Artificial Intelligence-Based Method for Constructing a Rectal Cancer MSI Prediction Model

Abstract

This invention discloses a method for constructing an artificial intelligence-based predictive model for rectal cancer MSI, belonging to the field of artificial intelligence technology. The method includes: acquiring multi-center MRI data and performing data quality control processing; and executing a definition verification mechanism based on the multi-center MRI data. Its key technical points are: constructing a dual-region fusion logistic regression model to address both IT and PT information; utilizing key physical features derived from OGSE-DWI and their feature differences to capture the biological essence of MSI tumors, such as large intercellular spaces and an active immune microenvironment; and reusing feature differences as a dynamic correction basis in the risk stratification stage, not only for model input to improve prediction accuracy but also for reclassifying intermediate-risk patients in the secondary risk stratification strategy, avoiding misjudgments caused by relying solely on probability thresholds, and simultaneously enhancing the model's biological rationality and prognostic stratification accuracy.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method for constructing an artificial intelligence-based rectal cancer MSI prediction model. Background Technology

[0002] Artificial intelligence (AI) is a technology that uses computers to simulate human intelligent behavior, including abilities such as learning, reasoning, recognition, and decision-making. Current mainstream methods are centered on machine learning, which has achieved significant results in areas such as image recognition, natural language processing, and medical diagnosis. It can automatically extract patterns from large amounts of data to predict and judge unknown samples. In the medical field, it is also used in conjunction with imaging technology. High-resolution pelvic magnetic resonance imaging (MRI) is the preferred examination method for preoperative non-invasive staging of rectal cancer, a consensus in domestic and international guidelines for rectal cancer diagnosis and treatment. Multimodal MRI technology can display the location, morphology, and depth of tumor invasion of rectal cancer to assess its T-staging. It can also identify poor prognostic factors such as extramural vascular invasion and positive circumferential resection margins. Clinically, surgical management strategies are usually formulated based on the MRI staging of rectal cancer.

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

[0004] First, the ADC values ​​in multicenter MRI data drift due to differences in equipment models and scanning parameters. For example, the ADC measured at Hospital A is 1.2 × 10⁻⁶. -3 mm² / s, but the same patient in Hospital B had a rate of 0.9 × 10⁻⁶. -3 The speed of mm² / s makes the model difficult to generalize; secondly, tumor region delineation is highly dependent on the subjective experience of different doctors, and the differences between delineations lead to more significant differences in subsequent feature extraction results. Moreover, traditional methods often ignore the peritumoral region and cannot reflect the immune microenvironment; thirdly, the features corresponding to the conventional imaging group have large redundancy and high dimensionality, and are also susceptible to noise interference. For example, most of the hundreds or thousands of texture features are unrelated to MSI, leading to overfitting; in addition, the existing risk stratification only uses probability as a one-size-fits-all approach. For example, the target patients with a traditional MSI probability of 68% are classified as uncertain, without utilizing relevant evidence such as peritumoral loosening, resulting in a high misjudgment rate. These problems together restrict the effective application of AI models from research to actual clinical practice. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for constructing an artificial intelligence-based rectal cancer MSI prediction model, the method comprising:

[0007] Collect multi-center MRI data and perform data quality control processing;

[0008] Among them, multicenter MRI data consists of image data corresponding to several scanning sequences;

[0009] Based on multicenter MRI data, a definition validation mechanism is executed to generate the tumor interior region and the peritumoral region;

[0010] 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.

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

[0012] 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;

[0013] 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.

[0014] Furthermore, the image data consists of imaging parameters for each scanning sequence;

[0015] The scanning sequences include at least: T2WI, OGSE-DWI, DWI, and TIWI+C.

[0016] Furthermore, the data quality control process involves: performing a scoring mechanism on the image data to filter out image data that does not meet the scoring criteria; periodically calibrating the equipment using a standard phantom; and image de-identification processing.

[0017] Furthermore, 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), and 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).

