pixel with the largest gray-level gradient value from the set of gray-level gradient values, and record it as the first gradient peak point; if the inner edge candidate range If a unique first gradient peak point exists within the inner edge, then that first gradient peak point is determined as the true inner edge point; if the inner edge candidate range is... There are multiple first gradient peaks in memory. For any first gradient peak... Calculate its relationship with the corresponding first outer edge sampling point Distance between The global average thickness determined during the process of identifying thickness anomaly regions is obtained. The first gradient peak point is calculated using the following formula. Thickness offset factor: ; Obtain the peak point located at the first gradient Connect the nearest inner edge points on both sides to the first gradient peak point. Find the nearest inner edge points on both sides and calculate the curvature of the line connecting them. Let be the curvature factor of the aforementioned shape; the probability of the aforementioned true inner edge point is calculated using the following formula: in, The probability of the true inner edge point is given above; the larger the value, the higher the corresponding first gradient peak point. The greater the probability that it is a true inner edge point; The first gradient peak point The larger the grayscale gradient value, the higher the value reflects the peak point of the first gradient. The more dramatic the grayscale change, the more likely it is to be a true tissue boundary (there is a significant difference in grayscale distribution between colorectal cancer tumors and non-tumor tissues inside the intestinal lumen). The first gradient peak point The thickness offset factor, the smaller its value, the closer it is to the first gradient peak point. The closer it conforms to the normal thickness characteristics of the intestinal wall (as a continuous and smooth tubular organ, the distance between the inner and outer edges of the same segment of the intestinal wall should show a continuous and gradual change trend, without sudden thickening or thinning). The first gradient peak point The morphological curvature factor, the smaller the value, the higher the peak value of the first gradient. The flatter the inner edge curve, the more it conforms to the true morphology of the intestinal wall (the inner edge of the true colorectal wall is a continuous and smooth arc, without sudden sharp angles or folds); after determining the true inner edge point probability of each first gradient peak point through the above method, the first gradient peak point with the highest true inner edge point probability is determined as the true inner edge point.
[0046] The error in manually annotating the outer edge mostly stems from high-signal artifacts in surrounding blood vessels and lymph nodes. Since the gradient direction of these artifacts is inconsistent with the anatomical direction of the intestinal wall (the normal direction of the inner edge sampling point), this embodiment of the application can eliminate artifacts by ensuring the consistency between the normal direction of the inner edge sampling point and the gradient direction of the sampling points adjacent to the second gradient peak point, thereby correcting the outer edge annotation error. Specifically, this is achieved through the correction module 120 as follows: The correction module 120 is also used to select inner edge sampling points in the outer edge line of the region to be corrected and determine the corresponding outer edge candidate range. Based on the number of second gradient peak points in the outer edge candidate range and the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point, the true outer edge point is determined in the second gradient peak points.
[0047] In this embodiment of the application, the aforementioned inner edge sampling points are sampling points selected in the circumferential direction of the inner edge line of the area to be corrected, based on the principle of equal arc length spacing, for locating the outer edge boundary.
[0048] In this embodiment, the aforementioned outer edge candidate range is a pixel interval formed by extending a preset number of pixels along the normal of the inner edge sampling point, centered on the intersection of the normal of the inner edge sampling point and the original outer edge line. The function of this outer edge candidate range is to reduce the search range of the real outer edge point and exclude interference from irrelevant areas such as surrounding blood vessels, lymph nodes, and adipose tissue.
[0049] For example, the above-mentioned selection of inner edge sampling points in the inner edge line of the region to be corrected and determination of their corresponding outer edge candidate range can be achieved as follows: Select multiple inner edge sampling points with equal arc lengths along the circumferential direction of the inner edge line of the region to be corrected; For each inner edge sampling point, determine the tangent of the inner edge line of the region to be corrected at the inner edge sampling point, determine the normal of the inner edge sampling point based on the tangent at the inner edge sampling point, and obtain the intersection of the normal of the inner edge sampling point and the outer edge line of the region to be corrected; Taking the intersection of the normal of the inner edge sampling point and the outer edge line of the region to be corrected as the center, extend along the normal of the inner edge sampling point by a preset number of pixels (the preset number of pixels is usually 5-8, which can be adjusted according to the resolution of the MRI image) to form the outer edge candidate range corresponding to the inner edge sampling point.
