A liver cancer tumor invasiveness prediction method based on a preoperative multi-modal model

Abstract

The application relates to a preoperative multi-modal model-based hepatocellular carcinoma tumor invasiveness prediction method, which comprises the following steps: collecting preoperative clinical data, contrast-enhanced ultrasound (CEUS) image data and magnetic resonance imaging (MRI) image data of a subject, and screening samples meeting preset conditions; pre-processing the CEUS image data and the MRI image data, extracting corresponding image features and screening stable features; the application combines preoperative CEUS, MRI and clinical multi-modal information, combines self-developed image registration and MAM calculation tools, realizes preoperative accurate prediction of tumor invasiveness, breaks through the limitation of the prior art that depends on postoperative information; through three-layer biological verification of pathology, immunohistochemistry and RNA sequencing, the model is endowed with strong interpretability, the association between the prediction result and the microscopic invasion mechanism of the tumor is determined, the stability is good, and the model can cover different genders, ages, liver backgrounds, tumor sizes and positions.

Description

Technical Field

[0001] This invention relates to the field of tumor prediction technology, specifically to a method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model. Background Technology

[0002] Hepatocellular carcinoma is a highly prevalent malignant tumor worldwide. Ablation therapy has become an important local treatment method due to its minimally invasive and highly effective nature. However, local progression of tumor (LTP) after ablation seriously affects the prognosis of patients, and tumor invasiveness is the core factor leading to LTP.

[0003] Existing LTP prediction technologies mostly rely on information such as the minimum ablation margin (MAM) measured by postoperative imaging, which can only provide follow-up suggestions and cannot guide preoperative treatment decisions. The use of single-modality imaging (such as CEUS or MRI alone) is usually difficult to fully capture the biological characteristics of tumors, resulting in limited predictive efficacy, lack of interpretability at the biological mechanism level, and inability to clearly define the correlation between prediction results and tumor micro-invasive characteristics. Fourth, the formulation of ablation strategies does not take into account the individual heterogeneity of tumors, and the use of a uniform MAM standard is not compatible with the biological characteristics of tumors.

[0004] Therefore, developing a preoperative, multimodal, highly accurate, and biointerpretable method for predicting tumor invasiveness is of significant clinical value for reducing the incidence of LTP and optimizing the efficacy of ablation therapy. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model. This method enables accurate preoperative prediction of tumor invasiveness, clarifies the biological mechanisms underlying the prediction results, and guides the development of individualized ablation strategies to reduce the incidence of LTP (Left-to-Tonify) tumors.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model, comprising the following steps: (1) Data collection and screening: collect preoperative clinical data, contrast-enhanced ultrasound (CEUS) imaging data, and magnetic resonance imaging (MRI) data of the subjects, and screen samples that meet the preset conditions; (2) Multimodal image preprocessing and feature extraction: CEUS image data and MRI image data were preprocessed respectively, and corresponding image features were extracted and stable features were selected; (3) Feature screening and multimodal model construction: Key clinical features are screened, single-modal prediction models are trained separately, and multimodal tumor invasiveness prediction models are constructed by fusion and output high / low invasiveness grouping results; (4) Validation of model biological interpretability: The correlation between the predicted results and the tumor microscopic invasion characteristics was verified through multi-layer biological experiments; (5) Individualized ablation strategy formulation: Based on the prediction results and the calculation results of the minimum ablation boundary (MAM), a differentiated ablation strategy is formulated.

[0007] Furthermore, the preset screening criteria in (1) are: the subject is pathologically confirmed to have hepatocellular carcinoma (HCC) with a tumor diameter ≤5cm, receives curative thermal ablation treatment, completes CEUS and / or MRI examination within 2 weeks before the procedure, and has a follow-up period of ≥24 months after the procedure; the exclusion criteria include cases with incomplete clinical data, imaging data quality that does not meet the analysis requirements, cases that receive palliative ablation treatment, and cases with other malignant tumors.

[0008] Furthermore, the clinical data in (1) includes basic information, medical history information, tumor characteristic information and serological indicators; among which, the basic information is gender and age; the medical history information is history of cirrhosis; the tumor characteristic information is tumor size, tumor location and adjacent blood vessels / capsule of the tumor; the serological indicators include platelet count, international normalized ratio, alanine aminotransferase, aspartate aminotransferase, total protein, serum albumin, total bilirubin, direct bilirubin and alpha-fetoprotein.

