Colorectal cancer MRI image segmentation method based on deep learning

Abstract

The invention discloses a colorectal cancer MRI (Magnetic Resonance Imaging) image segmentation method based on deep learning, particularly relates to the technical field of medical image processing, and is used for solving the problem of inconsistent development quality of key anatomical structures in existing rectal cancer image evaluation. Scanning parameters are adjusted based on a standardized protocol containing a T2WI fat-free suppression sequence, DWI and IVIM imaging, registration errors of a functional sequence are dynamically calibrated in combination with fascia layer edge continuity parameters, multi-modal features are fused through adaptive weight distribution, and a deep learning segmentation model is constructed to generate a minimum infiltration distance segmentation mask from a tumor to a fascia layer; further analyzing perfusion gradient characteristics of a tumor-fascia transition area to position a micro-infiltration area, and finally comparing a quantitative result with a clinical staging standard to generate a structured report containing TNM parameters and fascia invasion states; through multi-dimensional collaborative analysis of anatomy and functional images, a high-confidence-coefficient three-dimensional quantitative basis is provided for rectal cancer surgery incision edge planning.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and more specifically, to a method for segmenting colorectal cancer MRI images based on deep learning. Background Art

[0002] In the clinical diagnosis and treatment of rectal cancer, magnetic resonance imaging (MRI) technology has become an important tool for evaluating the local invasion range of tumors and the relationship of surrounding anatomical structures due to its excellent soft tissue resolution; through MRI image analysis, clinicians can obtain key information such as tumor location, invasion depth, and lymph node status, providing important basis for the formulation of surgical plans and the selection of comprehensive treatment strategies. Especially in the evaluation of the circumferential resection margin (CRM) of the rectal mesenteric fascia, the accuracy of imaging results is directly related to the planning of the surgical resection range, and thus affects the risk of local recurrence after surgery and the survival prognosis of patients.

[0003] However, existing MRI technologies still have significant limitations in the preoperative evaluation of rectal cancer. That is, due to the lack of a unified standard for scan parameter settings, imaging sequence selection, and operation specifications in different medical institutions, there are significant differences in the image display quality of key anatomical structures (such as the rectal mesenteric fascia). This inconsistency makes it difficult to accurately quantify the spatial relationship between the tumor and the fascia, resulting in errors in the judgment of the surgical margin status and affecting the achievement of the radical resection goal. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method for segmenting colorectal cancer MRI images based on deep learning to solve the problems raised in the above background art.

[0005] To achieve the above object, the present invention provides the following technical solutions: A method for segmenting colorectal cancer MRI images based on deep learning, comprising the following steps: S1. Based on a standardized protocol including T2WI non-fat-suppressed sequence, DWI, and IVIM imaging, adjust scan parameters to obtain pelvic multi-modal MRI data; S2. Analyze the registration error between DWI and IVIM imaging according to the edge continuity parameter of the fascia layer in the T2WI non-fat-suppressed sequence, and re-acquire DWI and IVIM imaging when the registration error does not meet the standard; S3. Dynamically allocate the fusion weights of DWI and IVIM imaging based on the edge continuity parameter; S4. Input the pelvic multi-modal MRI data and the fusion weights into a segmentation model to generate a segmentation mask of the minimum infiltration distance from the tumor to the fascia layer; S5. Analyze the perfusion gradient change of IVIM imaging in the tumor-fascia transition zone according to the segmentation mask and label the micro-infiltration area; S6. Compare the minimum infiltration distance and the range of the micro-infiltration area with the surgical safety threshold to generate a structured report containing TNM staging parameters and the status of fascial invasion.

[0006] In a preferred embodiment, based on a standardized protocol including T2WI non-fat-suppressed sequence, DWI, and IVIM imaging, adjust the scanning parameters to obtain pelvic multi-modal MRI data, including: The scanning parameters of the T2WI non-fat-suppressed sequence in the standardized protocol include a long echo time, a long repetition time, a preset thin slice thickness, and no fat suppression technique applied; The scanning parameters of DWI include applying a high diffusion-sensitizing gradient field under a single-exponential diffusion model; The scanning parameters of IVIM imaging include a multi-b value acquisition mode with multiple gradient field intensities; Adjusting the scanning parameters includes controlling the slice gap between the transverse and sagittal scans within a preset gap according to the spatial resolution requirements of the pelvic anatomical structure, and the field of view covering the pelvic target area.

[0007] In a preferred embodiment, analyze the registration error between DWI and IVIM imaging according to the edge continuity parameters of the fascial layer in the T2WI non-fat-suppressed sequence, and re-acquire DWI and IVIM imaging when the registration error does not meet the standard, including: Based on the edge continuity parameters of the mesorectal fascia layer in the T2WI non-fat-suppressed sequence, calculate the three-dimensional spatial displacement deviation and rotation angle deviation between DWI and IVIM imaging through a rigid registration algorithm; The edge continuity parameters calculate the continuity score by extracting the number of break points of the fascial layer edge contour and the adjacent contour spacing in the T2WI non-fat-suppressed sequence; The registration error not meeting the standard means that the displacement deviation exceeds the preset displacement threshold or the rotation angle deviation exceeds the preset angle threshold; When the registration error does not meet the standard, re-acquiring DWI and IVIM imaging includes updating the scanning positioning line according to the pelvic anatomical position and re-executing the scanning parameters of DWI and IVIM imaging in the standardized protocol.

[0008] In a preferred embodiment, the continuity score is calculated as the ratio with the number of break points as the numerator and the average spacing as the denominator.

