Artificial intelligence-based rectal cancer MSI prediction model construction method

Abstract

The invention discloses a rectal cancer MSI prediction model construction method based on artificial intelligence, and relates to the technical field of artificial intelligence. The method comprises the steps of collecting multi-center MRI data, and performing data quality control processing; executing a definition verification mechanism based on the multi-center MRI data; the method is technically characterized by comprising the following steps: constructing a double-region fusion logistic regression model for dealing with IT and PT information, and capturing the biological essence of large MSI tumor cell gap and active immune microenvironment by using key physical characteristics derived from OGSE-DWI and a characteristic difference value thereof; on the other hand, the feature difference value is multiplexed to serve as a dynamic correction basis in the risk layering stage, the dynamic correction basis is used for model input to improve prediction accuracy, middle risk group patients are classified again in the second-level risk layering strategy, misjudgment caused by dependence on a probability threshold value is avoided, and the prediction accuracy is improved. And the biological rationality and prognosis layering accuracy of the model are synchronously enhanced.

Description

TECHNICAL FIELD

[0001] The application relates to the field of artificial intelligence, and particularly to a rectal cancer MSI prediction model construction method based on artificial intelligence. BACKGROUND

[0002] Artificial intelligence is a technology for simulating human intelligent behavior by a computer, including learning, reasoning, recognition, decision-making and other capabilities. The current mainstream method takes machine learning as the core, and has achieved remarkable results in the fields of image recognition, natural language processing, medical diagnosis and the like, can automatically extract rules from a large amount of data, realize prediction and judgment on unknown samples, and is used in combination with imaging technology in the medical field; high-resolution pelvic magnetic resonance imaging (MRI) as the preferred examination method for preoperative non-invasive staging of rectal cancer is the consensus of rectal cancer diagnosis and treatment guidelines at home and abroad; multi-modal MRI technology can show the position, shape and tumor infiltration depth of rectal cancer to evaluate the T stage of rectal cancer, and can also identify adverse prognostic factors such as intravascular invasion and circumferential margin positivity of rectal cancer wall; in clinical practice, a corresponding surgical management strategy is usually formulated according to the MRI staging of rectal cancer.

[0003] The current non-invasive prediction of rectal cancer MSI status faces multiple technical bottlenecks:

[0004] Firstly, multi-center MRI data varies due to different equipment models and scanning parameters, resulting in ADC value drift, for example, the ADC value measured by A hospital is 1.2*10 -3 mm / s, while the ADC value of the same patient in B hospital is 0.9*10 -3 mm / s, making it difficult for the model to generalize; secondly, tumor region delineation is highly dependent on the subjective experience of different doctors, and the differences between the delineations lead to more significant differences in subsequent feature extraction results, and traditional methods often ignore the peritumoral region, which cannot reflect the immune microenvironment; thirdly, the features corresponding to the conventional image group are relatively redundant and have a high dimension, and are also susceptible to noise interference, for example, most of the hundreds or thousands of texture features are irrelevant to MSI, leading to overfitting; in addition, the existing risk stratification is a one-size-fits-all approach based on probability, for example, a patient with a traditional MSI probability of 68% is classified as an uncertain state, without using related evidence such as peritumoral looseness, resulting in a high misjudgment rate, and the above problems jointly restrict the effective application of AI models from research to actual clinical practice. SUMMARY

[0005] To achieve the above purpose, the technical scheme is as follows:

[0006] The rectal cancer MSI prediction model construction method based on artificial intelligence comprises the following steps:

[0007] Collecting multi-center MRI data and performing data quality control processing;

[0008] wherein the multi-center MRI data is image data corresponding to a plurality of scan sequences;

[0009] Based on the multi-center MRI data, a definition verification mechanism is performed to generate an intra-tumor region and a peritumoral region;

[0010] According to the intra-tumor region and the peritumoral region, a feature extraction action is performed to obtain original features and calculate key physical features and feature difference values thereof;

[0011] Upon receiving the original features, a multi-constraint driven feature selection strategy is triggered to screen out target features;

[0012] A dual-region fusion logistic regression model is constructed, and texture features and shape features in the target features, key physical features and feature difference values thereof are input, and an MSI probability is output, and a verification strategy is triggered synchronously;

[0013] According to the MSI probability, a first-level risk stratification strategy is performed to preliminarily divide different risk groups, and a feature difference value is introduced for dynamic correction processing: the target preliminarily divided into a medium risk group is extracted, a significant threshold is calculated according to the feature difference value, and a second-level risk stratification strategy is performed to determine whether to adjust the medium risk group.

[0014] Further, the image data is imaging parameters under each scan sequence;

[0015] The scan sequence at least includes: T2WI, OGSE-DWI, DWI and TIWI+C.

