Risk stratification method and system for radiochemotherapy of cervical cancer based on decoupling of tumor and host

Abstract

The application relates to a tumor-host decoupling-based radiochemotherapy risk stratification method and system for cervical cancer, and belongs to the technical field of medical image analysis, artificial intelligence auxiliary diagnosis and treatment, tumor efficacy prediction and clinical decision support, wherein the method acquires a pelvic multi-modal magnetic resonance image of a patient before treatment, sequentially carries out image preprocessing, tumor region of interest and host skeletal muscle region of interest segmentation, tumor habitat map construction, double-branch feature coding and decoupling representation learning, respectively represents tumor image phenotypes and host reserve phenotypes as potential factors, and generates a treatment failure risk score, low, medium and high risk grades and tumor-host prototypes through a risk output module. The application can effectively assist in pre-radiochemotherapy risk assessment, risk source explanation and multidisciplinary consultation shunting of cervical cancer.

Description

Technical Field

[0001] This invention belongs to the fields of medical image analysis, artificial intelligence-assisted diagnosis and treatment, tumor efficacy prediction and clinical decision support. It relates to a method and system for risk stratification of cervical cancer radiotherapy and chemotherapy based on tumor-host decoupling, and more particularly to a method and system for risk stratification of failure of standard concurrent chemoradiotherapy for locally advanced cervical cancer based on pretreatment multimodal pelvic MRI. Background Technology

[0002] Locally advanced cervical cancer (LACC) typically includes patients with stages IB3 to IVA (specific anatomical stages of cervical cancer) according to the 2018 FIGO staging system. Concurrent chemoradiotherapy (CCRT) is one of the important standard treatments for LACC, but a significant proportion of patients still experience local recurrence, distant metastasis, or disease-related death after standard CCRT; even patients at the same FIGO stage show significant differences in disease-free survival outcomes. Therefore, relying solely on anatomical staging and routine clinicopathological indicators is insufficient to adequately support individualized risk assessment before treatment.

[0003] In recent years, intensified treatment strategies such as combining programmed cell death 1 (PD-1) inhibitors with CCRT and sequential induction chemotherapy with CCRT have shown the potential to improve outcomes in high-risk LAC patients. However, intensified treatment is often accompanied by prolonged treatment cycles, monitoring of immune-related toxicities, increased acute toxicity, increased consumption of medical resources, and challenges in patient compliance. Therefore, identifying patients at high risk of failure after standard CCRT before treatment is an important technical requirement for multidisciplinary teams from oncology, radiation oncology, and radiology to determine whether intensified treatment or supportive treatment optimization is necessary.

[0004] Existing pre-treatment MRI radiomics and deep learning methods have been used to predict the response to or prognosis of cervical cancer radiotherapy and chemotherapy, but they still have the following shortcomings: First, most methods mainly focus on modeling the imaging phenotype of the tumor region, and rarely simultaneously characterize factors related to treatment tolerance, such as host nutrition, muscle reserve, anemia, low albumin, or immune status; Second, even if some methods incorporate host-related clinical indicators, they are mostly added as post-covariates to the prediction model, failing to achieve a structured fusion of tumor imaging phenotype and host reserve phenotype in the deep representation space; Third, most models output a single risk score, making it difficult to distinguish whether the increased risk is mainly due to increased tumor invasiveness or insufficient host reserve; Fourth, some interpretability analyses rely solely on significance heatmaps, lacking robustness checks such as parameter randomization and label randomization; Fifth, some methods are mainly based on single-center training or internal validation, with limited robustness across different scanner vendors, field strengths, geographical populations, and treatment eras; Sixth, some methods require additional examinations such as pre-treatment positron emission tomography (PET) and pathological immunohistochemistry as inputs or deployment conditions, limiting their accessibility in routine MRI clinical pathways.