[0018] Furthermore, the feature extraction actions performed are as follows:

[0019] Feature extraction was performed from each of the eight region-sequence combinations to obtain the original features;

[0020] 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.

[0021] 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.

[0022] 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.

[0023] Furthermore, the basis for triggering the multi-constraint-driven feature selection strategy is as follows:

[0024] 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;

[0025] 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;

[0026] 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.

[0027] Furthermore, the operational basis of the dual-region fusion logistic regression model is as follows:

[0028] 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.

[0029] Furthermore, 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 introduces an external validation queue.

[0030] Furthermore, in the implemented primary risk stratification strategy, when P(MSI) < 30%, the group is classified as low-risk; when 30% ≤ P(MSI) < 70%, the group is classified as medium-risk; and when 70% ≤ P(MSI), the group is classified as high-risk.

[0031] Furthermore, in the dynamic correction process, the basis for calculating the significance threshold D_Δ based on the feature difference ΔDeff is 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 its 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.

[0032] This invention provides a method for constructing an artificial intelligence-based rectal cancer MSI prediction model, which has the following beneficial effects:

[0033] (1) This scheme ensures that the apparent diffusion coefficient remains comparable across centers and time by establishing a unified 3T MRI multi-sequence acquisition protocol and using standard phantoms to calibrate the equipment monthly. It also introduces a scoring mechanism to control data noise and bias from the source, so that subsequent radiomics feature extraction is based on a reliable foundation, solves the problem of heterogeneity correlation of multicenter MRI images, achieves standardized data input, and ensures the model's generalization ability.

[0034] (1) This scheme independently delineates the tumor region and introduces intragroup correlation coefficient (ICC) to quantitatively assess consistency. Only when the ICC meets the standard is the intersection taken as the internal region IT of the tumor, thus controlling the subjective bias of the delineation. Then, based on the reliable IT, it expands outward to generate PT, covering the immune cell infiltration zone commonly found in MSI tumors. This organically combines clinical norms, statistical validation and biological knowledge, which not only ensures the reliability of subsequent radiomics feature extraction, but also provides a solid foundation for obtaining feature differences, which can improve the credibility and clinical applicability of the entire model to a certain extent.

[0035] (2) This scheme extracts biologically sensitive features that are highly correlated with MSI status from multiple sequence MRI, including the effective diffusion diameter Deff and its feature difference ΔDeff, which directly reflects the size of the intercellular space and the gradient of the immune microenvironment. On the other hand, by fusing the master constraint and two sets of sub-constraints, the original high-dimensional features are compressed to the required number while retaining key discriminant information, taking into account the balance between model accuracy, robustness and clinical deployment feasibility, and avoiding high-dimensional overfitting or unusability.

[0036] (3) This scheme adopts a linkage of primary and secondary risk stratification strategies to avoid the traditional grouping that only relies on the overall probability and does not consider the heterogeneity of the microenvironment. This reduces the risk of misjudgment to a certain extent and ensures the effectiveness of the prediction model after it is built.

[0037] (4) On the one hand, this scheme constructs a dual-region fusion logistic regression model to deal with IT and PT information. It uses key physical features derived from OGSE-DWI and their feature differences to capture the biological essence of MSI tumor cells with large gaps and active immune microenvironment. On the other hand, by reusing feature differences, it serves as a dynamic correction basis in the risk stratification stage. It is not only used as model input to improve prediction accuracy, but also used to reclassify patients in the intermediate-risk group in the secondary risk stratification strategy to avoid misjudgment caused by relying solely on probability thresholds, and simultaneously enhances the biological rationality of the model and the accuracy of prognostic stratification. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the method for constructing an artificial intelligence-based rectal cancer MSI prediction model in this invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figure 1 This embodiment provides a method for constructing an artificial intelligence-based MSI prediction model for rectal cancer. The goal of this method is to accurately determine whether a patient's rectal cancer belongs to the microsatellite instability (MSI) type, without surgery or biopsy, solely based on the patient's MRI examinations. MSI tumors respond well to immunotherapy but poorly to traditional chemotherapy; therefore, early non-invasive identification of MSI status has significant clinical value. The following steps are proposed to achieve this method:

[0041] S1. Acquire multi-center MRI data and perform data quality control processing; where multi-center MRI data consists of image data corresponding to several scanning sequences; the image data is the imaging parameters of each scanning sequence.