[0050] In this embodiment, the second gradient peak point is the pixel with the largest gray gradient value in the candidate range of the outer edge. Since there is a significant difference in gray distribution between colorectal tumor tissue and surrounding normal tissue (such as blood vessels and lymph nodes), the second gradient peak point theoretically corresponds to the real outer edge boundary. However, due to the high signal artifacts of blood vessels and lymph nodes, there may be multiple peak points, and further screening is needed to determine the real outer edge point.
[0051] In this embodiment, the aforementioned true outer edge points are the pixels in the region to be corrected that best match the true boundary between the colorectal tumor and the surrounding normal tissue. They are the basis for correcting the outer edge annotation error and constructing an accurate outer edge curve.
[0052] For example, the above method of determining the true outer edge point from the second gradient peak points based on the number of second gradient peak points in the outer edge candidate range and the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point can be implemented as follows: obtain the gradient value of each pixel in the outer edge candidate range, and record the pixel with the largest gradient value as the second gradient peak point; when there is only one second gradient peak point, record the second gradient peak point as the true outer edge point corresponding to the inner edge sampling point; when there is more than one second gradient peak point, determine the probability of the true outer edge point based on the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point, and determine the second gradient peak point with the highest probability of the true outer edge point as the true outer edge point.
[0053] When there is more than one second gradient peak point, the above method for determining the probability of the true outer edge point based on the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point can be implemented as follows: For the second gradient peak point, determine the adjacent sampling points on both sides of it within the candidate range of the outer edge along the normal direction of the inner edge sampling point; determine the gradient direction angle value of the adjacent sampling points and the normal direction angle value of the inner edge sampling point, and calculate the cosine value of the gradient direction angle value and the normal direction angle value; calculate the product of the gradient direction angle value and the cosine value of the normal direction angle value of the adjacent sampling points on both sides, and denot it as the probability of the true outer edge point.
[0054] The process of determining the true outer edge point described above will be explained in detail below in a specific embodiment: Based on the above-described method of determining the inner edge candidate range... Similarly, the method constructs the inner edge sampling point set. And determine any inner edge sampling point Corresponding outer edge alternative range ; Calculate the alternative range of the outer edge The grayscale gradient values of each pixel within the inner edge are calculated, and the pixel with the largest grayscale gradient value is selected and recorded as the second gradient peak point; if the outer edge is a candidate range... If a unique second gradient peak point exists within the memory, then that second gradient peak point is determined as the true outer edge point; if the outer edge candidate range is... There are multiple second gradient peaks in memory, and for any second gradient peak... Within the candidate outer edge range Sampling points along the inner edge of the area The normal direction is selected from the adjacent pixels on both sides. and As its adjacent sampling points, calculate its gradient direction angle value, and denot it as follows: The probability of the true outer edge point is calculated using the following formula: in, The higher the probability of the true outer edge point mentioned above, the higher the corresponding second gradient peak point. The greater the probability that it is a true outer edge point; The second gradient peak point adjacent sampling points The gradient direction angle value; The second gradient peak point adjacent sampling points The gradient direction angle value; Inner edge sampling point The angle value of the normal direction; adjacent sampling points gradient direction angle value With inner edge sampling points normal direction angle value The closer the cosine value is to 1, the better it reflects the gradient direction angle. Angle value with normal direction The more consistent the directions; adjacent sampling points gradient direction angle value With inner edge sampling points normal direction angle value The closer the cosine value is to 1, the better it reflects the gradient direction angle. Angle value with normal direction The more consistent the directions; the true outer boundary is due to the tubular structure of the intestinal wall, the more consistent the directions of the adjacent sampling points mentioned above. and The gradient direction angle value must be related to the normal direction of the inner edge sampling point. The directions of adjacent sampling points on both sides are consistent, while the gradient directions of artifacts are irregular. The above formula uses the product of the gradient direction angle value and the cosine of the normal direction angle value of adjacent sampling points on both sides as the probability of the true outer edge point. Select The second gradient peak point with the largest value Inner edge sampling point The corresponding actual outer edge point.