[0005] Therefore, there is an urgent need for a computer-aided risk stratification technique based on conventional pre-treatment multimodal MRI. This technique should extract tumor region imaging phenotypes and pelvic skeletal muscle-related host reserve phenotypes simultaneously, without requiring fluorodeoxyglucose positron emission tomography (FDG-PET), immunohistochemistry (IHC), or public omics data as necessary inputs. By decoupling representation learning, it can form distinguishable and interpretable intrinsic tumor factors and host reserve factors, thereby outputting failure risk, risk level, and tumor-host prototypes before CCRT, providing auxiliary information for enrichment and triage in multidisciplinary consultations. Summary of the Invention

[0006] The technical problem this invention aims to solve is: how to accurately identify high-risk patients who have failed treatment before standard CCRT using routine multimodal pelvic MRI in patients with locally advanced cervical cancer, and further differentiate whether the increased risk mainly stems from enhanced intrinsic tumor invasiveness or insufficient host reserve. Specifically, this includes:

[0007] 1. How to simultaneously obtain tumor region information and pelvic skeletal muscle host reserve information in the same pre-treatment MRI (magnetic resonance imaging) examination.

[0008] 2. How to encode tumor-related phenotypes and host reserve-related phenotypes into two relatively independent and interpretable latent subspaces.

[0009] 3. How to obtain biologically plausible intrinsic tumor factors and host reserve factors without requiring additional examinations such as FDG-PET and IHC as necessary inputs to the model.

[0010] 4. How to output a failure risk score, low / medium / high risk level, and four tumor-host prototypes that can be used for clinical multidisciplinary consultations, instead of just outputting a single, indivisible black-box risk value.

[0011] 5. How to improve the interpretability and safety of deep learning prediction results in clinical deployment through uncertainty estimation and attribution soundness checks.

[0012] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:

[0013] A risk stratification method for cervical cancer radiotherapy and chemotherapy based on tumor-host decoupling includes:

[0014] Input data acquisition: Collect pelvic multimodal MRI data that meets quality requirements within a specified time period before the patient's treatment;

[0015] Image preprocessing: Offset field correction, histogram matching, intensity normalization, unified resampling, and cross-sequence layer-by-layer registration are performed on the raw MRI data to ensure spatial and intensity consistency of the data.

[0016] Region of Interest (ROI) segmentation: Automatically segment and verify the tumor ROI, and automatically segment the four-compartment skeletal muscle complex of the pelvis to obtain the host skeletal muscle ROI, thus completing the precise localization of two key regions;

[0017] Input volume cropping and standardization: The image volume is cropped centered on the tumor and muscle regions of interest, resampled according to a preset size, and standardized host branch input data is constructed through masking.

[0018] Tumor habitat construction: Cluster analysis based on cross-sequence image intensity is performed within the tumor region to generate a voxel-level tumor habitat map, which serves as a supplementary input feature for subsequent risk stratification processes;

[0019] Dual-branch feature encoding: A deep encoder is used to independently encode tumor multimodal images, habitat information, and host muscle region information to generate corresponding region feature embeddings;

[0020] Decoupled representation learning: After embedding and fusing tumor and host features, it is projected into two sets of independent latent factors. Decoupled representation of intrinsic tumor factors and host reserve factors is achieved through independence regularization and auxiliary supervision.

[0021] Treatment risk stratification output: Based on the decoupled potential features, the risk output module outputs a treatment failure risk score;

[0022] Prototypes and Explainable Outputs: Based on the spatial distribution of potential factors, tumor and host prototypes are generated, and predictive uncertainty and explainable attribution results are output simultaneously to assist in clinical decision-making.

[0023] Furthermore, a risk stratification method for cervical cancer radiotherapy and chemotherapy based on tumor-host decoupling also includes:

[0024] Optional counterfactual analysis: Under conditions that meet data and confounding controls, conduct subgroup-level analysis to explore the heterogeneity of immunotherapy benefits, providing a reference for optimizing individualized treatment.

[0025] Furthermore, the pelvic multimodal MRI data includes fat-suppressed T2-weighted, enhanced T1-weighted, and diffusion-weighted imaging sequences.

[0026] Furthermore, k=3 clustering is performed within the tumor region of interest based on the intensity of the three sequences, and a voxel-level tumor habitat map is generated.

[0027] Furthermore, the two sets of independent potential factors are tumor intrinsic factors and host reserve factors. Based on the two-dimensional spatial distribution of tumor intrinsic factors and host reserve factors, four types of tumors and host prototypes are classified: standard treatable type, high tumor reserve type, frail care type, and dual vulnerable type.