[0042] To reduce image heterogeneity between different centers, standardized image acquisition protocols were used for MRI scanners of the same brand at different centers when acquiring image data. Each participant underwent at least one MRI examination during the study, with the first examination performed within one week before immunotherapy. The bowel was cleansed the night before the MRI examination.

[0043] MRI imaging equipment was used for data acquisition. This equipment employed a 3T MRI scanner and multi-channel body phased array coil parallel acquisition technology. The scanning sequences included at least T2WI, OGSE-DWI, DWI, and TIWI+C. The main imaging parameters were as follows: T2WI, a high-resolution T2WI sequence, used a small FOV fast spin-echo sequence covering 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 was a TDD-MRI sequence with b-values ​​of 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 conventional diffusion-weighted imaging, i.e., DWI sequence, with a corresponding b-value=1000s / mm²; TIWI+C is TIWI enhancement sequence, using VIBE multi-phase enhancement scanning sequence technology, FOV=36cm×30cm, matrix=215×352, TR=3.81ms, TE=1.38ms, flip angle=12°, slice thickness=3mm, interval=0mm, the contrast agent is 0.1mmol / kg Gd-DTPA, injected via dorsal vein of the hand at a flow rate of 2.0ml / s, followed by a bolus injection of 20ml of normal saline at the same flow rate. Enhancement scanning is performed simultaneously with contrast agent injection, for a total of 3 phases, each phase lasting 16s;

[0044] In addition, clinical data and follow-up data are also acquired simultaneously when creating relevant patient records;

[0045] The clinical data collection includes personal information such as age, gender, underlying diseases, and family history; clinical pathology and laboratory information such as pre-treatment laboratory test indicators (e.g., serum CEA, CA19-9 levels), tumor pathological type, degree of tissue differentiation, immunohistochemical staining and gene testing results for prognostic molecular markers, clinical MRI staging and clinical staging, perineural invasion, lymphovascular invasion, tumor deposition and circumferential resection margins, and prognostic status (i.e., recurrence, time and cause of death); MSI classification assessment can be conducted by two pathologists with more than 10 years of experience, referring to guideline standards. Any disagreements will be discussed and agreed upon. Follow-up data collection includes: a. Follow-up plan: follow-up will be conducted through outpatient records and telephone follow-ups; the follow-up plan is once every 3 months for the first 2 years after treatment, then... Follow-up was conducted every 6 months for the 2nd-3rd year post-treatment. Outpatient follow-up included physical examination and monitoring of tumor markers. Recurrence and metastasis of rectal cancer were monitored during follow-up. Colonoscopy was performed 1 year after treatment. Imaging examinations for high-risk patients included, but were not limited to, chest CT, abdominal MRI, cranial CT or MRI, bone scan, and neck ultrasound. When suspicious signs of recurrence or metastasis were present, pathological biopsy or PET / CT was used for confirmation. b. Follow-up endpoint: Each case was followed up for at least 2 years. c. Study endpoint: Progression-free survival was defined as the time from the date of surgery to the occurrence of tumor progression (in any aspect, including local / regional recurrence, distant metastasis) or death (from any cause). Censoring was defined as loss to follow-up or follow-up cutoff. The truncation value was defined as the time from the date of surgery to the date of loss to follow-up or follow-up cutoff.