[0009] In a preferred embodiment, dynamically allocate the fusion weights of DWI and IVIM imaging based on the edge continuity parameters, including: Determine the weight allocation rule according to the comparison result between the edge continuity parameters and the preset critical value; The preset critical value is set based on the statistical impact of the integrity of the fascial layer edge on the accuracy of multimodal registration in historical data, specifically as the critical threshold for the continuity score exceeding the clinically acceptable error boundary; Assigning fusion weights includes linearly adjusting the weight ratio of DWI and IVIM imaging according to the attenuation amplitude of the edge continuity parameter. The greater the attenuation amplitude, the greater the increase in the DWI weight.

[0010] In a preferred embodiment, the weight assignment rule is that when the edge continuity parameter is lower than the preset critical value, the fusion weight of DWI is increased to be higher than the fusion weight of IVIM imaging.

[0011] In a preferred embodiment, inputting pelvic multimodal MRI data and fusion weights into a segmentation model to generate a segmentation mask of the minimum infiltration distance from the tumor to the fascial layer, including: Fusing the anatomical boundary features of the T2WI non-fat-suppressed sequence, the diffusion signal features of DWI, and the perfusion parameters of IVIM imaging; The fusion process includes rigidly registering and aligning the spatial coordinates of the anatomical boundary features with the spatial coordinates of the diffusion signal features and perfusion parameters, and using the fusion weight as the weighting coefficient for each modal feature channel; The segmentation model extracts multimodal fusion features through a three-dimensional convolutional neural network and outputs a binary segmentation mask containing the tumor region and the fascial layer boundary; The calculation method of the minimum infiltration distance from the tumor to the fascial layer is to measure the Euclidean distance from each point along the tumor edge to the fascial layer boundary in the segmentation mask, and take the minimum value of all measured values as the final infiltration distance.

[0012] In a preferred embodiment, analyzing the perfusion gradient change of IVIM imaging in the tumor-fascia transition zone and annotating the micro-infiltration region according to the segmentation mask, including: Determining the spatial range of the tumor-fascia transition zone based on the minimum infiltration distance from the tumor to the fascial layer in the segmentation mask, and extracting the perfusion fraction and pseudo-diffusion coefficient of each voxel in the transition zone from the perfusion parameter map of IVIM imaging; Analyzing the perfusion gradient change includes dividing sampling windows in the direction from the tumor edge to the fascial layer at a preset step size, calculating the mean change rate of the perfusion fraction and the standard deviation of the pseudo-diffusion coefficient in each window, and generating a perfusion gradient distribution curve along the infiltration direction; Annotating the micro-infiltration region includes marking the corresponding voxels as high-risk micro-infiltration regions according to the continuous window regions in the perfusion gradient distribution curve where the change rate exceeds the preset change threshold and the standard deviation is lower than the preset stability threshold.

[0013] In a preferred embodiment, the preset change threshold and stability threshold are determined based on the ROC curve analysis of the microinvasive pathological results and perfusion gradient parameters in historical data, specifically the critical value when the Youden index is the largest.

[0014] In a preferred embodiment, the minimum invasion distance and the range of the microinvasive area are compared with the surgical safety threshold to generate a structured report including TNM staging parameters and the status of fascial invasion, including: When the minimum invasion distance is less than the preset distance threshold, mark the positive status of fascial invasion; Compare the ratio of the spatial range of the high-risk microinvasive area to the total tumor volume with the pathological staging standard, which is the microinvasion proportion threshold defined in the cancer staging manual. When the ratio exceeds the proportion threshold, up-regulate the T staging parameter; The generated structured report includes TNM staging parameters, a text description of the status of fascial invasion, and three-dimensional coordinate data of the high-risk area. The TNM staging parameters are comprehensively determined based on the independent analysis results of the minimum invasion distance, microinvasion proportion, and lymph node metastasis and distant metastasis; The structured report is integrated into the imaging archiving system in the Digital Imaging and Communications in Medicine (DICOM) standard format, including an interactive viewable infiltration risk assessment layer and a table of staging parameters.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By constructing a standardized multimodal MRI scanning protocol and an intelligent analysis process, the present invention effectively solves the problem of inconsistent imaging quality of key anatomical structures in the imaging evaluation of rectal cancer; based on the anatomical benchmark of the T2-weighted imaging without fat suppression sequence, combined with dynamic registration error analysis and fusion weight assignment technology, it significantly improves the spatial alignment accuracy and data fusion reliability of diffusion-weighted imaging and intravoxel incoherent motion imaging, avoiding the quantification limitation of the measurement of the infiltration distance between the tumor and the mesorectal fascia; through the collaborative analysis of perfusion gradient features and multimodal segmentation models, it realizes the accurate spatial positioning of the tumor microinvasive area.

[0016] 2. By deeply coupling imaging features with clinical staging criteria, a complete conversion chain from raw data to surgical strategies is constructed; based on the dynamic threshold determination of the invasion distance and microinvasion range, a visual report linked to TNM staging is automatically generated, which not only completely preserves the quantitative evidence of tumor invasion, but also realizes the intuitive mapping of imaging conclusions and surgical navigation through three-dimensional anatomical calibration; through the organic integration of standardized data streams and intelligent analysis models, the preoperative evaluation results have cross-institutional repeatability and clinical decision interpretability, providing a high-confidence imaging guidance paradigm for the achievement of the goal of radical resection of rectal cancer. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flowchart of a method for segmenting colorectal cancer MRI images based on deep learning according to the present invention. Specific implementation manners