[0028] Furthermore, a treatment failure risk score is output, and the risk is classified into low, medium, and high risk levels according to a threshold.

[0029] A stratification system for risk stratification of cervical cancer radiotherapy and chemotherapy based on tumor-host decoupling includes:

[0030] Data receiving module: Used to receive and verify the patient's pre-treatment pelvic multimodal MRI raw data, ensuring the integrity and validity of the input data.

[0031] Image preprocessing module: performs MRI data bias field correction, intensity normalization, spatial resampling and cross-sequence registration preprocessing operations, and outputs standardized image data.

[0032] Region of Interest (ROI) Segmentation Module: Used to obtain the tumor ROI and the host skeletal muscle ROI, where the tumor ROI is the cervical cancer lesion area and the host skeletal muscle ROI is the pelvic four-compartment skeletal muscle complex area.

[0033] Input volume cropping and standardization module: The image volume is cropped around the tumor and muscle regions of interest respectively, resampled according to the preset size, and standardized host branch input data is constructed through masking to provide a standardized and uniform input volume for subsequent dual-branch encoding.

[0034] Tumor habitat construction module: Based on cross-sequence intensity information of tumor regions, clustering and label alignment are performed to generate voxel-level tumor habitat map features.

[0035] Dual-branch encoding module: The module processes tumor multimodal images and habitat information and host muscle region information separately through independent encoders to generate corresponding region feature embeddings.

[0036] Decoupled representation module: Embedded and fused dual-branch features and decoupled them into independent latent factors. Regularization and auxiliary supervision ensure the independence and interpretability of the factors.

[0037] Risk output module: Outputs a treatment failure risk score based on decoupled potential factors.

[0038] Prototype triage module: Based on the spatial distribution mapping of potential factors, it generates four tumor and host prototypes to achieve patient subtype classification.

[0039] Uncertainty and Explanation Module: Outputs prediction uncertainty indicators and attribution explanation results, and completes prediction reliability verification.

[0040] Furthermore, the risk stratification system also includes an optional counterfactual analysis module, used to explore differences in the benefits of immune enhancement therapy at the subgroup level, providing exploratory evidence for clinical optimization.

[0041] This optional counterfactual analysis module is not a necessary step in the core risk stratification process, nor is it used to generate personalized treatment instructions.

[0042] Furthermore, the two sets of independent potential factors are tumor intrinsic factors and host reserve factors. The decoupling representation module adopts Hilbert-Schmidt independence criterion regularization and is equipped with biological indicators for auxiliary supervision to achieve independent decoupling of tumor intrinsic factors and host reserve factors.

[0043] Furthermore, the risk output module outputs a treatment failure risk score based on the decoupled potential factors, and then classifies it into low, medium, and high risk levels according to a threshold.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] 1. Improve the non-response discrimination capability of the 3-month standard CCRT. In a multicenter validation experiment, the area under the curve (AUC) of the pooled validation population was 0.81, with a 95% confidence interval (CI) of 0.78–0.85. The AUCs for the locked internal test set, the time-independent cohort, and the geographic external cohort were 0.83, 0.82, and 0.81, respectively. The AUC for the 1.5T out-of-distribution probe cohort in center 3 was 0.78.

[0046] 2. Compared with FIGO alone, clinical + laboratory alone, standard radiomics and tumor single-branch models, it has incremental value, with an AUC improvement of +0.0524 compared with tumor single-branch models in pooled validation.

[0047] 3. Improved clinical net benefit. Within the clinically relevant threshold range of 0.20–0.45, the net benefit of TumorHost-DT was consistently higher than that of the comparative model; at the threshold of 0.32, the improvement in net benefit (NB) compared to the tumor single-branch model was +0.038, and the improvement in net benefit compared to FIGO alone was +0.078.

[0048] 4. Risk stratification of progression-free survival (PFS) was achieved. In the primary analysis population of 1,350 patients using the standard CCRT, for every 0.1-point increase in the TumorHost-DT risk score, the adjusted hazard ratio (HR) for PFS was 1.30 (95% CI 1.18–1.43); the 2-year DFS rates for the low, intermediate, and high stratification levels were 81%, 64%, and 41%, respectively.