[0016] Further, the process of data quality control processing is: performing a scoring mechanism on the image data to exclude image data that does not meet the scoring standard; using a standard phantom periodic calibration device; image de-identification processing.

[0017] Further, the content of performing the definition verification mechanism is: on the imaging corresponding to T2WI, at least two preliminary tumor regions are outlined according to a known rectal cancer MRI report template, and relevant state data is recorded; the intraclass correlation coefficient ICC is calculated according to the two preliminary tumor regions; the intraclass correlation coefficient ICC is compared with a preset standard threshold IC_th, when ICC≥IC_th, the intersection of the two preliminary tumor regions is taken as the final region, that is, the intra-tumor region IT; after determining the intra-tumor region, the boundary of the intra-tumor region is extracted and expanded outwardly by M to form the peritumoral region PT.

[0018] Further, the feature extraction action performed is as follows:

[0019] Feature extraction is performed from eight region-sequence combinations to obtain original features;

[0020] wherein, the eight region-sequence combinations are: T2WI, OGSE-DWI, DWI and TIWI+C of the intra-tumor region IT, and T2WI, OGSE-DWI, DWI and TIWI+C of the peritumor region PT; and the original features at least include: first-order statistical features, shape features, texture features and high-order features after wavelet transform.

[0021] For the OGSE-DWI sequence, the product of the apparent diffusion coefficient ADC(t), the geometric factor and the reciprocal of pi is processed by square root to obtain the effective diffusion diameter Deff as the key physical feature according to the apparent diffusion coefficient ADC(t) obtained at a specific diffusion time t.

[0022] The effective diffusion diameter Deff corresponding to the peritumor region PT is subtracted from the effective diffusion diameter Deff corresponding to the intra-tumor region IT, and the difference is the feature difference ΔDeff.

[0023] Further, the multi-constraint driven feature selection strategy is triggered according to the following:

[0024] Perform the main constraint action, select Q features most relevant to the MSI state using the minimum redundancy maximum relevance mRMR method, and then use LASSO regression for compression to obtain Q / 5 features.

[0025] Perform sub-constraint A, construct a feature stability screening model, complete the first filtering action according to the output stability score, and retain the original features with a stability score exceeding a preset stability threshold.

[0026] Perform sub-constraint B, measure the extraction time and memory occupation of each original feature retained after sub-constraint A on the configured edge device, perform weighted calculation processing according to the number of original features retained after sub-constraint A, extraction time and memory occupation, output the comprehensive cost index, and perform the second filtering action to retain the features corresponding to the preset maximum acceptable cost value as the final target features.

[0027] Further, the operation of the dual-region fusion logistic regression model is as follows:

[0028] The negative cumulative value of e is summed with 1, and the reciprocal of the sum is the MSI probability P(MSI); wherein, each input feature value is multiplied by a corresponding regression coefficient obtained by supervised learning of the dual-region fusion logistic regression model, and the product is added to the intercept term to obtain the cumulative value.

[0029] Further, the triggered verification strategy at least includes: data division, cross-validation and external verification;Wherein, the data division is: sequentially divided into training set, verification set and test set according to 7:1:2;Cross-validation: 5-fold cross-validation is used in the training stage;External validation introduces an external validation queue.

[0030] Further, in the first-level risk stratification strategy, when P(MSI) is less than 30%, it is divided into a low-risk group;When 30%≤P(MSI)<70%, it is divided into a medium-risk group;When 70%≤P(MSI), it is divided into a high-risk group.

[0031] Further, in the dynamic correction process, the basis for calculating the significant threshold D_Δ according to the feature difference ΔDeff is: drawing the ROC curve of predicting MSI on the verification set according to the feature difference ΔDeff, calculating the Youden index, and taking the maximum value of the feature difference ΔDeff as the significant threshold D_Δ;And execute the second risk stratification strategy: when ΔDeff≥D_Δ, it is adjusted to the high-risk group;When ΔDeff≤-D_Δ, it is adjusted to the low-risk group;Otherwise, it continues to be kept in the medium-risk group.

[0032] The present application provides a rectal cancer MSI prediction model based on artificial intelligence, which has the following beneficial effects:

[0033] (1) The present application establishes a unified 3T MRI multi-sequence acquisition protocol, and uses a standard phantom to calibrate the equipment every month to ensure that the apparent diffusion coefficient is still comparable under the conditions of cross-center and cross-time, and introduces a scoring mechanism to control data noise and deviation from the source, so that the subsequent image feature extraction is based on reliable foundation, solves the problem of multi-center MRI image heterogeneity, realizes standardized data input, and guarantees the generalization ability of the model.