[0018] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] Embodiment: Figure 1 A method for segmenting colorectal cancer MRI images based on deep learning according to the present invention is provided, which includes the following steps: S1. Based on a standardized protocol including T2WI non-fat-suppressed sequence, DWI, and IVIM imaging, adjust the scanning parameters to obtain pelvic multi-modal MRI data; S2. Analyze the registration error between DWI and IVIM imaging according to the edge continuity parameters of the fascia layer in the T2WI non-fat-suppressed sequence, and re-acquire DWI and IVIM imaging when the registration error does not meet the standard; S3. Dynamically allocate the fusion weights of DWI and IVIM imaging based on the edge continuity parameters; S4. Input the pelvic multi-modal MRI data and the fusion weights into the segmentation model to generate a segmentation mask of the minimum infiltration distance from the tumor to the fascia layer; S5. Analyze the perfusion gradient change of IVIM imaging in the tumor-fascia transition zone according to the segmentation mask and label the micro-infiltration area; S6. Compare the minimum infiltration distance and the range of the micro-infiltration area with the surgical safety threshold to generate a structured report including TNM staging parameters and the status of fascia invasion.

[0020] S1. Based on a standardized protocol including T2WI non-fat-suppressed sequence, DWI, and IVIM imaging, adjust the scanning parameters to obtain pelvic multi-modal MRI data, including: The scanning parameters of the T2WI non-fat-suppressed sequence in the standardized protocol include a long echo time, a long repetition time, a preset thin slice thickness, and no fat suppression technology applied; The scanning parameters of DWI include applying a high diffusion-sensitizing gradient field under a single-exponential diffusion model; The scanning parameters of IVIM imaging include a multi-b value acquisition mode with multiple gradient field intensities; Adjusting the scanning parameters includes controlling the slice interval of the transverse and sagittal scans within a preset interval according to the spatial resolution requirements of the pelvic anatomical structure, and covering the pelvic target area with the field of view.

[0021] In the standardized protocol, the scanning parameters of the T2WI non-fat-suppressed sequence include setting the echo time in the long echo time range. For example, in a conventional magnetic resonance imaging device, the long echo time parameter optimized for pelvic anatomy can be set between 80 milliseconds and 120 milliseconds. The repetition time is set in the long repetition time range. For example, in a conventional magnetic resonance imaging device, the long repetition time parameter optimized for soft tissue contrast can be set between 3000 milliseconds and 5000 milliseconds. The slice thickness is set to a preset thin slice thickness parameter. For example, the slice thickness parameter is set to not exceed 3 millimeters to ensure high spatial resolution. And during the scanning process, the fat suppression technique is not applied, and the contrast between the mesorectal fascia layer and surrounding tissues is enhanced by retaining the natural signal of adipose tissue.

[0022] The scanning parameters of DWI include the DWI technique using a single exponential diffusion model. For example, a diffusion-sensitized gradient field with a high diffusion-sensitized gradient field strength is applied. The range of the b value corresponding to the high diffusion-sensitized gradient field strength is, for example, set between 800 seconds per square millimeter and 1000 seconds per square millimeter. The apparent diffusion coefficient is calculated by the single exponential model to quantify the degree of water molecule diffusion restriction.

[0023] The scanning parameters of IVIM imaging include a multi-b value acquisition mode using multi-gradient field strengths. For example, a gradient field strength combination including four different b values, such as set to 0 seconds per square millimeter, 50 seconds per square millimeter, 500 seconds per square millimeter, and 1000 seconds per square millimeter. The true diffusion effect and perfusion effect are separated by a double exponential model. Among them, the true diffusion effect is characterized by the diffusion coefficient, and the perfusion effect is jointly characterized by the perfusion fraction and the pseudo-diffusion coefficient.

[0024] Adjusting the scanning parameters includes controlling the slice gap between the transverse scan and the sagittal scan within a preset gap range according to the requirement of spatial resolution for pelvic anatomy. For example, the slice gap parameter is set to not exceed 1 millimeter to ensure the image continuity between adjacent slices during the three-dimensional reconstruction process. At the same time, the field of view range covers the pelvic target area, such as covering the anatomical area from the upper edge of the iliac crest to the end of the anal canal. The image artifacts caused by pelvic respiratory movement and intestinal peristalsis are suppressed by adjusting the field of view matrix and phase encoding direction oversampling.

[0025] Pelvic multi-modal magnetic resonance imaging data includes high-resolution anatomical images generated by the T2WI non-fat-suppressed sequence, diffusion signal intensity distribution maps generated by DWI, and perfusion parameter maps generated by IVIM imaging. The three-modal data are stored as three-dimensional volume data after spatial coordinate alignment for subsequent multi-modal feature fusion and segmentation modeling.

[0026] S2. Analyze the registration error between DWI and IVIM imaging according to the edge continuity parameter of the fascia layer in the T2WI non-fat-suppressed sequence. When the registration error does not meet the standard, re-acquire DWI and IVIM imaging, including: Based on the edge continuity parameter of the mesorectal fascia layer in the T2WI non-fat-suppressed sequence, calculate the three-dimensional spatial displacement deviation and rotation angle deviation between DWI and IVIM imaging through a rigid registration algorithm; The edge continuity parameter calculates the continuity score by extracting the number of break points on the edge contour of the fascia layer and the adjacent contour spacing in the T2WI non-fat-suppressed sequence. The calculation method of the continuity score is the ratio with the number of break points as the numerator and the average spacing as the denominator; The registration error not meeting the standard means that the displacement deviation exceeds the preset displacement threshold or the rotation angle deviation exceeds the preset angle threshold; When the registration error does not meet the standard, re-acquiring DWI and IVIM imaging includes updating the scan positioning line according to the pelvic anatomical position and re-executing the scan parameters of DWI and IVIM imaging in the standardized protocol.

