Prediction of RAS gene status in CRLM patients based on radiomics and semantic features

Abstract

The present invention belongs to the field of intelligent medical technology, and specifically relates to a model for predicting the RAS gene status of patients with advanced colorectal cancer liver metastasis based on imaging genomics features and its application. The present invention successfully constructed a model for predicting the RAS gene status of patients with advanced colorectal cancer liver metastasis based on imaging genomics features: the enhanced CT images of the patients before initial treatment were collected, imaging genomics features were extracted, 12 imaging genomics features were obtained using 1000 Lasso‑Logistic analyses, and an imaging genomics prediction model was constructed using a multivariate logistic regression method. The present invention utilizes enhanced CT images to achieve economical, non-invasive and rapid prediction of the RAS gene status of patients with colorectal cancer liver metastasis, assisting in genotyping, which can help clinicians make accurate individualized treatment decisions and has great clinical significance.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent medical technology, and specifically relates to a model for predicting the RAS gene status of patients with advanced colorectal cancer liver metastasis based on imaging genomics features and its application. Background Art

[0002] Colorectal cancer (CRC) is the third most common cause of cancer-related death worldwide. Previous studies have shown that approximately 25-30% of patients diagnosed with CRC develop liver metastases during their disease course. RAS (KRAS and NRAS) mutation status is increasingly being used to guide clinical decision-making in patients with CRC. A safe surgical margin of at least 10 mm is recommended for surgical treatment of patients with colorectal liver metastasis (CRLM). However, the optimal resection margin for patients with CRLM harboring RAS gene mutations remains unclear. Therefore, the selection of surgical margins for liver metastases in patients with CRLM harboring RAS gene mutations should differ. Furthermore, RAS gene mutations are associated with shorter disease-free survival (DFS) and overall survival (OS). RAS gene mutation status crucially influences treatment options for patients with advanced CRC.

[0003] For patients with advanced colorectal cancer, clarifying the RAS gene mutation status is essential for subsequent treatment. Currently, testing for RAS gene mutation status typically requires analysis of tissue specimens obtained through surgery or biopsy, an invasive and costly method. For patients whose tissue specimens are older and biopsy is inaccessible, non-invasive prediction of RAS gene status through radiomics is particularly important. The fundamental principle of radiomics is to analyze high-throughput feature information from medical images to reflect characteristics at the microscopic level (e.g., proteins, genes, and signaling pathways). The use of radiomics features in CT images to construct a RAS gene status prediction model has many unique advantages: (1) non-invasive: traditional RAS gene status detection methods rely on the analysis of tissue specimens, while this model uses CT image data and does not require invasive procedures on patients; (2) wide data sources: whole abdominal enhanced CT is one of the most common auxiliary examinations for patients with advanced colorectal cancer; (3) low cost: this model is a "secondary use" of CT images, and does not require additional treatment costs for patients; (4) short waiting period: enhanced CT images can be obtained instantly after the patient completes the CT scan, while traditional detection methods generally require 5-7 working days to determine the RAS gene mutation status. This non-invasive, widely sourced, low-cost, and short waiting period prediction method can further assist clinicians in making treatment decisions for CRLM patients. Summary of the Invention

[0004] To address the shortcomings of the prior art, the present invention provides a model for predicting the RAS gene status of CRLM patients based on radiomics features and its application. This model can be used to predict the RAS gene status of CRLM patients, assisting physicians in making more accurate disease diagnosis and treatment decisions.

[0005] In order to achieve the above objectives, the present invention provides the following technical solutions.

[0006] The present invention provides a method for constructing a model for predicting the RAS gene status of CRLM patients based on radiomics features, characterized in that the method is established based on radiomics features and the specific steps are as follows:

[0007] S1. Collect enhanced CT imaging data of patients with advanced colorectal cancer liver metastasis at initial diagnosis;

[0008] S2. Analyze the enhanced CT imaging data collected in S1, segment the region of interest, and extract image features;

[0009] S3. Using 1000 Lasso-Logistic analyses, we found that combinations of 12 imaging features and parameters were consistently repeated more than 900 times.