[0034] (1) The present application draws the tumor region independently, introduces the intraclass correlation coefficient ICC to quantitatively evaluate the consistency, and only when the ICC is up to standard, the intersection is taken as the internal region IT of the tumor to control the subjective deviation of the drawing;And then expand the PT based on the reliable IT to cover the common immune cell infiltration zone of MSI tumor, organically combine clinical standardization, statistical verification and biological cognition, not only guarantee the reliability of subsequent image feature extraction, but also provide a solid foundation for obtaining feature difference, which can improve the credibility and clinical applicability of the whole model to a certain extent.

[0035] (2) This scheme extracts biological sensitive features highly related to MSI status from multi-sequence MRI, including effective diffusion diameter Deff and its characteristic difference ΔDeff, which directly reflect the size of intercellular space and the gradient of immune microenvironment. On the other hand, by fusing the main constraint and two groups of sub-constraints, the original high-dimensional features are compressed to the required number while retaining key discriminative information, balancing the accuracy, robustness and feasibility of clinical deployment of the model, and avoiding high-dimensional overfitting or unusable situations.

[0036] (3) This scheme adopts a one-level and two-level risk stratification strategy linkage, avoiding the traditional grouping which only relies on the overall probability and does not consider the microenvironment heterogeneity, to a certain extent, reducing the risk of misjudgment and ensuring the effectiveness of the prediction model after construction.

[0037] (4) This scheme constructs a dual-region fusion logistic regression model for IT and PT information on the one hand, and uses the key physical features derived from OGSE-DWI and their characteristic differences to capture the biological nature of MSI tumor intercellular space and active immune microenvironment. On the other hand, by reusing the feature difference, it is used as a dynamic correction basis in the risk stratification stage, not only for model input to improve prediction accuracy, but also for reclassification of patients in the high-risk group in the two-level risk stratification strategy, avoiding misjudgment caused by relying only on probability threshold, and simultaneously enhancing the biological rationality of the model and the accuracy of prognosis stratification. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The figure is a flowchart of the method for constructing the MSI prediction model for rectal cancer based on artificial intelligence in the present application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0040] Please refer to Figure 1 The present embodiment provides a method for constructing an MSI prediction model for rectal cancer based on artificial intelligence. The goal of this method is to accurately determine whether the rectal cancer belongs to the microsatellite instability type (MSI) by using only the magnetic resonance imaging (MRI) examination performed on the patient, without surgery or biopsy. MSI type tumors have a good response to immunotherapy, but a poor response to traditional chemotherapy. Therefore, early non-invasive identification of MSI status has significant clinical value. To achieve this method, the following steps are proposed:

[0041] S1, collect multi-center MRI data and perform data quality control processing; wherein the multi-center MRI data is corresponding image data under a plurality of scanning sequences; the image data is imaging parameters under each scanning sequence;

[0042] To reduce image heterogeneity between different centers, the same brand MRI scanners of different centers use standardized image acquisition schemes when collecting image data. Each research subject receives at least one MRI examination during the research process. The first examination is performed within 1 week before immunotherapy. The bowel is cleaned the night before the MRI examination;

[0043] An MRI imaging device is used when collecting data. The device uses a 3T MRI scanner, multi-channel body phased array coil parallel technology to collect signals, and at least includes T2WI, OGSE-DWI, DWI and TIWI+C scanning sequences. The main imaging parameters are as follows: T2WI is a high-resolution T2WI sequence, which uses a small FOV fast spin echo sequence, covers the entire rectal cancer area, FOV=20cm*20cm, matrix=310*320, repetition time TR=5990ms, echo time TE=1001ms, slice thickness=3.0mm, interval=0mm, number of excitations NEX=2; OGSE-DWI is a TDD-MRI sequence, b value=0 / 350 / 750s / mm 2 , FOV=32cm*26cm; matrix=94*128, TR=5000ms, TE=598ms, slice thickness=4mm, interval=0, NEX=1 / 3; DWI is a conventional diffusion weighted imaging sequence, corresponding to b value=1000s / mm

[0044] In addition, clinical data and follow-up data are also obtained synchronously when creating the relevant files of the patients;