[0046] The data quality control process is based on the following:

[0047] S101. Implement a scoring mechanism on the image data and filter out image data that does not meet the scoring criteria;

[0048] S102. Use the standard phantom to calibrate the equipment monthly to ensure that the apparent diffusion coefficient ADC values ​​collected from different hospitals and at different times are comparable.

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

[0050] The scoring mechanism implemented in S101 can be implemented in the following way: a 5-point scoring system is used, where 1 = undiagnosable, 5 = perfect image, and 2, 3 and 4 are respectively close to perfect image. The average score is used as the final score. Only images with a final score ≥4 are included in the study; therefore, the scoring standard is whether it is ≥4.

[0051] Traditionally, the MRI equipment, scanning parameters, and operating habits used by different hospitals vary, resulting in incomparable image features and affecting the stability of the configured model. This solution establishes a unified 3T MRI multi-sequence acquisition protocol and uses standard phantoms to calibrate the equipment monthly, ensuring that the apparent diffusion coefficient remains comparable across centers and time periods. Furthermore, it introduces a scoring mechanism to control data noise and bias from the source, ensuring that subsequent radiomics feature extraction is based on a reliable foundation. This solves the problem of heterogeneity in multi-center MRI images, achieves standardized data input, and guarantees the model's generalization ability.

[0052] S2. Based on the multicenter MRI data after data quality control processing, a definition validation mechanism is executed to generate the intratumoral region IT and the peritumoral region PT; the specific content of the definition validation mechanism is as follows:

[0053] S201. On the T2WI imaging, based on the known rectal cancer MRI report template, delineate at least two preliminary tumor regions and record relevant status data; in this embodiment, two preliminary tumor regions are used. S202. Calculate the intra-group correlation coefficient (ICC) based on the two preliminary tumor regions. S203. Compare the intra-group correlation coefficient (ICC) 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 tumor interior region. When ICC < IC_th, arbitration is performed, and the tumor interior region is determined by expert evaluation. S204. After determining the tumor interior region, extract the boundary of the tumor interior region and extend it outward by M to form the peritumoral region; where M is a value greater than 0, and the unit is millimeters. In this embodiment, M = 5 mm.

[0054] Specifically, in S201, the known rectal cancer MRI report template can use the internationally recognized template, namely the ESMO guidelines. The two preliminary tumor regions delineated should not include necrosis, hemorrhage, or intestinal contents. Recorded relevant status data include: distance of the tumor from the anal verge, whether it invades the rectal mesofascia (mrCRM status), whether it has extramural vascular invasion (mrEMVI), and lymph node staging (N staging). In S202, when calculating the intraclass correlation coefficient (ICC), the variance of the difference between the two delineations and the variance of the error from multiple delineations of the same region are used. The sum of the variances of the differences between the two delineations and the variances of the error from multiple delineations of the same region is used as the denominator, and the variance of the differences between the two delineations is used as the numerator. The resulting ratio is the intraclass correlation coefficient (ICC). It should be noted that the variance of the two groups of delineation differences was calculated using ANOVA to measure the intergroup variance of the volume delineated by different doctors for the same group of tumors, reflecting the systematic differences among doctors; the variance of the error from multiple delineations of the same area was calculated by having each doctor repeatedly delineate a portion of the cases, averaging the variance of repeated measures, or estimated by the mean square of the residuals from ANOVA, reflecting operational random error; the principle of calculating the intragroup correlation coefficient (ICC) is as follows: the total variance is decomposed into the true difference between doctors and the random error of the doctors themselves, and the proportion of the former to the total variance is used to measure the consistency ratio. The closer the ICC is to 1, the more reliable the delineation results are; the peritumoral region obtained in S204 usually contains an area of ​​immune cell infiltration, and MSI tumors often show a stronger inflammatory response in this area.