[0010] S4. Use multivariate logistic regression method to construct a corresponding RAS gene status prediction model based on the 12 imaging genomics features obtained in S3.

[0011] Furthermore, the segmentation of the region of interest in S2 adopts a three-dimensional semi-automatic segmentation method to segment the region of interest in the portal venous phase CT image.

[0012] Preferably, the specific steps of the three-dimensional semi-automatic segmentation method are as follows:

[0013] (1) The DICOM format PVP images were imported into 3D-Slicer software, and then the selected ROIs were outlined using a soft tissue window (window width: 350HU, window level: 40HU). The FAST-MARCHING semi-automatic fast segmentation algorithm was used to segment all PVP images of each patient layer by layer to analyze the ROIs as a whole.

[0014] (2) Manually adjust and refine the outlined area layer by layer again, outline along the visible clear boundary of the lesion, and erase adjacent normal tissue structures such as bile ducts and blood vessels;

[0015] (3) The segmentation results of the delineated ROIs were reviewed and further refined by two experienced senior radiologists. Finally, the ROIs were exported to NRRD and MRML formats for storage and further analysis.

[0016] Furthermore, the extraction of image features in S2 includes:

[0017] (1) The first-order features describe the distribution of voxel intensities in ROIs;

[0018] (2) Shape-based features describe the three-dimensional intuitive features of ROIs, including size and shape at the two-dimensional and three-dimensional levels;

[0019] (3) Texture features extracted based on five texture matrices: a) grayscale co-occurrence matrix, b) grayscale region size matrix, c) grayscale run length matrix, d) neighborhood grayscale difference matrix, and e) grayscale correlation matrix;

[0020] (4) Wavelet features: Wavelet filtering is added to the original image to reduce noise while extracting detailed high-dimensional image group features.

[0021] Furthermore, the 12 image group features obtained by screening in S3 include:

[0022] wavelet.HHH_glszm_SizeZoneNonUniformityNormalized, wavelet.HLL_glszm_LowGrayLevelZoneEmphasis", wavelet.HLL_glszm_SmallAreaLowGrayLevelEmphasis", wavelet.LHL_firstorder_Skewness, wavelet.LHL_glcm_ClusterShade, wavelet.LHL_ glcm_Correlation, wavelet.LHL_glcm_MCC, wavelet.LLH_firstorder_Median, wavelet.LLH_glcm_Idn, wavelet.LLH_glszm_GrayLevelNonUniformityNormalized, wavelet.LLH_glszm_SmallAreaLowGrayLevelEmphasis, wavelet.LLH_glszm_ZoneEntropy.

[0023] Furthermore, the prediction model described in S4 is a radiomics prediction model. The radiomics score calculation formula is:

[0024] 3.5512-6.1777*wavelet.HHH_glszm_SizeZoneNonUniformityNormalized-1.8147*wavelet.HLL_glszm_LowGrayLevelZoneEmphasis+1.3087*wavelet.HL L_glszm_SmallAreaLowGrayLevelEmphasis-4.0329*wavelet.LHL_firstorder_Skewness-2.1221*wavelet.LHL_glcm_ClusterShade+1.0033*wavelet.LH L_glcm_Correlation-3.0967*wavelet.LHL_glcm_MCC+0.4027*wavelet.LLH_firstorder_Median+0.5078*wavelet.LLH_glcm_Idn-0.2159*wavelet.LLH_ glszm_GrayLevelNonUniformityNormalized+6.8791*wavelet.LLH_glszm_SmallAreaLowGrayLevelEmphasis+4.6685*wavelet.LLH_glszm_ZoneEntropy.

[0025] The cutoff value is set to 50%, that is, if the score calculated according to the above formula is greater than 50%, it is predicted that the patient has a RAS gene mutation type, otherwise it is predicted that the patient has a RAS gene wild type.

[0026] The present invention also provides a storage medium on which multiple instructions are stored, characterized in that the instructions are suitable for being loaded and executed by a processor to implement the steps of the above-mentioned method for predicting the RAS gene status of colorectal cancer liver metastasis based on the imaging genomics feature model.