[0027] The execution steps of the rigid registration algorithm include: selecting the anatomical landmark points on the edge of the mesorectal fascia layer in the T2WI non-fat-suppressed sequence as the registration reference points. The anatomical landmark points are the points with the maximum curvature or bifurcation points on the edge of the fascia layer. Manually or automatically identify the corresponding anatomical positions of the reference points in DWI and IVIM imaging. Calculate the three-dimensional translation matrix and rotation matrix of DWI and IVIM imaging relative to the T2WI non-fat-suppressed sequence through the least squares method. The displacement deviation is the square root of the sum of the squares of the displacement amounts in each axis of the translation matrix, and the rotation angle deviation is the sum of the absolute values of the rotation angles around each coordinate axis in the rotation matrix.

[0028] The extraction method of the number of break points is: calculate the signal intensity gradient pixel by pixel along the edge contour line of the fascia layer in the transverse image. When the gradient values of three consecutive pixels exceed the preset gradient threshold, it is determined as the starting position of the break point until the gradient value is lower than the threshold and it is determined as the ending position of the break point. When the interval between two adjacent break points on the same contour line is less than the preset distance threshold, they are merged into one break point; the adjacent contour spacing is the average Euclidean distance between the center lines of the edge contours of two adjacent fascia layers in the same anatomical plane; the calculation method of the continuity score is the ratio with the number of break points as the numerator and the average spacing as the denominator. For example, when the number of detected break points is 4 and the average spacing is 1.5 mm, the continuity score is 4 / 1.5≈2.67.

[0029] The method for setting the preset displacement threshold is as follows: Based on the spatial resolution of the magnetic resonance device and the clinically allowable anatomical positioning error, select twice the value of the device resolution as the displacement threshold. For example, when the device resolution is 0.5 mm, the displacement threshold is set to 1 mm. The method for setting the preset angle threshold is as follows: According to the statistical results of the pelvic organ movement amplitude, select the maximum rotation angle within the 95% confidence interval as the angle threshold. For example, when the maximum rotation angle obtained by statistics is 1.8 degrees, the angle threshold is set to 2 degrees.

[0030] The specific steps for updating the scan positioning line are as follows: Taking the latest image of the T2WI non-fat suppression sequence as a reference, redraw the scan centerlines of DWI and IVIM imaging in the sagittal and coronal views to ensure coverage of the target area of the mesorectal fascia layer, and adjust the upper and lower boundaries of the scan range according to the patient's body position change; Re-execute the scan parameters in the standardized protocol, including: the number of repetitions is 3 times, the slice thickness is 3 mm, the slice spacing is 0 mm, and the matrix size is 128×128.

[0031] S3. Dynamically allocate the fusion weights of DWI and IVIM imaging based on the edge continuity parameter, including: Determine the weight allocation rule according to the comparison result between the edge continuity parameter and the preset critical value; The weight allocation rule is that when the edge continuity parameter is lower than the preset critical value, increase the fusion weight of DWI to be higher than that of IVIM imaging; The preset critical value is set based on the statistical influence of the fascia layer edge integrity on the multi-modal registration accuracy in historical data, specifically as the critical threshold for the continuity score exceeding the clinically acceptable error boundary; Allocating the fusion weights includes linearly adjusting the weight ratio of DWI and IVIM imaging according to the attenuation amplitude of the edge continuity parameter. The greater the attenuation amplitude, the greater the increase in the DWI weight.

[0032] The edge continuity parameter is the continuity score, and its calculation method is the ratio of the number of break points of the fascia layer edge contour to the average value of the adjacent contour spacing.

[0033] For example, when the preset critical value is set to 2.0, if the calculated value of the edge continuity parameter is 1.5, then set the weight of DWI to 0.7 and the weight of IVIM imaging to 0.3; if the edge continuity parameter is 2.5, then keep the weights of DWI and IVIM imaging both at 0.5.

[0034] For example, the preset critical value is specifically obtained by retrospectively analyzing at least 100 sets of pelvic magnetic resonance imaging data, statistically analyzing the consistency between multimodal registration errors and surgical pathological results under different continuity scores, and selecting the maximum continuity score when the registration error exceeds the clinically acceptable error boundary as the critical threshold. For example, when the clinically acceptable error boundary is that the registration displacement deviation does not exceed 1 mm, the statistically obtained critical threshold is 2.0.

[0035] The attenuation amplitude is the relative difference between the edge continuity parameter and the preset critical value. For example, when the preset critical value is 2.0 and the actual parameter is 1.5, the attenuation amplitude is (2.0 - 1.5) / 2.0 = 25%. At this time, the increase amplitude of the DWI weight is proportional to the attenuation amplitude. For example, for every 10% increase in the attenuation amplitude, the DWI weight increases by 0.1, and the IVIM imaging weight decreases by 0.1.

[0036] The execution steps for allocating the fusion weight include: reading the edge continuity parameter and the preset critical value in the image fusion module, calculating the attenuation amplitude according to the difference between the current parameter and the critical value, updating the weight ratio according to the linear adjustment rule, and applying the updated weight to the weighted fusion process of the diffusion signal intensity of DWI and the perfusion parameter of IVIM imaging to generate a fused multimodal feature map. The weighted fusion process is to add the diffusion signal intensity value of DWI and the perfusion parameter value of IVIM imaging according to the weight ratio to generate fused image data containing anatomical and functional information.

