Radioactive enteritis prevention method based on AI model and image fusion

Abstract

The invention relates to and discloses a radiation enteritis prevention method based on an AI model and image fusion, which aims at the difficulty in prediction and intervention of radiation enteritis after radiotherapy of abdominal and pelvic tumors, predicts to be incorporated into 100 radiotherapy patients through historical retrospective and prospective prediction, systematically collects multi-modal data, carries out image fusion on the multi-modal data, and carries out image fusion on the multi-modal data. Utilizing interpretable machine learning to fuse radiomics parameters, flora markers and clinical variables, constructing a radiation enteritis dynamic risk prediction model, and verifying the universality of the model through 200 historical queues; key flora / metabolic targets are further screened, a coprophilous fungi transplantation intervention scheme is designed, the regulation and control effect of coprophilous fungi is verified in 50 high-risk patients, an AI auxiliary decision-making system and a microecological preparation are finally developed, a full-chain management strategy of'image-omics prediction-targeted intervention-clinical transformation 'is formed, and an innovative solution is provided for precise prevention and control of radiotherapy complications.

Description

Technical Field

[0001] This invention relates to the field of technology, and more specifically, to a method for preventing radiation enteritis based on AI model and image fusion. Background Technology

[0002] Radiation enteritis is an intestinal complication caused by radiotherapy for malignant tumors in the pelvic cavity, abdominal cavity, or retroperitoneum. It can affect the small intestine, colon, and rectum, hence it is also called radiation proctitis, colonitis, or small enteritis. Based on the radiation dose, duration, and onset speed, radiation sickness is generally classified into acute and chronic types. It is also classified into external radiation sickness and internal radiation sickness based on the location of the radiation source. In the early stages, intestinal mucosal cell renewal is inhibited; later, the small arterial walls swell and become occluded, causing intestinal ischemia and mucosal erosion. In the later stages, intestinal wall fibrosis occurs, leading to intestinal stenosis or perforation, and the formation of abscesses, fistulas, and intestinal adhesions within the abdominal cavity.

[0003] Current methods for preventing radiation enteritis still have several limitations: In terms of radiotherapy techniques, while advancements in radiotherapy have reduced the volume of intestine exposed to radiation, the mobility of the small intestine and individual anatomical differences lead to inaccurate dose assessment, and existing bioequivalence dosing models may not be applicable to patients with comorbidities. Regarding pharmacological prevention, amifostine's use is limited due to its significant side effects and short half-life; there is a lack of consensus on the selection and efficacy of probiotic strains; and other protective agents such as sucralfate lack large-scale evidence support. Physical barrier technologies such as hydrogels are effective in pelvic radiotherapy but are difficult to apply to the small intestine and are costly. Furthermore, existing strategies lack consideration for individualized factors such as genetic susceptibility and differences in gut microbiota, and there is a lack of targeted synergistic protective measures during concurrent chemoradiotherapy. These shortcomings highlight the inadequacies of current preventative measures. Summary of the Invention

[0004] To overcome the aforementioned deficiencies in existing technologies, this invention provides a method for preventing radiation enteritis based on AI models and image fusion. Addressing the challenge of predicting and intervening in radiation enteritis after radiotherapy for abdominal and pelvic tumors, this invention involves a retrospective and prospective study of 100 radiotherapy patients. Multimodal data is systematically collected, and interpretable machine learning is used to fuse radiomics parameters, microbial biomarkers, and clinical variables to construct a dynamic risk prediction model for radiation enteritis. The model's universality is validated through a 200-patient historical cohort. Furthermore, key microbial / metabolic targets are screened, and a fecal microbiota transplantation (FMT) intervention is designed. Its regulatory effect is verified in 50 high-risk patients. Finally, an AI-assisted decision-making system and microbial preparations are developed, forming a comprehensive management strategy encompassing "image-omics prediction-targeted intervention-clinical translation," providing an innovative solution for the precise prevention and control of radiotherapy complications.

[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution: a method for preventing radiation enteritis based on AI model and image fusion, characterized by comprising the following steps:

[0006] S1. A retrospective cohort was constructed by selecting 200 patients with malignant tumors who had previously received abdominopelvic radiotherapy; a prospective cohort was constructed by selecting 100 new patients who underwent abdominopelvic radiotherapy. Baseline data were collected before radiotherapy, and the occurrence of radiation enteritis was dynamically tracked during radiotherapy and 3 months after treatment.