[0045] Among them, the clinical data collection is personal information: age, gender, underlying disease and family history, etc.; clinical pathology and laboratory information: pre-treatment laboratory examination indexes, that is, serum CEA, CA19-9 level, tumor pathological type, tissue differentiation degree, results related to immunohistochemical staining and gene examination of prognostic molecular markers, clinical MRI staging and clinical staging, perineural invasion, lymphatic vessel invasion, tumor deposition and circumferential margin, prognosis, that is, recurrence, death time and cause; MSI typing evaluation can be selected by two pathologists with more than 10 years of pathological experience according to the guide standard, if there is a difference, a consensus discussion can be carried out; the follow-up data collection includes a. Follow-up plan: follow-up through outpatient records and telephone follow-up; the follow-up plan is once every 3 months in the first 2 years after treatment, then once every 6 months in the 2-3 years after treatment, the outpatient follow-up content includes physical examination and monitoring of tumor markers, based on follow-up monitoring of recurrence and metastasis of rectal cancer, colonoscopy is performed 1 year after treatment; high-risk patients image examination includes but is not limited to chest CT, abdominal MRI examination, brain CT or MRI examination, bone scan examination, neck ultrasound examination, etc.; when there is suspicious recurrence or metastasis sign, pathological biopsy or PET / CT is used for diagnosis; b. Follow-up endpoint: each case is followed up for at least 2 years; c. Research endpoint: progression-free survival is defined as the time from the date of surgery to tumor progression (in any way, including local / regional recurrence, distant metastasis) or (for any reason) death; the deletion is defined as the follow-up is lost or the follow-up is cut off, and the censored value is defined as the time between the date of surgery and the date of follow-up is lost or follow-up is cut off.

[0046] The process of data quality control processing is as follows:

[0047] S101, performing a scoring mechanism on image data, and screening out image data not meeting the scoring standard;

[0048] S102, using a standard phantom to calibrate the equipment every month to ensure that the apparent diffusion coefficient ADC values collected by different hospitals and at different times are comparable;

[0049] S103, image de-identification processing to protect patient privacy.

[0050] Among them, the scoring mechanism performed in S101 can be specifically taken in the following manner: 5-point scoring system; wherein, 1 = unable to diagnose, 5 = perfect image, 2, 3 and 4 are respectively close to the perfect image in turn, and the mean value of the scores is used as the final score, only the image with a final score ≥ 4 is included in the study; therefore, the scoring standard is whether ≥ 4.

[0051] The traditional magnetic resonance equipment, scan parameters and operation habits used by different hospitals are different, which leads to the difference of image features, and affects the stability of the configured model. The present scheme formulates a unified 3T MRI multi-sequence acquisition protocol, and calibrates the equipment every month using a standard phantom, to ensure that the apparent diffusion coefficient is still comparable under cross-center and cross-time conditions. A scoring mechanism is introduced to control data noise and bias from the source, so that the subsequent image feature extraction is based on a reliable foundation, solves the problem of multi-center MRI image heterogeneity, realizes standardized data input, and ensures the model generalization ability.

[0052] S2, based on the multi-center MRI data after data quality control processing, a definition verification mechanism is executed to generate the intratumoral region IT and the peritumoral region PT; wherein the specific content of the executed definition verification mechanism is as follows:

[0053] S201, on the imaging corresponding to T2WI, at least two preliminary tumor regions are delineated according to the known rectal cancer MRI report template, and the related state data is recorded; the two preliminary tumor regions are used in the present embodiment; S202, the intraclass correlation coefficient ICC is calculated according to the two preliminary tumor regions; S203, the intraclass correlation coefficient ICC is compared with the preset standard threshold IC_th, when ICC≥IC_th, it is determined that the consistency is good, and the intersection of the two preliminary tumor regions is taken as the final region, i.e. the intratumoral region; when ICC<IC_th, arbitration processing is performed, and the intratumoral region is obtained by expert judgment; S204, after the intratumoral region is determined, the boundary of the intratumoral region is extracted and expanded outward by M to form the peritumoral region; wherein M is a number greater than 0, and the unit is millimeter, M=5mm in the present embodiment;

[0054] Specifically, in S201, the known rectal cancer MRI report template can adopt the internationally recognized template, that is, the ESMO guideline, and the two groups of preliminary tumor regions drawn do not include necrosis, hemorrhage or intestinal cavity contents; the recorded related state data includes: the distance of the tumor from the anal margin, whether it invades the mesorectal fascia (mrCRM state), whether there is extramural vascular invasion (mrEMVI), and lymph node staging (N staging); in S202, when calculating the intragroup correlation coefficient ICC, the sum of the variance of the difference between the two groups of drawings and the error variance of multiple drawings of the same region is taken as the denominator part, and the variance of the difference between the two groups of drawings is taken as the numerator part, and the ratio obtained is the intragroup correlation coefficient ICC; it should be noted that the variance of the difference between the two groups of drawings is calculated by analysis of variance ANOVA to calculate the inter-group variance of the volume drawn by different doctors for the same group of tumors, reflecting the systematic difference between doctors; the error variance of multiple drawings of the same region is calculated by repeating the drawing of part of the cases by each doctor, and the variance of the repeated measurement is averaged, or estimated by the residual mean square of ANOVA, reflecting the random error of the operation; the principle of calculating the intragroup correlation coefficient ICC is: the total variation is decomposed into the real difference between doctors and the random error of the doctor himself, and the proportion of the former to the total variation is used to measure the consistency proportion, and the closer the ICC is to 1, the more reliable the drawing result is; the peritumoral region obtained in S204 usually contains the immune cell infiltration zone, and MSI tumors often show stronger inflammatory response in this region.