[0037] S4. Input the pelvic multimodal MRI data and the fusion weight into the segmentation model to generate a segmentation mask of the minimum infiltration distance from the tumor to the fascia layer, including: Fusing the anatomical boundary features of the T2WI non-fat-suppressed sequence, the diffusion signal features of DWI, and the perfusion parameters of IVIM imaging; The fusion process includes rigidly registering and aligning the spatial coordinates of the anatomical boundary features with the spatial coordinates of the diffusion signal features and the perfusion parameters, and using the fusion weight as the weighting coefficient for each modal feature channel; The segmentation model extracts multimodal fusion features through a three-dimensional convolutional neural network and outputs a binary segmentation mask containing the boundaries of the tumor region and the fascia layer; The calculation method of the minimum infiltration distance from the tumor to the fascia layer is to measure the Euclidean distance from the tumor edge to the fascia layer boundary point by point in the segmentation mask, and take the minimum value of all measured values as the final infiltration distance.

[0038] The execution steps of rigid registration alignment are as follows: Taking the image space coordinate system of the T2WI imaging without fat suppression sequence as the benchmark, the spatial coordinate systems of DWI and IVIM imaging are transformed to be consistent with the T2WI imaging without fat suppression sequence through the feature point matching algorithm. The feature point matching algorithm adopts the registration method based on mutual information, specifically calculating the similarity of the pixel intensity distributions of DWI or IVIM imaging and the T2WI imaging without fat suppression sequence at the same anatomical level, and optimizing the parameters of the rotation matrix and translation matrix by maximizing the mutual information value.

[0039] The fusion weight is used as the weighting coefficient for each modal feature channel, and its application method is as follows: In the fusion process, the diffusion signal features of DWI are multiplied by the corresponding fusion weight, the perfusion parameters of IVIM imaging are multiplied by the corresponding fusion weight, and the weighted eigenvalue is superimposed with the anatomical boundary features of the T2WI imaging without fat suppression sequence to generate a multi-modal fusion feature map; for example, when the fusion weight of DWI is 0.7 and the fusion weight of IVIM imaging is 0.3, each pixel signal value of DWI is multiplied by 0.7, each perfusion parameter value of IVIM imaging is multiplied by 0.3, and then added to the pixel value of the T2WI imaging without fat suppression sequence.

[0040] The structure of the three-dimensional convolutional neural network includes an input layer, three three-dimensional convolutional layers, three three-dimensional max-pooling layers, three three-dimensional deconvolutional layers, and an output layer. The input layer receives the multi-modal fusion feature map. The convolutional kernel size of each three-dimensional convolutional layer is 3×3×3, and the activation function is the ReLU function. The pooling window size of the three-dimensional max-pooling layer is 2×2×2. The three-dimensional deconvolutional layer restores the spatial resolution through upsampling. The output layer generates a binary segmentation mask through the Sigmoid function. The training data of the segmentation model is a publicly available dataset containing at least 200 pelvic magnetic resonance images and surgical pathology annotations. The cross-entropy loss function is used to optimize the network parameters during the training process.

[0041] The execution steps of point-by-point measurement include: Extracting the contour lines of the tumor region and the fascia layer boundary from the segmentation mask, expanding the tumor contour outward to contact the fascia layer contour through the morphological dilation algorithm, calculating the shortest distance from each point on the dilated tumor contour to the fascia layer contour, and recording the minimum value among all distances; for example, when a certain point on the tumor contour contacts the fascia layer contour after dilation, the Euclidean distance of this point is 0.5 mm. If this is the minimum measured value among all points, the final infiltration distance is 0.5 mm.

[0042] The generated segmentation mask and the minimum infiltration distance are visualized through a medical imaging workstation and stored in the DICOM format in the hospital image archiving system for clinical review.

[0043] S5. Analyze the perfusion gradient changes of IVIM imaging in the tumor-fascia transition zone based on the segmentation mask and label the microinvasion area, including: Determine the spatial range of the tumor-fascia transition zone based on the minimum invasion distance from the tumor to the fascia layer in the segmentation mask, and extract the perfusion fraction and pseudo-diffusion coefficient of each voxel in the transition zone from the perfusion parameter map of IVIM imaging; Analyze the perfusion gradient changes, including dividing sampling windows in the direction from the tumor edge to the fascia layer at a preset step size, calculating the mean change rate of the perfusion fraction and the standard deviation of the pseudo-diffusion coefficient in each window, and generating a perfusion gradient distribution curve along the invasion direction; Label the microinvasion area, including marking the corresponding voxels as high-risk microinvasion areas according to the continuous window areas in the perfusion gradient distribution curve where the change rate exceeds the preset change threshold and the standard deviation is lower than the preset stability threshold; The preset change threshold and stability threshold are determined based on the ROC curve analysis of the microinvasion pathological results and perfusion gradient parameters in historical data, specifically the critical values when the Youden index is the largest.

[0044] The segmentation mask is the binary image data including the boundary between the tumor area and the fascia layer; the spatial range of the tumor-fascia transition zone is defined as a region with a width of 1.5 times the minimum invasion distance extending from the tumor edge contour to the fascia layer. For example, when the minimum invasion distance is 2 mm, the transition zone range is a band-shaped area with an extension of 3 mm from the tumor edge; extract the perfusion fraction and pseudo-diffusion coefficient of each voxel in the transition zone from the perfusion parameter map of IVIM imaging, and the perfusion parameter map is the generated perfusion fraction map and pseudo-diffusion coefficient map.