[0055] The aforementioned identification module 130 is used to determine the actual lesion area based on the actual inner edge point and the actual outer edge point, and to use the actual lesion area as a deep learning label to construct an auxiliary diagnostic model, so as to identify colorectal cancer lesions through the auxiliary diagnostic model.
[0056] In the embodiments of this application, the aforementioned actual lesion area is the area that reflects the actual morphology and extent of colorectal cancer tumors after excluding errors in manual annotation.
[0057] For example, the above-mentioned determination of the true lesion region based on true inner and outer edge points can be achieved as follows: All corrected true inner edge points are smoothed to eliminate local noise or discontinuities, forming a continuous corrected inner edge line. This corrected inner edge line is then spliced with the normal inner edge line of the original lesion region, excluding the segment corresponding to the area to be corrected, to obtain a complete inner edge curve. Similarly, all corrected true outer edge points are smoothed to eliminate local noise or discontinuities, forming a continuous corrected outer edge line. This corrected outer edge line is then spliced with the normal outer edge line of the original lesion region, excluding the segment corresponding to the area to be corrected, to obtain a complete outer edge curve. The difference region between the complete outer edge curve and the complete inner edge curve in the MRI images of colorectal cancer patients is calculated to obtain the true lesion region after eliminating manual annotation errors.
[0058] In the embodiments of this application, the above-mentioned auxiliary diagnostic model is a model constructed based on a convolutional neural network, which takes MRI images of colorectal cancer patients as input and outputs lesion region segmentation results, and can realize the automatic identification and range determination of colorectal cancer lesions.
[0059] After obtaining the true lesion region after eliminating manual annotation errors through the above process, this embodiment of the application uses the true lesion region as a training label to construct an MRI image-assisted diagnostic model based on a convolutional neural network. By inputting the patient's rectal MRI sequence, the auxiliary diagnostic model can automatically learn the spatial distribution characteristics, gray-scale variation patterns, and edge morphological features of the tumor, thereby achieving automatic identification and segmentation of the lesion region. Specifically, 70% of the training label dataset was used as the training set, 15% as the validation set, and 15% as the test set. An improved U-Shaped Convolutional Neural Network (U-Net) was adopted, with a joint loss function of Dice Loss and Binary Cross-Entropy (BCE). The optimizer used was the Adaptive Moment Estimation (Adam). During training, the network utilized multi-scale convolutions to extract local and global features, combined with feature fusion and context information enhancement modules, to improve the ability to identify complex boundaries. This enabled the model to accurately identify the morphology and extent of rectal cancer lesions on multiple cross-sections, providing doctors with reliable auxiliary diagnostic information and improving the accuracy and consistency of lesion detection.
[0060] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A deep learning-based MRI image-assisted diagnostic system for colorectal cancer, characterized in that, The system includes: The acquisition module is used to acquire MRI images of colorectal cancer patients and the original lesion areas marked manually using the subtraction method. The correction module is used to determine the thickness distribution between the outer and inner edges of the original lesion area and identify areas with abnormal thickness, determine the morphological similarity between the outer and inner edges of the areas with abnormal thickness, and identify areas with abnormal thickness whose morphological similarity does not meet the preset conditions as areas to be corrected. The correction module is further configured to select a first outer edge sampling point in the outer edge line of the region to be corrected and determine its corresponding inner edge candidate range, and determine the real inner edge point in the first gradient peak point based on the number of first gradient peak points, gradient value, distance from the first outer edge sampling point and curvature of the line connecting the two nearest inner edge points in the inner edge candidate range; The correction module is further configured to select an inner edge sampling point in the outer edge line of the region to be corrected and determine its corresponding outer edge candidate range, and determine the true outer edge point in the second gradient peak point based on the number of second gradient peak points in the outer edge candidate range and the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point; The identification module is used to determine the actual lesion area based on the actual inner edge point and the actual outer edge point, and to use the actual lesion area as a deep learning label to construct an auxiliary diagnostic model, so as to identify colorectal cancer lesions through the auxiliary diagnostic model.