[0055] In the above scheme, the internationally recognized ESMO rectal cancer MRI report template is adopted, and the tumor region is independently drawn, the intragroup correlation coefficient ICC is introduced to quantitatively evaluate the consistency, and only when the ICC meets the standard, the intersection is taken as the IT, and the subjective deviation of the drawing is controlled; and then the PT is generated by expanding the reliable IT, covering the immune cell infiltration zone of MSI tumor, the whole scheme combines clinical specification, statistical verification and biological cognition organically, solves the key problems in traditional research such as arbitrary ROI definition, poor repeatability and lack of peritumoral region, not only guarantees the reliability of subsequent image feature extraction, but also lays an anatomical foundation for capturing tumor-microenvironment interaction signals such as feature difference, which can improve the credibility and clinical applicability of the whole model to a certain extent.

[0056] S3, performing feature extraction actions according to the tumor internal region IT and the peritumoral region PT, obtaining original features while calculating key physical features and feature difference values; wherein the performed feature extraction actions are as follows:

[0057] S301, performing feature extraction from eight region-sequence combinations to obtain original features;

[0058] Among them, eight area-sequence combinations are: T2WI, OGSE-DWI, DWI and TIWI+C of the internal tumor region IT, and T2WI, OGSE-DWI, DWI and TIWI+C of the peritumoral region PT; the obtained original features include: first-order statistical features, shape features, texture features and high-order features after wavelet transform; the first-order statistical features include: mean, standard deviation and entropy, etc.; the shape features are, for example, volume, sphericity, etc.; the texture features include, for example, gray level co-occurrence matrix GLCM and gray level run length matrix GLRLM, etc.; in addition, open source tools are used for feature extraction, and the open source tool selected in the embodiment is PyRadiomics, which complies with the international imageomics standard IBSI 1.0; the number of original features is much greater than 500;

[0059] S302, in response to the OGSE-DWI sequence, the product of the apparent diffusion coefficient ADC(t) obtained at a certain diffusion time t, the geometric factor and the reciprocal of pi is processed by taking the square root, and the result is the effective diffusion diameter Deff as the key physical feature; wherein the result is a physical quantity directly reflecting the size of the intercellular space, representing the equivalent space size in which water molecules can move freely; the intercellular space is simplified as a circular channel, and the geometric factor at this position is generated when the square of the circular radius of the cross section of the circular channel is converted to the diameter, and the geometric factor in the embodiment is 4; when water molecules diffuse in a circular channel, the diffusion speed is directly related to the diameter of the channel, and MSI tumors have loose cell arrangement and large intercellular space, so the effective diffusion diameter Deff is high, and MSS tumors are dense, so the effective diffusion diameter Deff is low;

[0060] S303, subtract the effective diffusion diameter Deff corresponding to the peritumoral region PT from the effective diffusion diameter Deff corresponding to the internal tumor region IT, and the difference value is the feature difference ΔDeff, which is used to capture the gradient change of the immune microenvironment.

[0061] The use of the above scheme can systematically extract high-dimensional imageomics features from the multi-sequence MRI of IT and PT, comprehensively depict the tumor structure and microenvironment; at the same time, the effective diffusion diameter Deff and its feature difference with clear biological significance are obtained, the size of the intercellular space and the gradient of immune infiltration are quantified, the problem that traditional imageomics only relies on statistical texture and lacks interpretable physical basis is solved, the AI feature is no longer a black box, but is closely related to the pathological nature of MSI; the effective diffusion diameter Deff and its feature difference are used as discriminant features to improve the prediction performance, and also provide correction basis for subsequent risk stratification, realizing the synergistic effect of physical priori and data-driven.