[0045] The preset step size is a fixed value set based on the spatial resolution of the magnetic resonance image. For example, when the image resolution is 1 mm × 1 mm × 3 mm, the preset step size is set to 2 mm; calculate the mean change rate of the perfusion fraction and the standard deviation of the pseudo-diffusion coefficient in each window. The calculation method of the mean change rate is the difference between the mean perfusion fractions of adjacent windows divided by the mean of the previous window. For example, when the mean of the previous window is 0.25 and the mean of the next window is 0.20, the change rate is (0.20 - 0.25) / 0.25 = -20%; the standard deviation is an index of the dispersion degree of all voxel pseudo-diffusion coefficient values in the same window. For example, when the pseudo-diffusion coefficient values are 1.2×10 -3 mm² / s, 1.3×10 -3 mm² / s, 1.1×10 -3 mm² / s, the standard deviation is 0.1×10 -3 mm² / s; generate a perfusion gradient distribution curve along the invasion direction. The horizontal axis of the curve is the distance from the tumor edge to the fascia layer, and the vertical axis is a combined index of the mean change rate of the perfusion fraction and the standard deviation of the pseudo-diffusion coefficient.

[0046] The microinfiltration region is marked by identifying consecutive window regions in the perfusion gradient distribution curve with a change rate exceeding a preset change threshold and a standard deviation lower than a preset stability threshold, and marking the corresponding voxels as high-risk microinfiltration regions.

[0047] The preset change threshold and stability threshold are determined based on the ROC curve analysis of the microinfiltration pathological results and perfusion gradient parameters in historical data. Specifically, by retrospectively analyzing at least 200 pelvic magnetic resonance imaging data with surgically proven microinfiltration, the ROC curve of the change rate of perfusion fraction and the standard deviation of the pseudo-diffusion coefficient for microinfiltration positivity is plotted, and the critical value at the maximum Youden index is selected as the threshold; for example, when the change rate threshold is -15% and the standard deviation threshold is 0.15×10 -3 mm² / s, the consecutive window regions satisfying the change rate ≤ -15% and the standard deviation ≤ 0.15×10 -3 mm² / s are marked as high-risk regions.

[0048] The marking results are overlaid on the anatomical image of the T2WI non-fat-suppressed sequence through color coding. Among them, the high-risk microinfiltration region is displayed as a red semi-transparent overlay, and an infiltration risk heat map is generated through the three-dimensional reconstruction function of the medical imaging workstation for clinicians to evaluate the tumor invasion range.

[0049] S6. Compare the minimum infiltration distance and the microinfiltration region range with the surgical safety threshold to generate a structured report containing TNM staging parameters and the status of fascia invasion, including: The surgical safety threshold is set as a preset distance threshold according to the fascia invasion determination criteria in international clinical guidelines. When the minimum infiltration distance is less than the preset distance threshold, the positive state of fascia invasion is marked; Compare the ratio of the spatial range of the high-risk microinfiltration region to the total tumor volume with the pathological staging standard, which is the microinfiltration proportion threshold defined in the cancer staging manual. When the ratio exceeds the proportion threshold, the T staging parameter is up-regulated; The generated structured report includes TNM staging parameters, a text description of the fascia invasion status, and three-dimensional coordinate data of the high-risk region. The TNM staging parameters are comprehensively determined based on the independent analysis results of the minimum infiltration distance, microinfiltration proportion, and lymph node metastasis and distant metastasis; The structured report is integrated into the imaging archiving system in the Digital Imaging and Communications in Medicine (DICOM) standard format, containing an interactive infiltration risk assessment layer and a staging parameter table.

[0050] The surgical safety threshold is set as a preset distance threshold according to the criteria for determining fascia invasion in international clinical guidelines. For example, in the international colorectal cancer treatment guidelines, the surgical safety threshold for fascia invasion is defined as 1 mm. When the minimum infiltration distance from the tumor to the fascia layer is less than 1 mm, the positive state of fascia invasion is marked. The marking method for the positive state of fascia invasion is to label it with a red warning sign at the corresponding anatomical position in the structured report and clearly record "high risk of fascia layer invasion" in the text description.

[0051] The ratio of the spatial range proportion of the high-risk microinvasion area to the total tumor volume is compared with the pathological staging criteria. The pathological staging criteria are the microinvasion proportion thresholds defined in the cancer staging manual. For example, in the colorectal cancer staging guidelines of the American Joint Committee on Cancer (AJCC), when the microinvasion proportion exceeds 20% is used as the threshold for upstaging the T stage. The calculation method for the spatial range proportion of the high-risk microinvasion area is to divide the total number of voxels in the marked high-risk microinvasion area by the total number of voxels in the tumor area. For example, when the number of voxels in the high-risk area is 500 and the total number of voxels in the tumor volume is 2000, the proportion is 25%. At this time, it is determined that the threshold of 20% is exceeded and the T stage parameter is upstaged from T2 to T3. The generated structured report includes TNM staging parameters, text description of the fascia invasion status, and three-dimensional coordinate data of the high-risk area. The TNM staging parameters are comprehensively determined based on the independent analysis results of the minimum infiltration distance, microinvasion proportion, and lymph node metastasis and distant metastasis. Specifically: the T stage parameter is jointly determined by whether the minimum infiltration distance is lower than the surgical safety threshold and whether the proportion of the high-risk microinvasion area exceeds the pathological staging criteria. The N stage parameter is based on the independent imaging analysis results of pelvic lymph nodes, and the M stage parameter is based on the detection results of distant metastasis foci in whole-body CT or PET-CT images. For example, when the minimum infiltration distance is 0.8 mm (lower than the 1 mm threshold) and the microinvasion proportion is 25% (exceeding the 20% threshold), the T stage is determined as T3. If 1 metastatic lymph node is found in the lymph node analysis, the N stage is N1, and when there is no distant metastasis, the M stage is M0.