[0027] The present invention also provides a terminal, characterized in that it includes: a processor, a storage medium communicatively connected to the processor, and the storage medium is suitable for storing multiple instructions; the processor is suitable for calling the instructions in the storage medium to execute the steps of the above-mentioned method for predicting the RAS gene status of colorectal cancer liver metastasis based on the imaging genomics feature model.

[0028] The present invention also provides a prediction model for the RAS gene status of patients with advanced colorectal cancer liver metastasis, characterized in that it is constructed through the steps of the above-mentioned method for predicting the RAS gene status of patients with advanced colorectal cancer liver metastasis based on the imaging genomics feature model.

[0029] The present invention also provides a device for predicting the RAS gene status of patients with colorectal cancer liver metastasis, characterized in that it includes the above-mentioned RAS gene status prediction model for patients with advanced colorectal cancer liver metastasis.

[0030] Compared with the prior art, the present invention has the following beneficial effects.

[0031] As described above, based on the imaging genomics features before treatment, the present invention constructs a prediction model for the RAS gene status of patients with colorectal cancer liver metastasis, which has the following beneficial effects.

[0032] First, through the method of the present invention, the imaging genomics features of enhanced CT images can be used to more accurately predict the RAS gene status, that is, whether the patient carries a mutated RAS gene, for patients with advanced colorectal cancer liver metastasis before they receive systemic anti-tumor treatment, thereby assisting doctors in providing more accurate judgments and treatment decisions for subsequent anti-tumor treatment.

[0033] Secondly, detecting the mutation status of the RAS gene typically requires analysis of tissue specimens obtained through surgery or biopsy, a method that is not only invasive but also expensive and requires a long wait for results. In this invention, we utilize enhanced CT images, which is one of the most common examinations for cancer patients. Furthermore, the enhanced CT examination method itself is noninvasive and can generate images immediately, providing immediate results. The method of the present invention can achieve economical, noninvasive, and rapid prediction of RAS gene status.

[0034] The method of the present invention, as an auxiliary pathological diagnosis method, can help clinicians make accurate individualized treatment decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1. Using 1000 Lasso-Logistic analyses, 12 image group features were screened. Using 1000 Lasso-Logistic analyses, 12 image features and parameter combinations appeared more than 900 times, namely (1) parameters, (2) wavelet.HHH_glszm_SizeZoneNonUniformityNormalized, (3) wavelet.HLL_glszm_LowGrayLevelZoneEmphasis, (4) wavelet.HLL_glszm_SmallAreaLowGrayLevelEmphasis, (5) wavelet.LHL_firstorder_Skewness, (6) wavelet.LHL_glcm_ ClusterShade, (7) wavelet.LHL_glcm_Correlation, (8) wavelet.LHL_glcm_MCC, (9) wavelet.LLH_firstorder_Median, (10) wavelet.LLH_glcm_Idn, (11) wav elet.LLH_glszm_GrayLevelNonUniformityNormalized, (12) wavelet.LLH_glszm_SmallAreaLowGrayLevelEmphasis, (13) wavelet.LLH_glszm_ZoneEntropy.

[0036] Figure 2 . ROC curves of the radiomics prediction model in the training set (solid line) and validation set (dashed line). DETAILED DESCRIPTION

[0037] The present invention will be described in detail below in conjunction with specific examples. The following examples will contribute to understanding of the present invention, but these examples are only for illustration of the present invention and the present invention is not limited to these contents. The operating methods in the examples are all conventional operating methods in the art. Example

[0038] 1. Research methods and results

[0039] 1.1 Case collection.

[0040] According to the international ISBER criteria, patients with colorectal cancer liver metastasis at the time of initial diagnosis from January 2014 to October 2019 were retrospectively collected.