[0062] S4, under the condition of receiving the original features, a multi-constraint driven feature selection strategy is triggered to screen the target features;

[0063] Wherein, the basis of triggering the multi-constraint driven feature selection strategy is as follows:

[0064] S401, execute the main constraint action, select Q features most relevant to the MSI state using the minimum redundancy maximum relevance (mRMR) method, and then use LASSO regression for further compression to obtain Q / 5 features;

[0065] Wherein, the value of Q is 500 in this embodiment, so Q / 5 = 100;

[0066] S402, execute the sub-constraint A, construct a feature stability screening model, complete the first filtering action according to the output stability score, and retain the original features with a stability score exceeding a preset stability threshold;

[0067] Wherein, the basis for constructing the feature stability screening model is: ;

[0068] In the formula, S i represents the stability score of the i-th feature, which measures the consistency of the feature expression in different image sequences and regions, and N represents the total number of feature sources. In this embodiment, N represents the number of region-sequence combinations, i.e. N = 8, fi i (a) represents the value of the i-th feature in the a-th region-sequence combination, and b and a have the same meaning, except that a≠b, and p represents the Pearson correlation coefficient, which measures the degree of linear correlation between two variables; It should be noted that the sub-constraint A executed above adopts the robustness priority principle in feature engineering. High-dimensional imageomics features are easily affected by noise, scanning parameters, and delineation errors. Only retaining features stable across conditions can improve model generalization. Using the average absolute correlation can avoid the offset of positive and negative correlations, and emphasizes that regardless of direction, as long as the change trend is consistent, it is considered stable; The preset stability threshold in the first filtering action can be set to 0.6 in this embodiment, to ensure that it is consistent in different images and regions.

[0069] S403, execute the sub-constraint B, measure the extraction time and memory occupancy of each original feature retained after sub-constraint A on the configured edge device, and perform weighted calculation and processing according to the number of original features retained after sub-constraint A, extraction time and memory occupancy, output the comprehensive cost index, and perform the second filtering action, retain the features corresponding to the maximum value of the preset acceptable cost, and use them as the final target features;

[0070] Wherein, the edge device is configured in the construction system matched with the construction method; it should be noted that the weighted summation method is adopted in S403 to fuse multiple conflict targets into a single scalar as a sub-constraint B of edge computing friendly, ensuring that the final model can be run in real time on the local server of the hospital; in the second filtering action, the preset maximum acceptance cost can be set to 100 in this embodiment, which can make the final target feature not more than 50 in actual operation.

[0071] Through the above technical solution, biological sensitive features highly related to MSI status are extracted from multi-sequence MRI, especially the effective diffusion diameter Deff and its feature difference ΔDeff calculated based on OGSE-DWI, which directly reflect the size of intercellular space and the gradient of immune microenvironment; on the other hand, by fusing the main constraint and the two groups of sub-constraints, the original high-dimensional features are compressed to the required number while retaining the key discriminant information, balancing the accuracy, robustness and clinical deployment feasibility of the model, so that the overall scheme effectively solves the problems of traditional imageomics methods, such as being easily disturbed by noise, poor generalization ability and difficult to land in clinical practice, avoiding high-dimensional overfitting or unusable conditions.

[0072] S5, constructing a double-region fusion logistic regression model, inputting the texture features and shape features in the target features, the key physical features and their feature differences obtained in S3, and the double-region fusion logistic regression model outputting an MSI probability P(MSI) and triggering a verification strategy at the same time;

[0073] Wherein, the operation basis of the double-region fusion logistic regression model is:

[0074] The negative cumulative value of e is summed with 1, and the reciprocal of the sum is the MSI probability P(MSI); each input feature value is multiplied by a corresponding regression coefficient obtained by supervised learning of the double-region fusion logistic regression model, and the product is added to the intercept term to obtain the cumulative value; wherein, the formula for obtaining the cumulative value is: cumulative value = β0+ β1×x1+ β2×x2+....+ β k ×x k ; In the formula, β0 is the intercept term, β1, β2,..., β k : regression coefficients of each input feature; x1, x2,..., x k : input feature values, i.e. texture features and shape features in the target features, key physical features and their feature differences obtained in S3; among the key physical features, there are: effective diffusion diameter Deff corresponding to PT in the peritumoral region and effective diffusion diameter Deff corresponding to IT in the intratumoral region; it should be noted that the above double-region fusion logistic regression model realizes the simultaneous consideration of accuracy and interpretability.

[0075] The triggered verification strategy at least includes data division, cross-validation, and external verification.

[0076] The data division, cross-validation, and external verification form a three-level verification system. The data division divides the data into a training set, a verification set, and a test set in the order of 7:1:2, ensuring that the model development and final evaluation are separated. The cross-validation adopts 5-fold cross-validation in the training phase, that is, the training set is divided into 5 parts, and 4 parts are used for training and 1 part is used for verification in turn. After repeating 5 times, the AUC mean value and standard deviation are taken to reduce the random division bias. The external verification is to introduce an independent external verification queue, which can come from different hospitals or different devices, to evaluate the generalization ability of the double-region fusion logistic regression model in the real heterogeneous environment. The verification strategy given above takes into account the accuracy and interpretability, conforms to the medical AI model verification specification, such as the TRIPOD statement, thereby effectively preventing overfitting.