[0052] The structured report is integrated into the image archiving system in the format of Digital Imaging and Communications in Medicine (DICOM) structured report (DICOM-SR) format, which contains an interactive viewable infiltration risk assessment layer and a staging parameter table. The infiltration risk assessment layer is a color semi-transparent overlay generated by superimposing the three-dimensional coordinate data of the high-risk microinvasion area on the anatomical image of the T2WI non-fat-suppressed sequence. The high-risk area is shown in dark red and supports clicking to view the specific coordinate values. The staging parameter table is stored in an independent DICOM-SR module, which contains TNM staging parameters, text description of the fascia invasion status, and data source citation information, such as marked "T3N1M0, positive fascia invasion, data source: MRI multimodal analysis and CT images".

[0053] The generation process of the structured report includes: retrieving the minimum invasion distance data, the coordinates of the microinvasion high-risk area, the results of independent lymph node analysis, and the results of whole-body imaging examinations from the medical imaging archiving system. Through the report generation engine, the above data is filled into the DICOM-SR standard template, automatically matching the anatomical position description fields with the staging parameter logic rules, and finally outputting an electronically report that can be interactively viewed and transmitted back to the imaging archiving system for clinical review. The template configuration of the report generation engine is based on the staging rules in international clinical guidelines. For example, the determination logic and text description format of the TNM fields are preset according to the AJCC guidelines.

[0054] It should be noted that the magnetic resonance imaging (MRI) technology adopted in the present invention includes the T2-weighted imaging without fat suppression sequence (abbreviated as T2WI without fat suppression sequence), diffusion-weighted imaging (abbreviated as DWI), and intravoxel incoherent motion imaging (abbreviated as IVIM imaging).

[0055] The T2WI without fat suppression sequence obtains high-signal-to-noise anatomical images by setting the scan parameters of the echo time (TE) of 80 - 120 ms and the repetition time (TR) of 3000 - 5000 ms, and clearly shows the edge morphology of the fascia layer using the natural signal contrast between adipose tissue and the fascia layer. Its slice thickness is controlled within 3 mm, and the field of view (FOV) is set to 20 - 24 cm to cover the pelvic target area.

[0056] DWI adopts a single-exponential diffusion model, applies a diffusion-sensitive gradient field with a b value of 800 - 1000 s / mm², and quantifies the degree of water molecule diffusion restriction in the tumor area by calculating the apparent diffusion coefficient (ADC value), which is used to detect high-cell-density malignant tumor lesions and localize metastatic lymph nodes.

[0057] IVIM imaging acquires data based on multiple b values (at least including 0, 50, 500, 1000 s / mm²), uses a double-exponential model to separate the true diffusion coefficient D value and the perfusion-related parameter f value, and identifies subclinical microinvasion lesions by analyzing the D value spatial distribution gradient in the tumor-fascia transition area. The scanning orientations of the three sequences all include the transverse and sagittal positions, and the slice gap is ≤1 mm to ensure the three-dimensional reconstruction accuracy. During the scanning process, respiratory gating and phase-encoding direction oversampling are used to suppress motion artifacts.

[0058] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product.

[0059] Those of ordinary skill in the art can realize that the modules and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application of the technical solution and the inventive constraints. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0060] In addition, in each embodiment of this application, the functional modules can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module.

[0061] As mentioned above, the above is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed in this application can easily think of changes or substitutions, which should all be covered by the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

[0062] Finally: The above is only the preferred embodiment of the present invention and is not used to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for segmenting colorectal cancer MRI images based on deep learning, characterized in that, It includes the following steps: S1. Based on a standardized protocol including T2WI non-fat-suppressed sequence, DWI, and IVIM imaging, adjust the scanning parameters to obtain pelvic multimodal MRI data; S2. Analyze the registration error between DWI and IVIM imaging according to the edge continuity parameters of the fascia layer in the T2WI non-fat-suppressed sequence, and re-acquire DWI and IVIM imaging when the registration error fails to meet the standard; S3. Dynamically assign the fusion weights of DWI and IVIM imaging based on the edge continuity parameters; S4. Input the pelvic multimodal MRI data and the fusion weights into the segmentation model to generate a segmentation mask of the minimum infiltration distance from the tumor to the fascia layer; S5. Analyze the perfusion gradient change of IVIM imaging in the tumor-fascia transition zone according to the segmentation mask and label the micro-infiltration area; S6. Compare the minimum infiltration distance and the range of the micro-infiltration area with the surgical safety threshold to generate a structured report containing TNM staging parameters and the status of fascia invasion; 2. The colorectal cancer MRI image segmentation method based on deep learning according to claim 1, wherein, Based on a standardized protocol including T2WI non-fat-suppressed sequence, DWI, and IVIM imaging, adjust the scanning parameters to obtain pelvic multimodal MRI data, including: The scanning parameters of the T2WI non-fat-suppressed sequence in the standardized protocol include a long echo time, a long repetition time, a preset thin slice thickness, and no fat suppression technique applied; The scanning parameters of DWI include applying a high diffusion-sensitizing gradient field under a single-exponential diffusion model; The scanning parameters of IVIM imaging include a multi-b value acquisition mode with multiple gradient field intensities; Adjusting the scanning parameters includes controlling the slice spacing of transverse and sagittal scans within a preset spacing according to the spatial resolution requirements of pelvic anatomy, and covering the pelvic target area with the field of view; 3. A method for segmenting MRI images of colorectal cancer based on deep learning according to claim 1, characterized in that, Analyze the registration error between DWI and IVIM imaging according to the edge continuity parameters of the fascia layer in the T2WI non-fat-suppressed sequence, and re-acquire DWI and IVIM imaging when the registration error fails to meet the standard, including: Based on the edge continuity parameters of the mesorectal fascia layer in the T2WI non-fat-suppressed sequence, calculate the three-dimensional spatial displacement deviation and rotation angle deviation between DWI and IVIM imaging through a rigid registration algorithm; The edge continuity parameters calculate the continuity score by extracting the number of break points on the edge contour of the fascia layer and the adjacent contour spacing in the T2WI non-fat-suppressed sequence; The registration error not meeting the standard means that the displacement deviation exceeds the preset displacement threshold or the rotation angle deviation exceeds the preset angle threshold; When the registration error fails to meet the standard, re-acquiring DWI and IVIM imaging includes updating the scanning positioning line according to the pelvic anatomical position and re-executing the scanning parameters of DWI and IVIM imaging in the standardized protocol; 4. A method for segmenting colorectal cancer MRI images based on deep learning according to claim 3, characterized in that, The calculation method of the continuity score is the ratio with the number of break points as the numerator and the average spacing as the denominator; 5. A method for segmenting MRI images of colorectal cancer based on deep learning according to claim 1, characterized in that, Dynamically assign the fusion weights of DWI and IVIM imaging based on the edge continuity parameters, including: Determine the weight assignment rule according to the comparison result between the edge continuity parameters and the preset critical value; The preset critical value is set based on the statistical impact of the edge integrity of the fascia layer on the multimodal registration accuracy in historical data, specifically the critical threshold for the continuity score exceeding the clinically acceptable error boundary; Assigning fusion weights includes linearly adjusting the weight ratio of DWI and IVIM imaging according to the attenuation amplitude of the edge continuity parameter. The greater the attenuation amplitude, the greater the increase in the DWI weight.