[0055] The above scheme uses the internationally recognized ESMO rectal cancer MRI report template and independently delineates the tumor region. It introduces intragroup correlation coefficient (ICC) to quantify consistency, and only when the ICC meets the standard is the intersection taken as the internal tumor region IT, thus controlling the subjective bias of the delineation. Then, based on the reliable IT, it expands outward to generate PT, covering the immune cell infiltration zone commonly found in MSI tumors. The overall scheme organically combines clinical standards, statistical validation, and biological understanding, solving key problems in traditional studies such as arbitrary ROI definition, poor reproducibility, and missing peritumoral regions. It not only ensures the reliability of subsequent radiomics feature extraction, but also lays an anatomical foundation for capturing tumor-microenvironment interaction signals, such as feature differences, which can improve the credibility and clinical applicability of the entire model to a certain extent.

[0056] S3. Perform feature extraction based on the internal tumor region (IT) and the peritumoral tumor region (PT), obtaining the original features while calculating key physical features and their feature differences; the feature extraction actions performed are as follows:

[0057] S301. Extract features from the eight region-sequence combinations to obtain the original features;

[0058] The eight region-sequence combinations are: T2WI, OGSE-DWI, DWI, and TIWI+C for the tumor intratumoral region (IT), and T2WI, OGSE-DWI, DWI, and TIWI+C for the peritumoral region (PT). The resulting raw features include: first-order statistical features, shape features, texture features, and higher-order features after wavelet transform. First-order statistical features include: mean, standard deviation, and entropy; shape features include volume and sphericity; texture features include gray-level co-occurrence matrix (GLCM) and gray-level run-length matrix (GLRLM). Furthermore, open-source tools are used for feature extraction; in this embodiment, PyRadiomics is selected, which follows the international radiomics standard IBSI 1.0. The number of raw features is far greater than 500.

[0059] S302. 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 ADC(t), geometric factor, and the reciprocal of pi is taken, and the result is the effective diffusion diameter Deff, which is used as a key physical feature. This result is a physical quantity that directly reflects the size of the intercellular space, representing the equivalent space size in which water molecules can move freely. The intercellular space is simplified to a circular channel, and the geometric factor is generated when converting the square of the radius of the circular channel cross-section to the diameter. In this embodiment, the geometric factor is 4. When water molecules diffuse in a circular channel, their diffusion rate is directly related to the channel diameter. MSI tumors have a high effective diffusion diameter Deff because the cells are loosely arranged and the gaps are large, while MSS tumors are dense and therefore have a low effective diffusion diameter Deff.

[0060] S303. Subtract the effective diffusion diameter Deff corresponding to the IT region inside the tumor from the effective diffusion diameter Deff corresponding to the PT region around the tumor. The difference is the characteristic difference ΔDeff, which is used to capture the gradient changes of the immune microenvironment.

[0061] The above-mentioned approach can systematically extract high-dimensional radiomics features from multiple sequence MRI of IT and PT, comprehensively characterizing tumor structure and microenvironment. At the same time, it obtains the effective diffusion diameter Deff and its feature difference with clear biological significance, quantifies the size of intercellular spaces and immune infiltration gradient, and solves the problem that traditional radiomics relies only on statistical texture and lacks interpretable physical basis. This makes AI features no longer a black box, but closely related to the pathological nature of MSI. The effective diffusion diameter Deff and its feature difference not only serve as discriminative features to improve predictive performance, but also provide a correction basis for subsequent risk stratification, realizing the synergistic effect of physical prior and data-driven approach.

[0062] S4. Upon receiving the original features, trigger the multi-constraint driven feature selection strategy to filter and obtain the target features;

[0063] The criteria for triggering the multi-constraint-driven feature selection strategy are as follows:

[0064] S401. 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 further compress them to obtain Q / 5 features.

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

[0066] S402. 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.