[0049] 5. Demonstrates calibration and robustness. Geographic external cohort calibration intercept -0.05, calibration slope 0.94, Brier score 0.19, Hosmer-Lemeshow P=0.28; overall performance is stable across manufacturers, field strengths, treatment eras, and subgroup stratifications.

[0050] 6. Provides biological interpretability. z_T is associated with metabolic load in positron emission tomography (PET), and z_H is associated with hemoglobin, albumin, lymphocyte count, and skeletal muscle markers; IHC and public database analyses support the high-risk phenotype corresponding to hypoxia, immune rejection, and high-proliferative characteristics.

[0051] 7. Supports multidisciplinary consultation and triage language. The model can break down a single "high-risk" label into four tumor-host prototypes, making it easier for MDTs to distinguish between tumor-driven risk, host vulnerability risk, and dual risk.

[0052] 8. Reduce reliance on additional inspections and annotations. The host branch uses automatic skeletal muscle segmentation results on CE-T1WI; FDG-PET and IHC are not required inputs to the model.

[0053] 9. Improve interpretability. Validate GradientSHAP attribution plots through parameter randomization and label randomization sanity checks to reduce the risk of unreliable significance plots misleading clinical interpretation. Attached Figure Description

[0054] The invention will now be further described with reference to the accompanying drawings:

[0055] Figure 1 This is a flowchart of a risk stratification method for cervical cancer radiotherapy and chemotherapy based on tumor-host decoupling, as described in this invention. Detailed Implementation

[0056] I. Overall Concept;

[0057] This invention proposes a TumorHost-DT (Tumor-Host Disentangled Transformer) risk stratification process. For example... Figure 1 As shown, this workflow takes pre-treatment pelvic MRI as input, and proceeds through image preprocessing, region of interest (ROI) and host skeletal muscle ROI segmentation, tumor habitat construction, bi-branch encoding, decoupled representation learning, risk stratification output module, and uncertainty and attribution output to obtain a non-response risk score 3 months after standard CCRT. It can further provide low / intermediate / high risk levels, long-term progression-free survival (PFS) risk indicators, and tumor-host prototypes. It should be noted that ROI is an abbreviation for Region of Interest, which can be simply understood as "the area circled on the image that needs to be analyzed in detail later." In this application, two types of areas need to be circled: the first is the region of interest, which is the location of the cervical cancer lesion, used to analyze the imaging characteristics of the tumor itself; the second is the host skeletal muscle region of interest, which is the location of several groups of skeletal muscles in the pelvis, used to analyze the patient's own physical reserves, such as muscle mass, nutritional status, and treatment tolerance. The region of interest in host skeletal muscle does not refer to the presence of tumors within skeletal muscle, nor is it part of a tumor region; rather, it refers to these muscles as separate objects of analysis.

[0058] The core of this invention is not simply to add clinical variables, but to structurally decouple intrinsic tumor factors and host reserve factors within the representation space, so that the model can simultaneously provide explanations for both the level and source of risk.

[0059] II. As shown in Table 1, the steps of the risk stratification method for cervical cancer radiotherapy and chemotherapy based on tumor-host decoupling;

[0060] Table 1

[0061]

[0062]

[0063] III. Decoupling auxiliary monitoring objectives;

[0064] During the training phase, measurable biological readouts can be used as auxiliary supervision targets. z_T can correspond to tumor descriptors such as tumor size, FIGO stage, and histological type; z_H can correspond to host reserve-related readouts such as pre-treatment hemoglobin, albumin, lymphocyte count, and skeletal muscle index.

[0065] FDG-PET, IHC, and public database omics indicators can be used for biological anchoring or supporting analyses, but are not required inputs to the model, auxiliary supervision targets, hyperparameter selection variables, threshold selection variables, or primary validation endpoints.

[0066] IV. As shown in Table 2, four types of tumors and their host prototypes;

[0067] Table 2

[0068]

[0069] The aforementioned prototype is intended to assist multidisciplinary consultations in discussing risk sources and management priorities, and does not replace the comprehensive judgment of doctors.