[0077] S6, according to the output MSI probability P(MSI), a primary risk stratification strategy is executed to preliminarily divide into different risk groups. When P(MSI) < 30%, it is divided into a low-risk group, indicating that it is likely to be MSS and suitable for traditional chemotherapy. When 30% ≤ P(MSI) < 70%, it is divided into a medium-risk group, indicating that it cannot be determined and needs to be further determined by gene detection and other means. When 70% ≤ P(MSI), it is divided into a high-risk group, indicating that it is likely to be MSI and considering immunotherapy.

[0078] Then the feature difference ΔDeff in S3 is introduced for dynamic correction processing:

[0079] The target preliminarily divided into a medium-risk group is extracted, and the feature difference ΔDeff is used to draw a ROC curve for predicting MSI on the verification set, calculate the Youden index, and take the maximum value of the feature difference ΔDeff as the significant threshold D_Δ, for example: significant threshold D_Δ = 1 μm; and a secondary risk stratification strategy is executed: when ΔDeff ≥ D_Δ, it is upgraded to a high-risk group; when ΔDeff ≤ -D_Δ, it is downgraded to a low-risk group; otherwise, it continues to be kept as a medium-risk group. The above-mentioned primary and secondary risk stratification strategies are linked to avoid the traditional grouping relying only on the overall probability without considering the microenvironment heterogeneity, thereby reducing the risk of misjudgment to a certain extent and ensuring the effectiveness of the prediction model after construction.

[0080] By adopting the technical scheme, on one hand, a double-region fusion logistic regression model for responding to IT and PT information is constructed, and the key physical characteristics derived from OGSE-DWI and the feature difference value thereof are used to capture the biological nature of large MSI tumor cell gap and active immune microenvironment; on the other hand, the feature difference value is reused as a dynamic correction basis in the risk stratification stage, which is not only used for model input to improve prediction accuracy, but also used for reclassification of the medium-risk group patients in the secondary risk stratification strategy, so as to avoid misjudgment caused by only relying on the probability threshold, and make the feature difference value serve as a discriminant factor and a calibration tool.

[0081] In addition, survival analysis verification can also be added after S6:

[0082] Collect target, that is, subsequent overall survival OS data of patients, draw Kaplan-Meier survival curve; expected result: the survival period of the high-risk group after receiving immunotherapy is significantly longer than that of other groups; even if not treated, MSI-like patients can have better prognosis due to strong tumor immunogenicity; use Log-rank test to determine whether the difference between groups is significant, and use Cox proportional hazards model to correct confounding factors such as age and TNM stage; since the above scheme is not the core scheme of the embodiment, it will not be described here; the predicted result is linked with the real clinical outcome to prove its actual value; the classification of patients is not directly for the patients themselves but for the processing of their corresponding data or information.

[0083] The above embodiments can be realized wholly or partially by software, hardware, firmware or any combination thereof. When realized by software, the above embodiments can be realized wholly or partially in the form of a computer program product. Those skilled in the art can realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized by hardware or software depends on the specific application and design constraints of the technical scheme.

[0084] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, which can be located in one place or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiment according to actual needs.

[0085] The above is merely specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for constructing an artificial intelligence-based MSI prediction model for rectal cancer, characterized in that, The method includes: Collect multi-center MRI data and perform data quality control processing; Among them, multicenter MRI data consists of image data corresponding to several scanning sequences; Based on multicenter MRI data, a definition validation mechanism is executed to generate the tumor interior region and the peritumoral region; Feature extraction is performed based on the tumor's internal and peritumoral regions, acquiring the original features while simultaneously calculating key physical features and their feature differences. Upon receiving the original features, a multi-constraint-driven feature selection strategy is triggered to filter and obtain the target features; Construct a dual-region fusion logistic regression model, inputting texture features and shape features, key physical features and their feature differences from the target features, and outputting the MSI probability, while simultaneously triggering a verification strategy; Based on the MSI probability, a first-level risk stratification strategy is implemented to initially divide different risk groups. Feature differences are introduced for dynamic correction: targets initially classified as medium-risk groups are extracted, significance thresholds are calculated based on feature differences, and a second-level risk stratification strategy is implemented to determine whether to adjust the medium-risk groups.

2. The method for constructing an artificial intelligence-based rectal cancer MSI prediction model according to claim 1, characterized in that: The image data consists of imaging parameters for each scan sequence; The scanning sequences include at least: T2WI, OGSE-DWI, DWI, and TIWI+C.

3. The method for constructing an artificial intelligence-based rectal cancer MSI prediction model according to claim 1, characterized in that: The data quality control process is as follows: a scoring mechanism is applied to the image data to filter out image data that does not meet the scoring criteria; the equipment is periodically calibrated using a standard phantom; and image de-identification processing is performed.