[0007] S2. Clinical data and CT images were collected from patients included in the retrospective cohort to delineate the gut microbiome sensitive region of interest (ROI); gut microbiota-metabolomics data were collected from patients included in the prospective cohort.

[0008] S3. Perform multi-omics feature screening and fusion modeling feature engineering, specifically including:

[0009] (1) Univariate analysis: Analysis of the texture entropy value of the high-dose area of ​​the intestine;

[0010] (2) Multifactor dimensionality reduction: Key image omics features were screened using the LASSO regression algorithm;

[0011] (3) Microbiome-metabolism network analysis: The correlation between microbiome and metabolites was calculated using the SParCC algorithm;

[0012] (4) Multimodal feature fusion: stacking traditional radiomics, depth features, microbial α diversity, butyrate concentration and clinical variables;

[0013] (5) Model construction: XGBoost, LightGBM and random forest algorithms were used to construct a multi-omics fusion model, and five-fold cross-validation was used to optimize the hyperparameters;

[0014] (6) Model validation;

[0015] S4. Mechanism analysis and intervention verification of gut microenvironment: causal pathways were constructed using structural equation modeling, and the interaction between gut microbiota functional genes and host metabolism was analyzed through MetaCyc pathway enrichment.

[0016] S5. Statistical Analysis: Validate differences between groups by testing continuous and categorical variables, and test model performance using ROC curves, calibration curves, and Hosmer-Lemeshow.

[0017] Furthermore, the intestinal dose-sensitive region ROI mentioned in step S2 includes a metrological gradient ROI and a target-adjacent ROI. The metrological gradient ROI includes the intestinal region covered by dose lines of 20Gy, 30Gy and 40Gy, and the target-adjacent ROI includes the intestinal segment extending 1cm beyond the PTV.

[0018] Furthermore, by processing time-series metabolite changes during radiotherapy using LSTM, combined with baseline characteristics, the risk of radiation enteritis was predicted.

[0019] Furthermore, in step S3, the generalization of the model is verified through a retrospective cohort; the performance of the traditional imaging model and the multi-omics fusion model is compared through the Delong test; and the clinical net benefit is evaluated through decision curve analysis.

[0020] Furthermore, the causal pathway described in step S4 is as follows: radiation dose, dysbiosis, metabolic disorder, Th17 / Treg imbalance, and radiation enteritis.

[0021] In summary, the present invention has the following beneficial effects:

[0022] (1) Multimodal fusion: For the first time, radiomics, microbiome-metabolomics and AI dynamic modeling are integrated to break through the bottleneck of single-dimensional prediction;

[0023] (2) Mechanism-intervention closed loop: from model prediction → FMT validation → product development, forming a complete chain of translational medicine;

[0024] (3) Strong clinical applicability: The AI ​​system directly connects to the radiotherapy planning system to achieve real-time linkage of "drawing-prediction-intervention";

[0025] (4) The method provided in this application uses cross-omics data to accurately analyze the mechanism of radiation enteritis, providing an innovative tool for personalized regulation of the gut microenvironment. Attached Figure Description

[0026] Figure 1 This is a flowchart of the radiation enteritis prevention method based on AI model and image fusion in an embodiment of the present invention;

[0027] Figure 2 This is a diagram showing the results of radiomics feature screening in the radiation enteritis prevention method based on AI model and image fusion in this embodiment of the invention.

[0028] Figure 3 This is a heatmap used to display differences in image group characteristics in the radiation enteritis prevention method based on AI model and image fusion in this embodiment of the invention;

[0029] Figure 4 This is a principal component analysis diagram of PCA in the radiation enteritis prevention method based on AI model and image fusion in an embodiment of the present invention. Detailed Implementation

[0030] The following is in conjunction with the appendix Figure 1-4 The present invention will be described in further detail below.

[0031] Example: First, cohort design and data collection were conducted to construct the cohort: Retrospective cohort (n=200): Patients with malignant tumors (colorectal cancer, cervical cancer, prostate cancer, etc.) who had previously received abdominopelvic radiotherapy (dose ≥45Gy) at our hospital were included, screened according to inclusion / exclusion criteria (age 18-80 years, no history of intestinal surgery, PS score ≤1, survival >3 months). Prospective cohort (n=100): Newly enrolled patients scheduled for abdominopelvic radiotherapy were included. Baseline data (feces, serum, imaging) were collected before radiotherapy, and the occurrence of radiation enteritis (RE) was dynamically tracked during radiotherapy and 3 months after treatment (CTCAE v4.0 classification, divided into RE 2+ group (≥2 grade) and RE 2- group (<2 grade)). Exclusion criteria: severe liver and kidney disease, immunodeficiency, use of antibiotics / probiotics in the past 3 months, missing follow-up data, or refusal of informed consent.