[0070] V. As shown in Table 3, the module composition of the cervical cancer radiotherapy and chemotherapy risk stratification system based on tumor-host decoupling;

[0071] Table 3

[0072]

[0073] Example 1: Standard CCRT Failure Risk Stratification;

[0074] For LAC patients who meet the inclusion criteria, pretreatment T2FS, CE-T1WI, and DWI images were obtained. After preprocessing, tumor ROI and muscle ROI segmentation, and tumor habitat construction, the tumor multimodal volume and CE-T1WI muscle complex were input into TumorHost-DT. The standard non-response risk score 3 months after CCRT was output and divided into low, medium, and high risk layers according to thresholds of 0.32 and 0.40.

[0075] Example 2: Tumor-host prototype shunt;

[0076] Based on the model outputs z_T and z_H, patients are categorized into standard treatable, high tumor reserve, frail care, or dual vulnerable groups. This output is used in MDT discussions to identify risk sources, such as tumor-driven, host-driven, or dual-driven.

[0077] Example 3: Model Interpretation and Quality Control;

[0078] Inference uncertainty was quantified by deep ensemble of 5 models and 20 Monte Carlo Dropout (MC dropout) runs for each ensemble member; voxel-level GradientSHAP attribution maps were generated for tumor and host branches respectively, and the attribution maps were verified to be dependent on the model’s real learning signal by parameter randomization and label randomization integrity checks.

[0079] Example 4: Heterogeneity analysis of the benefits of optional treatment enhancement;

[0080] With both a PD-1 intensified treatment cohort and a standard CCRT control cohort available, methods such as the Double Robust Learner (DR-learner), the Counterfactual Prediction Network based on Wasserstein distance (TARNet+Wasserstein), the X-learner, Inverse Probability of Treatment Weighting (IPTW), or propensity score matching can be used to estimate the heterogeneity of treatment effects at low, medium, and high risk levels or at the tumor-host prototype subgroup level. This embodiment is an exploratory subordinate module and is not intended as an automated treatment recommendation at the individual level.

[0081] Boundary description:

[0082] 1. This invention is positioned as a tool for enrichment and risk source interpretation in pre-treatment multidisciplinary consultations, and does not replace doctors in making diagnostic or treatment decisions.

[0083] 2. FDG-PET, IHC, and public omics data can be used for biological anchoring, but do not constitute necessary inputs for the model.

[0084] 3. The PD-1 counterfactual treatment effect analysis is an exploratory, subgroup-level analysis and should not be presented as a validated individualized treatment recommendation.

[0085] 4. Risk thresholds and network structures can be replaced or recalibrated based on training data, clinical scenarios, or regulatory requirements, but this does not affect the core concept of “tumor-host dual-branch input, decoupled representation, and prototype output”.

[0086] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be included within the scope of protection of the present invention. Furthermore, all content not described in detail in this specification is prior art known to those skilled in the art.

Claims

1. A risk stratification method for cervical cancer radiotherapy and chemotherapy based on tumor-host decoupling, characterized in that, include: Input data acquisition: Collect pelvic multimodal MRI data of the patient within a specified time before treatment; Image preprocessing: Offset field correction, histogram matching, intensity normalization, uniform resampling, and cross-sequence layer-by-layer registration were performed on the raw MRI data; Region of Interest (ROI) segmentation: Automatically segment and verify the tumor ROI, and automatically segment the four-compartment skeletal muscle complex of the pelvis to obtain the host skeletal muscle ROI; Input volume cropping and standardization: The image volume is cropped centered on the tumor region of interest and the host skeletal muscle region of interest, resampled according to a preset size, and then masked to construct a standardized host branch input, which is used as the input for subsequent host branch feature encoding; Tumor habitat construction: Cluster analysis based on cross-sequence image intensity is performed in the tumor region of interest to generate a voxel-level tumor habitat map, which serves as a supplementary input feature for the subsequent risk stratification process; Dual-branch feature encoding: A deep encoder is used to independently encode the tumor multimodal image, the voxel-level tumor habitat map, and the host branch input information to generate corresponding regional feature embeddings; Decoupled representation learning: After embedding and fusing tumor and host features, it is projected into two sets of independent latent factors. Decoupled representation of intrinsic tumor factors and host reserve factors is achieved through independence regularization and auxiliary supervision. Treatment risk stratification output: Based on the decoupled potential features, output a treatment failure risk score; Prototypes and Explainable Outputs: Based on the spatial distribution of potential factors, tumor and host prototypes are generated, and prediction uncertainty and explainable attribution results are output simultaneously.