[0009] Furthermore, the CEUS image preprocessing and feature extraction in (2) specifically includes: using optical flow technology to process CEUS dynamic video, and combining long short-term memory network (LSTM) to extract the perfusion time sequence features of the tumor arterial phase, portal venous phase and delayed phase; and using the CEUS dynamic video automatic annotation tool to complete the accurate annotation of the tumor region, with the annotated region serving as the effective range for feature extraction.

[0010] Furthermore, the MRI image preprocessing and feature extraction in (2) specifically includes: acquiring multi-parameter MRI images using a 3.0T MRI scanner, and having two professional physicians independently annotate the regions of interest (ROI) of the tumor arterial phase and delayed phase, as well as the ROI of the ablation zone (AZ) after surgery using 3D-slicer software; mapping the ROI of the AZ after surgery to the preoperative MRI using a self-developed MRI elastic registration tool before and after ablation, and then calculating the MAM using the morphological outward contact point detection method; expanding the tumor ROI by 1cm to obtain the tumor periphery, extracting radiomics features using the pyradiomics tool, and screening features with an intragroup correlation coefficient (ICC) ≥ 0.8 as stable features.

[0011] Furthermore, the MRI elastic registration tool before and after ablation is developed based on 3D-slicer, adopts the VoxelMorph framework combined with differential homeomorphism constraint, and the registration error is ≤0.5 voxels; the morphological expansion step size of the MAM calculation tool is 0.1 mm, and the calculation accuracy reaches 0.01 mm.

[0012] Furthermore, step (3) specifically includes: using univariate and multivariate logistic regression analysis to screen independent risk features of LTP in clinical data, the screened features being tumor size, adjacent blood vessels, and adjacent capsule; constructing an LSTM single-modal prediction model based on CEUS perfusion time-series features, and constructing a multilayer perceptron (MLP) single-modal prediction model based on MRI stable radiomics features; and using logistic regression algorithm to fuse the prediction probabilities of the two single-modal models and the screened clinical features to construct a multimodal tumor invasiveness prediction model.

[0013] Furthermore, the multi-layered biological experimental verification in (4) includes: pathological verification, comparing the differences in the incidence of microvascular invasion (MVI), the incidence of trabecular hepatocellular carcinoma (MTM), and the degree of tumor differentiation between the high / low invasiveness groups; immunohistochemical verification, comparing the differences in the positive expression rates of cytokeratin-7 (CK-7) and phosphatidylinositol proteoglycan-3 (GPC-3) between the two groups; and RNA sequencing verification, comparing the differences in the gene expression levels of vascular endothelial growth factor A (VEGFA), matrix metalloproteinase-9 (MMP-9), and heat shock protein 70 homolog 1A (HSPA1A) between the two groups.

[0014] Furthermore, the differentiated ablation strategy in (5) is as follows: for tumors predicted as highly invasive by the multimodal model, an ablation strategy with MAM ≥ 8 mm is adopted; for tumors predicted as low invasive, a conventional ablation strategy with MAM = 5 mm is adopted; and after using an 8 mm ablation boundary for highly invasive tumors, there is no significant difference in the levels of alanine aminotransferase, aspartate aminotransferase, total bilirubin, direct bilirubin, and the incidence of complications after surgery compared with the conventional strategy group.

[0015] Furthermore, the multimodal tumor invasiveness prediction model showed an area under the curve (AUC) ≥ 0.809 on both the internal and external test sets, and an AUC ≥ 0.774 in subgroups of different genders, ages, liver backgrounds, tumor sizes, and tumor locations; the model sensitivity was ≥ 0.789, specificity was ≥ 0.765, and negative predictive value was ≥ 0.926, demonstrating good stability and generalization ability.

[0016] Compared with the prior art, the technical solution of this application has the following beneficial effects: 1. This invention integrates preoperative CEUS, MRI, and clinical multimodal information, combined with self-developed image registration and MAM calculation tools, to achieve accurate preoperative prediction of tumor invasiveness, breaking through the limitations of existing technologies that rely on postoperative information; through three layers of biological validation of pathology, immunohistochemistry, and RNA sequencing, it endows the model with strong interpretability, clearly defines the correlation between prediction results and tumor micro-invasive mechanisms, has good stability, and can cover subgroups of different genders, ages, liver backgrounds, tumor sizes, and locations.