[0041] Inclusion criteria: 1) patients with pathologically confirmed advanced colorectal cancer liver metastasis; 2) age 18-80 years; 3) Eastern Cooperative Oncology Group performance status 0-2; 4) completed RAS gene mutation status testing in surgical or biopsy specimens; 5) availability of lung and abdominal enhanced CT images before treatment, with CT image slice thickness ≤2 mm; 6) the interval between lung and abdominal enhanced CT examination and surgical or biopsy specimen acquisition was no more than 30 days (range, 4-30 days).

[0042] Exclusion criteria: 1) patients with other tumors; 2) patients with CT images with artifacts due to metal implants or movement; 3) the boundaries of liver metastases in CT images were too blurred to accurately delineate the edges.

[0043] Data Collection: Patients' names, gender, age, and RAS mutation status were collected from hospital medical records. All included patients provided written informed consent. The collection, testing, and follow-up of tumor specimens from this study were approved by the ethics committee of the hospital where the patients were treated.

[0044] According to the inclusion and exclusion criteria, a total of 158 patients (88 mutant patients and 70 wild-type patients) were finally included, and all patients were randomly divided into validation set and training set in a ratio of 7:3.

[0045] 1.2 CT image collection

[0046] CT image collection is the first step in radiomics. First, a large number of imaging images (DICOM format) are obtained from CT images. Then these images are preprocessed, including image reconstruction, noise reduction, grayscale standardization, etc., to ensure the standardization and consistency of feature data acquisition and reconstruction parameters, including radiation dose, scanning plan, reconstruction algorithm and scanning layer thickness.

[0047] Prior to initial systemic treatment, patients underwent enhanced CT of the lungs and abdomen. CT acquisition parameters and conditions were as follows: Standard operating procedures were used using a 64-slice spiral CT machine from various manufacturers, including GE, Phillips, Siemens, and Toshiba. The tube voltage was 120 kVp (range, 100-140 kVp), the tube current was 333 mA (range, 100-752 mA), the CT slice thickness was 2 mm, and standard reconstruction techniques were used. The contrast agent iohexol was administered at a dose of 1.2–1.5 mL / kg body weight, intravenously at a rate of 2.5 mL / second, followed by a 20–30 mL infusion of normal saline. All patients underwent CT scanning in the supine position with inspiratory breath-hold. Portal venous phase (PVP) images were acquired at approximately 60–70 seconds. During image screening, image quality was assessed visually by visual inspection of each slice. Finally, all CT images were desensitized using DICOM cleaner software, removing private information such as patient name, gender, age, examination date, hospital registration number, CT examination number, and hospital name. Different patients were labeled with random numbers. All CT images were stored in DICOM format.

[0048] 1.3 CT image segmentation

[0049] CT image segmentation primarily utilizes 3D-Slicer software (www.slicer.org), employing a three-dimensional semi-automated segmentation method to segment regions of interest (ROIs) in portal venous phase (PVP) CT images. ROIs are generally defined during image processing as regions of interest (ROIs) delineated from the original CT image using boxes, circles, ellipses, or irregular polygons for subsequent analysis. In this study, ROIs were selected as non-confluent liver metastatic lesions with the largest cross-sectional area and clear boundaries.

[0050] First, DICOM-formatted PVP images were imported into 3D-Slicer software. Selected ROIs (ROIs) were then outlined using a soft tissue window (window width: 350 HU, window level: 40 HU). All PVP images for each patient were segmented slice by slice using the FAST-MARCHING semiautomatic segmentation algorithm for comprehensive ROI analysis. The outlined regions were then manually adjusted and refined slice by slice, following the visible boundaries of the lesion and erasing adjacent normal tissue structures such as bile ducts and blood vessels. The final ROI segmentation results were reviewed and further refined by two experienced senior radiologists. Finally, the ROIs were exported to NRRD and MRML formats for storage and further analysis.

[0051] 1.4 Image group feature extraction and screening of CT images.