2. The deep learning-based MRI image-assisted diagnostic system for colorectal cancer according to claim 1, characterized in that, The correction module achieves the determination of the thickness distribution between the outer and inner edges of the original lesion area and identifies areas of abnormal thickness by performing the following methods: Multiple second outer edge sampling points of equal arc length are selected along the circumferential direction on the outer edge line of the original lesion area; For each of the second outer edge sampling points, calculate the distance between the second outer edge sampling point and its nearest inner edge point along the normal direction, and record it as the local thickness of the second outer edge sampling point and the corresponding nearest inner edge point; Calculate the average value of the local thickness of all the second outer edge sampling points and the corresponding nearest inner edge point, and record it as the global average thickness. Based on the global average thickness and the preset adjustment ratio, determine the normal thickness range. The second outer edge sampling points and the nearest inner edge points corresponding to the local thickness that exceeds the normal thickness range are determined as abnormal points. Connect the second outer edge sampling points that are consecutively adjacent among the abnormal points to obtain the abnormal outer edge line, and connect the nearest inner edge points that are consecutively adjacent among the abnormal points to obtain the abnormal inner edge line. The region formed by the abnormal outer edge line and the abnormal inner edge line is identified as the thickness abnormal region.
3. The deep learning-based MRI image-assisted diagnostic system for colorectal cancer according to claim 2, characterized in that, The correction module determines the morphological similarity of the outer and inner edges of the thickness anomaly region by executing the following method, and identifies the thickness anomaly region whose morphological similarity does not meet the preset conditions as the region to be corrected, including: For each of the thickness anomalous regions, the curvature of the outer edge line and the inner edge line of the anomalous region are calculated respectively, and the ratio of the smaller value to the larger value of the curvature is determined and denoted as the curvature similarity. The similarity between the outer edge of the anomaly and the inner edge of the anomaly is determined based on the dynamic time warping algorithm and denoted as curve similarity. Based on the curvature similarity and the curve similarity, the morphological similarity between the abnormal outer edge line and the abnormal inner edge line is determined, and the thickness abnormal region with the morphological similarity less than a first preset threshold is determined as the region to be corrected.
4. The deep learning-based MRI image-assisted diagnostic system for colorectal cancer according to claim 1, characterized in that, The correction module selects a first outer edge sampling point in the outer edge of the region to be corrected and determines its corresponding inner edge candidate range by executing the following method: Multiple first outer edge sampling points of equal arc length are selected along the circumferential direction of the outer edge line of the region to be corrected; For each of the first outer edge sampling points, the tangent line of the outer edge line of the region to be corrected at the first outer edge sampling point is determined, the normal line of the first outer edge sampling point is determined based on the tangent line at the first outer edge sampling point, and the intersection point of the normal line of the first outer edge sampling point and the outer edge line of the region to be corrected is obtained. Taking the intersection of the normal of the first outer edge sampling point and the inner edge line of the area to be corrected as the center, a preset number of pixels are extended along the normal of the first outer edge sampling point to form the inner edge candidate range corresponding to the first outer edge sampling point.
5. The deep learning-based MRI image-assisted diagnostic system for colorectal cancer according to claim 4, characterized in that, The correction module determines the true inner edge point from the first gradient peak points by executing the following method: based on the number of first gradient peak points in the inner edge candidate range, the gradient value, the distance to the first outer edge sampling point, and the curvature of the line connecting the two nearest inner edge points. Obtain the gradient value of each pixel within the inner edge candidate range, and record the pixel with the largest gradient value as the first gradient peak point; When there is one and only one first gradient peak point, the first gradient peak point is recorded as the real inner edge point corresponding to the first outer edge sampling point; When there is more than one first gradient peak point, the probability of the true inner edge point is determined based on the gradient value of the first gradient peak point, the distance to the first outer edge sampling point, and the curvature of the line connecting the two nearest inner edge points. The first gradient peak point with the highest probability of the true inner edge point is determined as the true inner edge point.