6. The method for segmenting colorectal cancer MRI images based on deep learning according to claim 5, wherein, The weight assignment rule is that when the edge continuity parameter is lower than the preset critical value, the fusion weight of DWI is increased to be higher than that of IVIM imaging.

7. A method for segmenting colorectal cancer MRI images based on deep learning according to claim 1, characterized in that, Input the pelvic multimodal MRI data and fusion weights into the segmentation model to generate a segmentation mask of the minimum infiltration distance from the tumor to the fascia layer, including: Fusing the anatomical boundary features of the T2WI non-fat-suppressed sequence, the diffusion signal features of DWI, and the perfusion parameters of IVIM imaging; The fusion process includes rigidly registering and aligning the spatial coordinates of the anatomical boundary features with the spatial coordinates of the diffusion signal features and perfusion parameters, and using the fusion weights as the weighting coefficients for each modal feature channel; The segmentation model extracts multimodal fusion features through a three-dimensional convolutional neural network and outputs a binary segmentation mask containing the boundaries between the tumor region and the fascia layer; The calculation method of the minimum infiltration distance from the tumor to the fascia layer is to measure the Euclidean distance from each point along the tumor edge to the fascia layer boundary in the segmentation mask, and take the minimum value of all measured values as the final infiltration distance.

8. A method for segmenting MRI images of colorectal cancer based on deep learning according to claim 1, characterized in that, Analyze the perfusion gradient change of IVIM imaging in the tumor-fascia transition zone based on the segmentation mask and label the microinfiltration region, including: Determine the spatial range of the tumor-fascia transition zone based on the minimum infiltration distance from the tumor to the fascia layer in the segmentation mask, and extract the perfusion fraction and pseudo-diffusion coefficient of each voxel in the transition zone from the perfusion parameter map of IVIM imaging; Analyzing the perfusion gradient change includes dividing sampling windows in the direction from the tumor edge to the fascia layer at a preset step size, calculating the mean change rate of the perfusion fraction and the standard deviation of the pseudo-diffusion coefficient in each window, and generating a perfusion gradient distribution curve along the infiltration direction; Labeling the microinfiltration region includes marking the corresponding voxels as high-risk microinfiltration regions according to the continuous window regions in the perfusion gradient distribution curve where the change rate exceeds the preset change threshold and the standard deviation is lower than the preset stability threshold.

9. A method for segmenting MRI images of colorectal cancer based on deep learning according to claim 8, characterized in that, The preset change threshold and stability threshold are determined based on the ROC curve analysis of the microinfiltration pathological results and perfusion gradient parameters in historical data, specifically the critical value when the Youden index is the largest.

10. A method for segmenting colorectal cancer MRI images based on deep learning according to claim 1, characterized in that, Compare the minimum infiltration distance and the range of the microinfiltration region with the surgical safety threshold to generate a structured report containing TNM staging parameters and the status of fascia invasion, including: Mark the positive status of fascia invasion when the minimum infiltration distance is less than the preset distance threshold; Compare the ratio of the spatial range of the high-risk microinfiltration region to the total tumor volume with the pathological staging standard. The pathological staging standard is the microinfiltration proportion threshold defined in the cancer staging manual. When the ratio exceeds the proportion threshold, the T staging parameter is up-regulated; The generated structured report includes TNM staging parameters, a text description of the fascia invasion status, and three-dimensional coordinate data of the high-risk region. The TNM staging parameters are comprehensively determined based on the independent analysis results of the minimum infiltration distance, microinfiltration proportion, and lymph node metastasis and distant metastasis; The structured report is integrated into the imaging archiving system in the Digital Imaging and Communications in Medicine (DICOM) standard format, containing an interactive viewable infiltration risk assessment layer and a table of staging parameters.

Images