[0067] The feature stability screening model is based on the following: ;

[0068] In the formula, S i The stability score of the i-th feature is used to measure the consistency of its expression across different image sequences and regions. N represents the total number of feature sources; in this embodiment, it represents the number of region-sequence combinations, i.e., N=8. i (a) This represents the value of the i-th feature in the a-region sequence combination. b has the same meaning as a, except that a≠b. ρ represents the Pearson correlation coefficient, used to measure the degree of linear correlation between two variables. It should be noted that the sub-constraint A executed above adopts the principle of robustness priority in feature engineering. High-dimensional image omics features are easily affected by noise, scanning parameters, and delineation errors. Only retaining cross-conditionally stable features can improve the generalization of the model. Using average absolute correlation can avoid the cancellation of positive and negative correlations. It is emphasized that as long as the trend of change is consistent, regardless of the direction, 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 performs consistently in different images and regions.

[0069] S403. 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 a comprehensive cost index. After a second filtering action, retain the features whose comprehensive cost index does not exceed the preset maximum acceptable cost and use them as the final target features.

[0070] The aforementioned edge devices are configured in the construction system that supports this construction method. It should be noted that in the above S403, a weighted summation method is used to merge multiple conflicting targets into a single scalar as an edge computing-friendly sub-constraint B, ensuring that the final model can run in real time on the hospital's local server. In the second filtering action, the preset maximum value of the acceptance cost can be set to 100 in this embodiment, which can ensure that the final target features do not exceed 50 in actual operation.

[0071] The above technical solution extracts biologically sensitive features highly correlated with MSI status from multi-sequence MRI, especially the effective diffusion diameter Deff and its feature difference ΔDeff calculated based on OGSE-DWI, which directly reflects the intercellular space size and immune microenvironment gradient. On the other hand, by fusing the principal constraint and two sub-constraints, the original high-dimensional features are compressed to the required number while retaining key discriminative information, balancing model accuracy, robustness, and clinical deployment feasibility. This overall solution effectively solves the problems of traditional radiomics methods being susceptible to noise interference, having poor generalization ability, and being difficult to implement clinically, avoiding high-dimensional overfitting or unusability.

[0072] S5. Construct a dual-region fusion logistic regression model, inputting texture and shape features from the target features, key physical features obtained in S3 and their feature differences. The dual-region fusion logistic regression model outputs the MSI probability P(MSI) and simultaneously triggers the verification strategy.

[0073] The operational basis of the dual-region fusion logistic regression model is as follows:

[0074] Summing the negative cumulative value of e to the power of 1 and taking the reciprocal of the sum gives the MSI probability P(MSI). Multiplying each input feature value by a corresponding regression coefficient obtained from supervised learning using a two-region fusion logistic regression model, and adding the resulting product to the intercept term, yields the cumulative value. The formula for 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 for each input feature; x1, x2, ..., x k The input feature values ​​are the texture and shape features in the target features, the key physical features obtained in S3 and their feature differences; the key physical features include the effective diffusion diameter Deff corresponding to the peritumoral region PT and the effective diffusion diameter Deff corresponding to the tumor internal region IT; it should be noted that the above dual-region fusion logistic regression model achieves both accuracy and interpretability.

[0075] The triggered validation strategies include at least: data partitioning, cross-validation, and external validation;

[0076] The aforementioned data partitioning, cross-validation, and external validation form a three-tiered validation system. Data partitioning involves dividing the data into training, validation, and test sets in a 7:1:2 ratio to ensure separation between model development and final evaluation. Cross-validation employs a 5-fold cross-validation approach during training, dividing the training set into five equal parts and alternating between four parts for training and one part for validation, repeating this process five times before calculating the mean and standard deviation of the AUC to reduce random partitioning bias. External validation involves introducing an independent external validation cohort, which can come from different hospitals or different devices, to assess the generalization ability of the dual-region fusion logistic regression model in real-world heterogeneous environments. Therefore, the validation strategy described above balances accuracy and interpretability, conforms to medical AI model validation standards such as the TRIPOD declaration, and effectively prevents overfitting.