6. The deep learning-based MRI image-assisted diagnostic system for colorectal cancer according to claim 5, characterized in that, The correction module determines the probability of the true inner edge point based on the gradient value of the first gradient peak point, the distance to the first outer edge sampling point, and the curvature of the line connecting the two nearest inner edge points by executing the following method: Multiple second outer edge sampling points of equal arc length are selected along the circumferential direction on the outer edge line of the original lesion area. The average distance between each second outer edge sampling point and its nearest inner edge point along the normal direction is calculated and recorded as the global average thickness. Calculate the absolute value of the difference between the distance between the first gradient peak point and the first outer edge sampling point and the global average thickness, and denot it as the thickness offset factor; Connect the first gradient peak point and its nearest inner edge points on both sides and calculate the curvature of the connecting line, which is denoted as the shape curvature factor. Calculate the product of the thickness offset factor and the shape curvature factor, and denot it as the thickness shape comprehensive factor at the first gradient peak point; The ratio of the gradient value of the first gradient peak point to the thickness morphology synthesis factor is recorded as the probability of the true inner edge point.
7. The deep learning-based MRI image-assisted diagnostic system for colorectal cancer according to claim 1, characterized in that, The correction module selects inner edge sampling points and determines their corresponding outer edge candidate ranges within the outer edge line of the region to be corrected by executing the following method: Multiple inner edge sampling points of equal arc length are selected along the circumferential direction of the inner edge line of the region to be corrected; For each of the inner edge sampling points, the tangent line of the outer edge line of the region to be corrected at the inner edge sampling point is determined, the normal line of the inner edge sampling point is determined based on the tangent line at the inner edge sampling point, and the intersection point of the normal line of the inner edge sampling point and the outer edge line of the region to be corrected is obtained. Taking the intersection of the normal of the inner edge sampling point and the outer edge of the region to be corrected as the center, a preset number of pixels are extended along the normal of the inner edge sampling point to form the outer edge candidate range corresponding to the inner edge sampling point.
8. The deep learning-based MRI image-assisted diagnostic system for colorectal cancer according to claim 7, characterized in that, The correction module determines the true outer edge point from the second gradient peak points by executing the following method: based on the number of second gradient peak points in the outer edge candidate range and the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point. Obtain the gradient value of each pixel within the outer edge candidate range, and record the pixel with the largest gradient value as the second gradient peak point; When there is one and only one second gradient peak point, the second gradient peak point is recorded as the real outer edge point corresponding to the inner edge sampling point; When there is more than one second gradient peak point, the probability of the true outer edge point is determined based on the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point, and the second gradient peak point with the highest probability of the true outer edge point is determined as the true outer edge point.
9. The deep learning-based MRI image-assisted diagnostic system for colorectal cancer according to claim 8, characterized in that, The correction module achieves the determination of the true outer edge point probability based on the consistency between the normal direction of the inner edge sampling point and the gradient direction of the adjacent sampling points of the second gradient peak point by executing the following method, including: For the second gradient peak point, the adjacent sampling points on both sides of it are determined within the outer edge candidate range along the normal direction of the inner edge sampling point; Determine the gradient direction angle value of the adjacent sampling points and the normal direction angle value of the inner edge sampling points, and calculate the cosine value of the gradient direction angle value and the normal direction angle value; The product of the gradient direction angle value and the cosine of the normal direction angle value of the adjacent sampling points on both sides is calculated and denoted as the true outer edge point probability.
10. The deep learning-based MRI image-assisted diagnostic system for colorectal cancer according to claim 1, characterized in that, The identification module determines the actual lesion region based on the actual inner edge point and the actual outer edge point by executing the following method: The corrected true inner edge points are smoothed to form a continuous corrected inner edge line. The corrected inner edge line is then spliced with the normal inner edge line of the original lesion area, excluding the segment corresponding to the area to be corrected, to obtain a complete inner edge curve. The corrected true outer edge points are smoothed to form a continuous corrected outer edge line. The corrected outer edge line is then spliced with the normal outer edge line of the original lesion area, excluding the segment corresponding to the area to be corrected, to obtain a complete outer edge curve. The difference region between the complete outer edge curve and the complete inner edge curve in the MRI images of the colorectal cancer patient is calculated to obtain the true lesion region after removing manual annotation errors.