4. The method for constructing an artificial intelligence-based rectal cancer MSI prediction model according to claim 2, characterized in that: The definition verification mechanism is implemented as follows: On T2WI imaging, at least two preliminary tumor regions are delineated based on the known rectal cancer MRI report template, and relevant status data are recorded; the intra-group correlation coefficient (ICC) is calculated based on the two preliminary tumor regions; the intra-group correlation coefficient (ICC) is compared with the preset standard threshold (IC_th); when ICC ≥ IC_th, the intersection of the two preliminary tumor regions is taken as the final region, i.e., the tumor internal region (IT); after determining the tumor internal region, the boundary of the tumor internal region is extracted and extended outward (M) to form the peritumoral region (PT).

5. The method for constructing an artificial intelligence-based rectal cancer MSI prediction model according to claim 2, characterized in that: The feature extraction actions performed are as follows: Feature extraction was performed from each of the eight region-sequence combinations to obtain the original features; Among them, the eight region-sequence combinations are: T2WI, OGSE-DWI, DWI and TIWI+C of the tumor internal region IT, and T2WI, OGSE-DWI, DWI and TIWI+C of the peritumoral region PT; the original features include at least: first-order statistical features, shape features, texture features and higher-order features after wavelet transform. For OGSE-DWI sequences, based on the apparent diffusion coefficient ADC(t) obtained at a specific diffusion time t, the square root of the product of the apparent diffusion coefficient ADC(t), the geometric factor, and the reciprocal of pi is taken to obtain the effective diffusion diameter Deff, which is a key physical feature. The effective diffusion diameter Deff corresponding to PT in the peritumoral region is subtracted from the effective diffusion diameter Deff corresponding to IT in the internal tumor region. The resulting difference is the characteristic difference ΔDeff.

6. The method for constructing an artificial intelligence-based rectal cancer MSI prediction model according to claim 1, characterized in that: The criteria for triggering the multi-constraint-driven feature selection strategy are as follows: Execute the main constraint action, use the minimum redundancy maximum correlation mRMR method to select Q features most relevant to the MSI state, and then use LASSO regression to compress them to obtain Q / 5 features; Execute sub-constraint A, construct a feature stability screening model, and complete the first filtering action based on the output stability score, retaining the original features whose stability scores exceed the preset stability threshold; Execute sub-constraint B on the configured edge device, measure the extraction time and memory usage of each original feature retained after sub-constraint A, perform weighted calculation based on the number of original features retained after sub-constraint A, extraction time and memory usage, output comprehensive cost index, and after a second filtering action, retain the features corresponding to the comprehensive cost index that do not exceed the preset maximum acceptable cost, and use them as the final target features.

7. The method for constructing an artificial intelligence-based rectal cancer MSI prediction model according to claim 1, characterized in that: The operating principle of the dual-region fusion logistic regression model is as follows: The MSI probability P(MSI) is obtained by summing the negative cumulative value of e to the power of 1 and taking the reciprocal of the summation result; where, the cumulative value is obtained by multiplying each input feature value by a corresponding regression coefficient obtained by supervised learning from the dual-region fusion logistic regression model and adding the resulting product to the intercept term.

8. The method for constructing an artificial intelligence-based rectal cancer MSI prediction model according to claim 1, characterized in that: The triggered validation strategies include at least: data partitioning, cross-validation, and external validation; where data partitioning is performed by dividing the data into training, validation, and test sets in a 7:1:2 ratio; cross-validation is performed using 5-fold cross-validation during the training phase; and external validation is performed by introducing an external validation queue.

9. The method for constructing an artificial intelligence-based rectal cancer MSI prediction model according to claim 1, characterized in that: In the first-level risk stratification strategy, when P(MSI) < 30%, it is classified as a low-risk group; when 30% ≤ P(MSI) < 70%, it is classified as a medium-risk group; and when 70% ≤ P(MSI), it is classified as a high-risk group.

10. The method for constructing an artificial intelligence-based rectal cancer MSI prediction model according to claim 9, characterized in that: In the dynamic correction process, the significance threshold D_Δ is calculated based on the feature difference ΔDeff as follows: the ROC curve of the predicted MSI is plotted on the validation set based on the feature difference ΔDeff, the Youden index is calculated, and the feature difference ΔDeff corresponding to the maximum value is taken as the significance threshold D_Δ; and a two-level risk stratification strategy is implemented: when ΔDeff ≥ D_Δ, it is upgraded to the high-risk group; when ΔDeff ≤ -D_Δ, it is downgraded to the low-risk group; otherwise, it remains in the medium-risk group.

Images