[0077] S6. Based on the output MSI probability P(MSI), a first-level risk stratification strategy is implemented to initially divide the patient into different risk groups. When P(MSI) < 30%, the patient is classified as a low-risk group, indicating that it is very likely to be MSS and suitable for traditional chemotherapy. When 30% ≤ P(MSI) < 70%, the patient is classified as a medium-risk group, indicating that it is uncertain and further determination is needed through genetic testing and other means. When 70% ≤ P(MSI), the patient is classified as a high-risk group, indicating that it is very likely to be MSI and immunotherapy should be considered.

[0078] Then, the feature difference ΔDeff from S3 is introduced for dynamic correction:

[0079] Targets initially classified as medium-risk groups are extracted. Based on the feature difference ΔDeff, the ROC curve of the predicted MSI is plotted on the validation set. The Youden index is calculated, and the feature difference ΔDeff corresponding to its maximum value is taken as the significance threshold D_Δ, for example: significance threshold D_Δ = 1μm. A two-level risk stratification strategy is implemented: if ΔDeff ≥ D_Δ, it is upgraded to the high-risk group; if ΔDeff ≤ -D_Δ, it is downgraded to the low-risk group; otherwise, it remains in the medium-risk group. The above-mentioned use of primary and secondary risk stratification strategies in conjunction avoids the traditional grouping that only relies on overall probability and does not consider the heterogeneity of the micro-environment, thus reducing the risk of misjudgment to a certain extent and ensuring the effectiveness of the prediction model after its construction.

[0080] By adopting the above technical solutions, on the one hand, a dual-region fusion logistic regression model for dealing with IT and PT information is constructed. It utilizes key physical features derived from OGSE-DWI and their feature differences to capture the biological essence of MSI tumor cells, which have large intercellular spaces and an active immune microenvironment. On the other hand, by reusing feature differences, it serves as a dynamic correction basis in the risk stratification stage. It is not only used as model input to improve prediction accuracy, but also used to reclassify patients in the intermediate-risk group in the secondary risk stratification strategy, avoiding misjudgments caused by relying solely on probability thresholds. This makes feature differences both a discriminant factor and a calibration tool.

[0081] Additionally, survival analysis validation can be added after S6:

[0082] The target data collected was the patients' subsequent overall survival (OS) to plot Kaplan-Meier survival curves. Expected outcomes: the high-risk group, after receiving immunotherapy, had significantly longer survival than other groups; even without treatment, MSI-like patients may have a better prognosis due to strong tumor immunogenicity. The Log-rank test was used to determine the significance of differences between groups, and a Cox proportional hazards model was used to correct for confounding factors such as age and TNM stage. Since this approach is not the core of this embodiment, it will not be elaborated upon further. This section links the predicted results to actual clinical outcomes, demonstrating its practical value. Patient classification is not directly applied to the patients themselves but rather involves processing their corresponding data or information.

[0083] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0084] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0085] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this 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; The multicenter MRI data consists of image data corresponding to several scanning sequences; the image data consists of imaging parameters for each scanning sequence; the scanning sequences include at least: T2WI, OGSE-DWI, DWI, and TIWI+C; Based on multicenter MRI data, a definition verification mechanism is executed to generate the tumor interior region and the peritumoral region. The definition verification mechanism is as follows: on the T2WI imaging, according to the known rectal cancer MRI report template, at least two preliminary tumor regions are delineated 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 a 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 interior region (IT); after determining the tumor interior region, the boundary of the tumor interior region is extracted and extended outward (M) to form the peritumoral region (PT). 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; A dual-region fusion logistic regression model is constructed, taking texture and shape features, key physical features, and their feature differences as inputs from the target features, and outputting the MSI probability while simultaneously triggering a validation strategy. The operational basis 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; 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 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.

3. The method for constructing an artificial intelligence-based rectal cancer MSI prediction model according to claim 1, 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.

4. 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.

5. 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.

6. 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.

7. The method for constructing an artificial intelligence-based rectal cancer MSI prediction model according to claim 6, 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