[0052] Radiomic features of each patient's ROI were extracted using the Pyradiomics method. Radiomic features extracted from the delineated ROIs can quantitatively assess tumor characteristics such as intensity, shape, and texture. These radiomic features can be categorized into three types: 1) first-order features, shape-based features, and texture features. First-order features describe the distribution of voxel intensity within the ROI; 2) shape-based features describe the three-dimensional, intuitive characteristics of the ROI, including size and shape in both two and three dimensions. These features are different from the grayscale intensity distribution in the ROIs; 3) Texture features extracted based on five texture matrices: (1) Gray Level Co-occurrence Matrix (GLCM), (2) Gray Level Size Zone Matrix (GLSZM), (3) Gray Level Run Length Matrix (GLRLM), (4) Neighbouring Gray Tone Difference Matrix (NGTDM), and (5) Gray Level Dependence Matrix (GLDM). In addition, we added wavelet filtering to the original image to reduce noise while extracting detailed high-dimensional image group features.

[0053] After extracting imaging features from each patient's ROI, we screened and reduced the dimensionality of all imaging features. First, we removed imaging features with identical values ​​across all patients and lacking discriminatory power. Second, to validate the robustness and reproducibility of the imaging features, we randomly selected 30 ROIs and calculated the intraclass correlation coefficient (ICC) and concordance correlation coefficient (CCC) for each imaging feature. To calculate the ICC, two radiologists semi-automatically delineated the ROIs layer by layer on the same patient (using the same method as previously described). These two radiologists were blinded to the patient's RAS mutation status, except for the clinical diagnosis of colorectal liver metastasis. To calculate the CCC, ROIs were delineated and segmented for the 30 randomly selected patients, and then the same delineation and segmentation steps were repeated two weeks later by the same radiologist. Finally, imaging features with an ICC or CCC value below 0.75 were excluded from subsequent analyses.

[0054] 1.5 RAS gene mutation detection

[0055] DNA was obtained from formalin-infiltrated, paraffin-embedded tissue sections using a DNA sample preparation kit. Real-time polymerase chain reaction (RT-PCR) was used to determine KRAS exons 2, 3, and 4 and NRAS exons 2, 3, and 4 mutations.

[0056] 1.6 Build a prediction model based on image group features.

[0057] In this study, we constructed a RAS gene status prediction model based on radiomics features. All patients were randomly divided into a training set and a validation set in a ratio of 7:3. Using 1000 Lasso-Logistics analysis, 12 radiomics features and parameter combinations appeared consistently more than 900 times. These 12 features are: Figure 1 ).

[0058] The calculation formula of the constructed radiomics prediction model is:

[0059] 3.5512-6.1777*wavelet.HHH_glszm_SizeZoneNonUniformityNormalized-1.8147*wavelet.HLL_glszm_LowGrayLevelZoneEmphasis+1.3087*wavelet.HL L_glszm_SmallAreaLowGrayLevelEmphasis-4.0329*wavelet.LHL_firstorder_Skewness-2.1221*wavelet.LHL_glcm_ClusterShade+1.0033*wavelet.LH L_glcm_Correlation-3.0967*wavelet.LHL_glcm_MCC+0.4027*wavelet.LLH_firstorder_Median+0.5078*wavelet.LLH_glcm_Idn-0.2159*wavelet.LLH_ glszm_GrayLevelNonUniformityNormalized+6.8791*wavelet.LLH_glszm_SmallAreaLowGrayLevelEmphasis+4.6685*wavelet.LLH_glszm_ZoneEntropy.

[0060] The cutoff value is set to 50%, that is, if the score calculated according to the above formula is greater than 50%, it is predicted that the patient has a RAS gene mutation type, otherwise it is predicted that the patient has a RAS gene wild type.

[0061] 1.7 Comparison of the predictive performance of radiomics models.

[0062] The area under the receiver operating characteristic (ROC) curve (AUC) was used to evaluate the ability of the model constructed based on radiomics features to predict the RAS gene status in patients with advanced colorectal cancer liver metastasis. A radiomics RAS gene status prediction model was constructed based on these 12 radiomics features. The AUC of this radiomics model was 0.84 in the training set and 0.79 in the validation set ( Figure 2 ).