2. The method for risk stratification of cervical cancer radiotherapy and chemotherapy based on tumor-host decoupling as described in claim 1, characterized in that, Also includes: Optional counterfactual analysis: Under conditions that meet data and confounding controls, conduct subgroup-level analysis to explore the heterogeneity of immunotherapy benefits, providing a reference for optimizing individualized treatment.

3. The method for risk stratification of cervical cancer radiotherapy and chemotherapy based on tumor-host decoupling as described in claim 1, characterized in that: The pelvic multimodal MRI data includes fat-suppressed T2-weighted, enhanced T1-weighted, and diffusion-weighted imaging sequences.

4. The method for risk stratification of cervical cancer radiotherapy and chemotherapy based on tumor-host decoupling as described in claim 1, characterized in that: Within the tumor region of interest, k=3 clustering is performed based on the intensity of three sequences to form a voxel-level tumor habitat map.

5. The method for risk stratification of cervical cancer radiotherapy and chemotherapy based on tumor-host decoupling as described in claim 1, characterized in that: The two sets of independent potential factors are tumor intrinsic factors and host reserve factors. Based on the two-dimensional spatial distribution of tumor intrinsic factors and host reserve factors, four types of tumors and host prototypes are classified: standard treatable type, high tumor reserve type, frail care type, and dual vulnerable type.

6. The method for risk stratification of cervical cancer radiotherapy and chemotherapy based on tumor-host decoupling as described in claim 1, characterized in that: Output a treatment failure risk score and classify it into low, medium, and high risk levels according to thresholds.

7. The risk stratification system for cervical cancer radiotherapy and chemotherapy risk stratification based on tumor-host decoupling as described in claim 1, characterized in that, include: Data receiving module: Used to receive and verify the patient's pre-treatment pelvic multimodal MRI raw data, ensuring the integrity and validity of the input data; Image preprocessing module: performs MRI data bias field correction, intensity normalization, spatial resampling and cross-sequence registration preprocessing operations, and outputs normalized image data; Region of Interest (ROI) segmentation module: used to obtain the tumor ROI and the host skeletal muscle ROI; Input volume cropping and standardization module: The image volume is cropped around the tumor region of interest and the host skeletal muscle region of interest respectively, resampled according to the preset size, and standardized host branch input data is constructed through masking to provide a standardized and uniform input volume for subsequent host branch feature encoding; Tumor habitat construction module: performs clustering and label alignment based on cross-sequence intensity information of tumor regions of interest to generate voxel-level tumor habitat map features; Dual-branch encoding module: The module processes tumor multimodal images and voxel-level tumor habitat map features and host branch input information separately through independent encoders to generate corresponding region feature embeddings; Decoupled representation module: Embedded and fused dual-branch features and decoupled them into independent latent factors, and ensured factor independence and interpretability through regularization and auxiliary supervision; Risk output module: Outputs a treatment failure risk score based on decoupled potential factors; Prototype triage module: Based on the spatial distribution mapping of potential factors, four tumor and host prototypes are generated to achieve patient subtype classification; Uncertainty and Explanation Module: Outputs prediction uncertainty indicators and attribution explanation results, and completes prediction reliability verification.

8. The risk stratification system according to claim 7, characterized in that, Also includes: Optional counterfactual analysis module: used to explore differences in the benefits of immune enhancement therapy at the subgroup level, providing exploratory evidence for clinical optimization.

9. The risk stratification system according to claim 7, characterized in that: The two sets of independent potential factors are tumor intrinsic factors and host reserve factors. The decoupling representation module adopts Hilbert-Schmidt independence criterion regularization and is equipped with biological indicators for auxiliary supervision to achieve independent decoupling of tumor intrinsic factors and host reserve factors.

10. The risk stratification system according to claim 7, characterized in that: The risk output module outputs a treatment failure risk score based on the decoupled potential factors, and then classifies the risk into low, medium, and high risk levels according to a threshold.

Images