[0017] 2. The 8mm MAM individualized ablation strategy developed based on the prediction results of this invention can reduce the incidence of LTP in highly invasive tumors without significantly increasing postoperative complications, and has direct clinical application value. It proposes a feature extraction scheme that extends the MRI tumor ROI by 1cm to solve the problem that the features of a single tumor region cannot reflect vascular invasion and MVI. The multimodal fusion algorithm takes into account the blood perfusion advantages of CEUS and the spatial texture advantages of MRI, and improves the prediction accuracy. Attached Figure Description

[0018] Figure 1 This is a flowchart of the multimodal image preprocessing and feature extraction process of the present invention; Figure 2 This is a flowchart of the feature selection and multimodal model construction process of the present invention; Figure 3 This is a flowchart illustrating the bio-interpretability verification process of the model of this invention. Detailed Implementation

[0019] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1-3 This embodiment presents a method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model, comprising the following steps: 1. Data collection and filtering Preoperative clinical data, CEUS imaging data, and MRI imaging data of subjects were collected, and subjects met the following criteria: pathologically confirmed HCC ≤ 5cm and received curative thermal ablation treatment, CEUS and / or MRI examinations were completed within 2 weeks before the procedure, and follow-up was conducted for ≥ 24 months after the procedure. Cases with incomplete clinical data, poor image quality, and palliative ablation were excluded.

[0021] Clinical data include gender, age, history of cirrhosis, tumor size, tumor location (adjacent blood vessels / capsule), and serological markers (platelets, international normalized ratio, alanine aminotransferase, aspartate aminotransferase, total protein, serum albumin, total bilirubin, direct bilirubin, and alpha-fetoprotein).

[0022] 2. Multimodal image preprocessing and feature extraction CEUS Image Processing: Optical flow technology and Long Short-Term Memory (LSTM) network are used to process CEUS dynamic videos, extract tumor perfusion time-series features, and complete accurate annotation of tumor regions through CEUS dynamic video automatic annotation tool.

[0023] MRI Image Processing: Multiparameter MRI images were acquired using a 3.0T MRI scanner. Two physicians independently annotated the preoperative arterial and delayed phase ROIs of the tumor using 3D-slicer software, and simultaneously annotated the postoperative ablation zone (AZ) ROI. Using a self-developed MRI elastic registration tool before and after ablation, based on elastic registration algorithm and deformation field optimization technology, the ROI of the postoperative AZ was mapped to the preoperative MRI. Then, the MAM was calculated using an MRI-based automatic MAM calculation tool with morphological outward contact point detection method. The tumor ROI was expanded by 1cm to obtain the surrounding area of ​​the tumor, and the pyradiomics tool was used to extract radiomics features, and stable features with an intragroup correlation coefficient (ICC) ≥0.8 were screened.

[0024] 3. Feature selection and multimodal model construction Clinical feature screening: Tumor size, adjacent blood vessels, and adjacent capsule were screened as independent clinical features of LTP risk through univariate and multivariate logistic regression analysis.

[0025] Single-modal model training: An LSTM prediction model (CEUS model) is constructed based on CEUS temporal features, and a multilayer perceptron (MLP) prediction model (MRI model) is constructed based on MRI stable radiomics features.

[0026] Multimodal model fusion: Using logistic regression algorithm, the predicted probabilities of CEUS model, MRI model and screened clinical features are fused to construct a multimodal tumor invasiveness prediction model and output the results of high / low tumor invasiveness grouping.

[0027] 4. Validation of interpretability of the model organism Subject samples with pathological specimens were collected, and pathological analysis was used to verify the differences in microvascular invasion (MVI), trabecular hepatocellular carcinoma (MTM), and tumor differentiation between the high / low invasiveness groups.

[0028] Immunohistochemical analysis was used to verify the differences in the positive expression rates of cytokeratin-7 (CK-7) and phosphatidylinositol proteoglycan-3 (GPC-3) between the two groups.

[0029] RNA sequencing analysis was used to verify the differences in expression levels of invasion-related genes such as vascular endothelial growth factor A (VEGFA), matrix metalloproteinase-9 (MMP-9), and heat shock protein 70 homolog 1A (HSPA1A) between the two groups.

[0030] 5. Development of individualized ablation strategies For tumors predicted by the model to be highly invasive, an ablation strategy was formulated based on the MAM calculation results, using a minimum ablation boundary of 8 mm (MAM≥8 mm); for low-invasive tumors, a conventional 5 mm MAM ablation strategy was adopted.

[0031] Example (a) Data Sources Training set: 502 HCC samples from 3 hospitals from January 2015 to December 2022, of which 225 cases contained only CEUS data (training set 1) and 277 cases contained only MRI data (training set 2).

[0032] Validation set: 180 HCC samples from the aforementioned hospitals during the same period, all including CEUS, MRI, and complete clinical data.

[0033] Internal test set: 250 HCC samples from the aforementioned hospitals during the same period, including CEUS, MRI, clinical data, and postoperative MRI data.

[0034] External test set: 276 HCC samples from another center from January 2017 to August 2023, to verify the model's generalization ability.