Claims

1. A method for constructing a model for predicting the RAS gene status of CRLM patients based on radiomics features, characterized in that: The method is based on radiomics features and the specific steps are as follows: S1. Collect enhanced CT imaging data of patients with advanced colorectal cancer liver metastasis at initial diagnosis; S2. Analyze the enhanced CT imaging data collected in S1, segment the region of interest, and extract image features; S3. Using 1000 Lasso-Logistic analyses, we found that combinations of 12 imaging features and parameters consistently occurred at least 900 times. The 12 imaging features included: wavelet.HHH_glszm_SizeZoneNonUniformityNormalized; wavelet.HLL_glszm_LowGrayLevelZoneEmphasis, wavelet.HLL_glszm_SmallAreaLowGrayLevelEmphasis, wavelet.LHL_firstorder_Skewness, wavelet.LHL_glcm_ClusterShade, wavelet.LHL_g lcm_Correlation, wavelet.LHL_glcm_MCC, wavelet.LLH_firstorder_Median, wavelet.LLH_glcm_Idn, wavelet.LLH_glszm_GrayLevelNonUniformityNormalized, wavelet.LLH_glszm_SmallAreaLowGrayLevelEmphasis, wavelet.LLH_glszm_ZoneEntropy; S4. Use multivariate logistic regression to construct a corresponding RAS gene status prediction model based on the 12 radiomics features obtained in S3; The prediction model is a radiomics prediction model, and the radiomics score calculation formula is: 3.5512-6.1777*wavelet.HHH_glszm_SizeZoneNonUniformityNormalized 1.8147*wavelet.HLL_glszm_LowGrayLevelZoneEmphasis+1.3087*wavelet.HLL_glszm_SmallAreaLowGrayLevelEmphasis-4.0329*wavele t.LHL_firstorder_Skewness-2.1221*wavelet.LHL_glcm_ClusterShade+1.0033*wavelet.LHL_glcm_Correlation-3.0967*wavelet.LHL_g lcm_MCC+0.4027*wavelet.LLH_firstorder_Median+0.5078*wavelet.LLH_glcm_Idn-0.2159*wavelet.LLH_glszm_GrayLevelNonUniformit yNormalized+6.8791*wavelet.LLH_glszm_SmallAreaLowGrayLevelEmphasis+4.6685*wavelet.LLH_glszm_ZoneEntropy; Set the cutoff value to 50%.

2. The method for constructing a model for predicting the RAS gene status of CRLM patients based on radiomics features according to claim 1, characterized in that: The segmentation of the region of interest described in S2 uses a three-dimensional semi-automatic segmentation method to segment the region of interest in the portal venous phase CT image.

3. The method for constructing a model for predicting the RAS gene status of CRLM patients based on radiomics features according to claim 1, characterized in that: The image features extracted in S2 include: (1) The first-order features describe the distribution of voxel intensities in ROIs; (2) Shape-based features describe the three-dimensional intuitive features of ROIs, including size and shape at the two-dimensional and three-dimensional levels; (3) Texture features extracted based on five texture matrices: a) grayscale co-occurrence matrix, b) grayscale region size matrix, c) grayscale run length matrix, d) neighborhood grayscale difference matrix, and e) grayscale correlation matrix; (4) Wavelet features: Wavelet filtering is added to the original image to reduce noise while extracting detailed high-dimensional image group features.

4. A storage medium having a plurality of instructions stored thereon, characterized in that: The instructions are suitable for being loaded and executed by a processor to implement the steps of the method for constructing a model for predicting the RAS gene status of CRLM patients based on imaging genomics features as described in any one of claims 1-3.

5. A terminal, characterized in that: include: A processor and a storage medium communicatively connected to the processor, the storage medium being suitable for storing a plurality of instructions; the processor being suitable for calling the instructions in the storage medium to execute the steps of the method for constructing a model for predicting the RAS gene status of CRLM patients based on radiomics features as described in any one of claims 1 to 3.

6. A device for predicting the RAS gene status of patients with colorectal cancer liver metastasis, characterized in that: The method comprises constructing a prediction model by the steps of the method for constructing a model for predicting the RAS gene status of CRLM patients based on imaging genomics features according to any one of claims 1 to 3.

Images