[0032] Multimodal data acquisition and preprocessing: Image data: CT image acquisition: Scanning was performed using [instrument model, such as Siemens SOMATOM Force], parameters [supplementary: 120kVp, slice thickness 3mm], delineating dose-sensitive areas of interest (ROIs) in the intestine (e.g., rectosigmoid junction, high-dose areas of the small intestine), including: dose gradient ROIs: intestinal regions covered by dose lines of 20Gy, 30Gy, 40Gy, etc.; target-adjacent ROIs: intestinal segments extending 1cm beyond the PTV. Consistency verification: ROIs were delineated by two radiologists using Pinnacle software, and intra-class / inter-class ICC was calculated (>0.75 indicates acceptable consistency). Omics data: Traditional radiomics: PyRadiomics was used to extract texture features (GLCM, GLSZM) and morphological features (volume, surface area). Deep radiomics: Pre-trained ResNet-50 was used to extract high-level semantic features (output from fully connected layers). Gut microbiota: 16S rRNA sequencing (V3-V4 region, QIIME2 analysis of microbiota composition), metagenomic sequencing (HUMAnN3 annotation of the KEGG pathway). Metabolomics: LC-MS / MS targeted quantification of serum short-chain fatty acids (SCFAs), secondary bile acids, and tryptophan metabolites. Clinical and dose data: age, BMI, radiotherapy regimen (total dose, fractionation), irradiated intestinal volume (V30, V45), PS score.

[0033] Multi-omics feature screening and fusion modeling: Univariate analysis: Radiomics: t-test / Mann-Whitney U screening for differential features in the RE 2+ group (e.g., increased texture entropy in the high-dose intestinal region); Microbiota / Metabolomics: LEfSe analysis for differential bacterial genera (e.g., reduced Bacteroides) and metabolites (decreased butyrate levels). Multivariate dimensionality reduction: LASSO regression screening for key radiomics features (λ determined by 10-fold cross-validation); Microbiota-metabolism network analysis: SParCC calculation of the correlation between microbiota and metabolites (|r|>0.6, p<0.05); Multimodal feature fusion: stacking traditional radiomics, deep features, microbiota α diversity, butyrate concentration, and clinical variables. Model construction: Algorithm selection: XGBoost, LightGBM, and Random Forest (RF) to construct a multi-omics fusion model, with 5-fold cross-validation to optimize hyperparameters. Dynamic prediction: LSTM processing of time-series metabolite changes during radiotherapy (e.g., weekly butyrate fluctuations), combined with baseline features to predict RE risk. Interpretability: SHAP value quantifies feature contribution (e.g., "RE risk increases 2.1-fold when V45>50%)", Grad-CAM visualizes high-risk intestinal imaging regions. Validation strategy: Retrospective cohort validation of model generalization (AUC, sensitivity, specificity); Delong test compares the performance of traditional imaging models vs. multi-omics fusion models; Decision curve analysis (DCA) assesses net clinical benefit.

[0034] Mechanism analysis and intervention validation of the gut microenvironment: Causal inference: Structural equation modeling (SEM) was used to construct a causal pathway of "radiotherapy dose → dysbiosis → metabolic disorder → Th17 / Treg imbalance → RE occurrence"; MetaCyc pathway enrichment analysis was conducted to examine the interaction between gut microbiota functional genes and host metabolism (e.g., butyrate synthesis gene deletion and inflammation exacerbation). Mechanism validation: Metagenomic source tracing analysis was used to determine the colonization rate of the recipient gut microbiota; metabolomics was used to dynamically monitor the recovery of butyrate and secondary bile acids.

[0035] Clinical translation and system deployment, AI-assisted decision-making system: Develop the Django platform to integrate dose distribution analysis, microbiome / metabolite detection data, and output RE risk grading in real time; Visualization module: Heat map marking of high-risk areas in the gut (based on Grad-CAM) and generation of personalized interventions.

[0036] Statistical analysis, differences between groups: continuous variables: t-test (normal distribution) or Mann-Whitney U test (non-parametric); categorical variables: chi-square test or Fisher's exact test. Model performance: ROC curve (AUC, 95% CI), calibration curve (Brier score), Hosmer-Lemeshow test. Causal validation: SEM goodness of fit (RMSEA < 0.08, CFI > 0.90).