[0035] Biovalidation set: 223 HCC samples with pathological results, 139 samples with MVI data, 171 samples with MTM data, 166 samples with CK-7 data, 178 samples with GPC-3 data, and 63 samples with RNA sequencing data.

[0036] (II) Specific Implementation Steps 1. Image processing parameter settings CEUS processing: Optical flow technology was used to extract perfusion features in the arterial, portal, and delayed phases. The LSTM network was configured with 3 hidden layers and the activation function was tanh.

[0037] MRI processing: 3.0T MRI scanner, gadotate dimeglumine injection dose 0.2mmol / kg, ROI labeling was performed using 3D-slicer software, after expanding outward by 1cm, 1323 radiomics features were extracted, and features with ICC≥0.8 were selected for model training.

[0038] The flexible registration tool is developed based on 3D-slicer and adopts the VoxelMorph framework combined with differential homeomorphism constraints, with a registration error of ≤0.5 voxels. The MAM calculation tool adopts a morphological expansion step of 0.1mm and a calculation accuracy of 0.01mm.

[0039] 2. Model Training and Validation CEUS model: trained on training set 1, AUC=0.831 (95%CI 0.757-0.893), accuracy=0.818.

[0040] MRI model: The model was trained on training set 2. The feature of 1cm outward expansion in the arterial phase was selected to construct the model. The AUC was 0.852 (95% CI 0.799-0.896) and the accuracy was 0.816.

[0041] Multimodal model: In the validation set, the CEUS model, MRI model and clinical features (tumor size, adjacent blood vessels, adjacent capsule) were fused. The AUC was 0.823 (95% CI 0.753-0.881) and the accuracy was 0.783.

[0042] 3. Biovalidation Results Pathological verification: The proportions of MVI, MTM, and poor differentiation in the high-invasive group were 40.9%, 39.7%, and 35.9%, respectively, all of which were significantly higher than those in the low-invasive group (p<0.05).

[0043] Immunohistochemical verification: The positive rates of CK-7 and GPC-3 in the high-invasive group were 25.5% and 69.2%, respectively, which were significantly higher than those in the low-invasive group (p<0.05).

[0044] RNA sequencing verification showed that the expression levels of VEGFA, MMP-9, and HSPA1A were significantly increased in the highly invasive group (p < 0.01).

[0045] 4. Implementation and Effects of Ablation Strategies For highly invasive tumors using the MAM≥8mm strategy, the incidence of LTP in the external test group was 10.5%, which was not significantly different from the low-invasive group (6.1%) (p=0.613). Furthermore, the postoperative ALT, AST, TB, DB levels and complication rates were not significantly different from the conventional strategy group (p>0.05).

[0046] (III) Model Performance Evaluation Internal test set: Multimodal model AUC=0.809 (95% CI 0.738-0.872), sensitivity=0.789, specificity=0.777, negative predictive value=0.926.

[0047] External test set: Multimodal model AUC=0.811 (95% CI 0.771-0.868), sensitivity=0.825, specificity=0.765, negative predictive value=0.937.

[0048] Subgroup analysis: In subgroups of different genders, ages, liver backgrounds, tumor sizes, and locations, the model AUC was ≥0.774, indicating good stability. It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0049] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model, characterized in that, Includes the following steps: (1) Data collection and screening: collect preoperative clinical data, contrast-enhanced ultrasound (CEUS) imaging data, and magnetic resonance imaging (MRI) data of the subjects, and screen samples that meet the preset conditions; (2) Multimodal image preprocessing and feature extraction: CEUS image data and MRI image data were preprocessed respectively, and corresponding image features were extracted and stable features were selected; (3) Feature screening and multimodal model construction: Key clinical features are screened, single-modal prediction models are trained separately, and multimodal tumor invasiveness prediction models are constructed by fusion and output high / low invasiveness grouping results; (4) Validation of model biological interpretability: The correlation between the predicted results and the tumor microscopic invasion characteristics was verified through multi-layer biological experiments; (5) Individualized ablation strategy formulation: Based on the prediction results and the calculation results of the minimum ablation boundary (MAM), a differentiated ablation strategy is formulated.

2. The method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model according to claim 1, characterized in that, The preset filtering conditions in (1) are: Subjects were pathologically confirmed to have hepatocellular carcinoma (HCC) with a tumor diameter ≤5cm, and underwent curative thermal ablation treatment. CEUS and / or MRI examinations were completed within 2 weeks before the procedure, and the follow-up period after the procedure was ≥24 months. Exclusion criteria included cases with incomplete clinical data, imaging data that did not meet the analysis requirements, cases that received palliative ablation therapy, and cases with other malignant tumors.

3. The method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model according to claim 1, characterized in that, The clinical data in (1) include basic information, medical history information, tumor characteristic information and serological indicators; The basic information includes gender and age; The patient's medical history includes cirrhosis. Tumor characteristics include tumor size, tumor location, and the presence of adjacent blood vessels / capsule. Serological markers include platelet count, international normalized ratio, alanine aminotransferase, aspartate aminotransferase, total protein, serum albumin, total bilirubin, direct bilirubin, and alpha-fetoprotein.

4. The method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model according to claim 1, characterized in that, The CEUS image preprocessing and feature extraction in (2) specifically include: Optical flow technology was used to process CEUS dynamic video, and Long Short-Term Memory Network (LSTM) was used to extract perfusion time sequence features of tumor arterial phase, portal venous phase and delayed phase; The CEUS dynamic video automatic annotation tool is used to accurately annotate the tumor region, and the annotated region serves as the effective range for feature extraction.

5. The method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model according to claim 1, characterized in that, The MRI image preprocessing and feature extraction in (2) specifically include: Multiparameter MRI images were acquired using a 3.0T MRI scanner. Two professional physicians independently annotated the regions of interest (ROI) in the arterial and delayed phases of the tumor and the ROI in the ablation zone (AZ) after surgery using 3D-slicer software. The ROI of the postoperative ablation zone (AZ) was mapped to the preoperative MRI using a self-developed MRI elastic registration tool before and after ablation, and the MAM was calculated using the morphological outward contact point detection method. The tumor ROI was expanded by 1 cm to obtain the surrounding area of ​​the tumor. The pyradiomics tool was used to extract radiomics features, and features with an intragroup correlation coefficient (ICC) ≥ 0.8 were selected as stable features.

6. The method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model according to claim 5, characterized in that, The MRI elastic registration tool before and after ablation is developed based on 3D-slicer and adopts the VoxelMorph framework combined with differential homeomorphism constraint, with a registration error of ≤0.5 voxels; the morphological expansion step size of the MAM calculation tool is 0.1 mm, and the calculation accuracy reaches 0.01 mm.

7. The method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model according to claim 1, characterized in that, Specifically, (3) includes: Univariate and multivariate logistic regression analyses were used to screen independent risk features of LTP in clinical data. The screened features were tumor size, adjacent blood vessels, and adjacent capsule. An LSTM single-modal prediction model was constructed based on CEUS perfusion time-series features, and a multilayer perceptron (MLP) single-modal prediction model was constructed based on MRI stable image omics features. A multimodal tumor invasiveness prediction model was constructed by fusing the predicted probabilities of two unimodal models and the selected clinical features using a logistic regression algorithm.

8. The method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model according to claim 1, characterized in that, The multi-layer biological experimental verification in (4) includes: Pathological verification was performed to compare the differences in the incidence of microvascular invasion (MVI), trabecular hepatocellular carcinoma (MTM), and tumor differentiation between the high and low invasiveness groups. Immunohistochemical verification was performed to compare the differences in the positive expression rates of cytokeratin-7 (CK-7) and phosphatidylinositol proteoglycan-3 (GPC-3) between the two groups. RNA sequencing was used to verify and compare the differences in gene expression levels of vascular endothelial growth factor A (VEGFA), matrix metalloproteinase-9 (MMP-9), and heat shock protein 70 homolog 1A (HSPA1A) between the two groups.

9. The method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model according to claim 1, characterized in that, The differentiated ablation strategy in (5) is as follows: For tumors predicted as highly invasive by the multimodal model, an ablation strategy with a MAM ≥ 8mm was adopted. For tumors predicted to be low-invasive, a conventional ablation strategy with a MAM of 5mm was adopted. After using an 8mm ablation border for highly invasive tumors, there were no significant differences in postoperative alanine aminotransferase (ALT), aspartate aminotransferase (AST), total bilirubin, direct bilirubin levels, and complication rates compared to the conventional strategy group.

10. The method for predicting the invasiveness of hepatocellular carcinoma based on a preoperative multimodal model according to claim 1, characterized in that, The multimodal tumor invasiveness prediction model showed an area under the curve (AUC) ≥ 0.809 on both the internal and external test sets, and an AUC ≥ 0.774 in subgroups of different genders, ages, liver backgrounds, tumor sizes, and tumor locations. The model also demonstrated a sensitivity ≥ 0.789, specificity ≥ 0.765, and negative predictive value ≥ 0.926, exhibiting good stability and generalization ability.

Images