[0037] This application retrospectively collected clinical treatment data from 100 patients with abdominopelvic tumors, detailing the patients' radiotherapy efficacy, the occurrence of radiation enteritis (RE), and other adverse reactions (such as decreased white blood cell and platelet counts). Based on the patients' pre-radiotherapy CT images and dose measurement information in the treatment plan, feature profiles were segmented, and radiomics features, including PTV regions (100%, 105%, 108%) and different dose line regions (20 Gy, 30 Gy, 40 Gy), were extracted, totaling over 3000 radiomics features. After removing features with zero variance using analysis of variance, t-tests were used to screen for features significantly associated with the occurrence of radiation enteritis. Further LASSO regression analysis ultimately selected 32 key features for model construction.

[0038] This application employed several mainstream machine learning algorithms (including LASSO, logistic regression, KNN, SVM, GBDT, XGBoost, and LightGBM) for modeling, and evaluated the models using ROC analysis. The results showed that all models exhibited good predictive performance, with the AUC values ​​of the five mainstream modeling methods exceeding 0.85, indicating that the radiomics-based models have high predictive accuracy for the occurrence of radiation enteritis. This result not only validates the effectiveness of radiomics features in predicting radiation enteritis but also provides important guidance for subsequent prospective studies and clinical interventions. The results are as follows... Figure 2-4 As shown.

[0039] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1.A method for preventing radiation enteritis based on AI model and image fusion, characterized in that, The method comprises the following steps: S1, selecting 200 patients with malignant tumors who have received abdominal and pelvic radiotherapy in the past to construct a retrospective cohort; selecting 100 new patients who are undergoing abdominal and pelvic radiotherapy to construct a prospective cohort, collecting baseline data before radiotherapy, and dynamically tracking the occurrence of radiation enteritis during radiotherapy and 3 months after treatment; S2, collecting clinical data and CT images of the patients in the retrospective cohort, and drawing the intestinal dose-sensitive region ROI; Collecting microbiota-metabolomics data of the patients in the prospective cohort; S3, performing multi-omics feature screening and fusion modeling feature engineering, specifically including: (1) Single factor analysis: analyzing the texture entropy value of the high-dose region of the intestine; (2) Multivariate dimensionality reduction: screening key imageomics features through LASSO regression algorithm; (3) Microbial-metabolic network analysis: calculating the correlation between microorganisms and metabolites through SParCC algorithm; (4) Multi-modal feature fusion: stacking traditional imageomics, deep features, microbial alpha diversity, butyric acid concentration and clinical variables; (5) Model construction: constructing a multi-omics fusion model using XGBoost, LightGBM and random forest algorithm, and optimizing the hyperparameters using five-fold cross-validation; (6) Model verification; S4, intestinal microenvironment mechanism analysis and intervention verification: constructing causal pathways through structural equation modeling, and analyzing the interaction between microbial functional genes and host metabolism through MetaCyc pathway enrichment analysis; S5, statistical analysis: verifying the differences between groups through continuous variables and categorical variables, and verifying the model performance through ROC curve, calibration curve and Hosmer-Lemeshow test. 2.The AI model and image fusion-based radiation enteritis prevention method of claim 1, wherein, The intestinal dose-sensitive region ROI in step S2 includes a dose gradient ROI and a target region adjacent ROI, the dose gradient ROI includes an intestinal region covered by 20Gy, 30Gy and 40Gy dose lines, and the target region adjacent ROI includes an intestinal canal within a range of 1cm expanded outside the PTV. 3.The AI model and image fusion-based radiation enteritis prevention method of claim 1, wherein, The changes of metabolites during radiotherapy are processed by LSTM, and the baseline features are combined to predict the risk of radiation enteritis. 4.The AI model and image fusion-based radiation enteritis prevention method of claim 1, wherein, In step S3, the generalization of the model is verified through the retrospective cohort; the performance of the traditional image model and the multi-omics fusion model is compared through Delong test; and the clinical net benefit is evaluated through decision curve analysis. 5.The AI model and image fusion-based radiation enteritis prevention method of claim 1, wherein, The causal pathways in step S4 are as follows: radiotherapy dose, microbial imbalance, metabolic disorder, Th17 / Treg imbalance, and radiation enteritis.

Images