Comprehensive epithelial ovarian cancer model based on combination of clinical features, CT image features and radiomics features and training method
By combining clinical features, CT imaging features and radiomic features, the problem of early diagnosis and subtype distinction between ovarian cancer in the prior art is solved, and higher diagnostic accuracy is achieved, providing personalized treatment. Scientific basis.
Patent Information
- Application Number
- CN202410961240.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-11-12
AI Technical Summary
The existing ovarian cancer diagnosis methods are difficult to accurately diagnose and distinguish epithelial ovarian cancer subtypes in the early stage. Imaging examinations have limitations in distinguishing subtypes, and the existing technology is difficult to meet the needs of personalized treatment.
A comprehensive diagnostic model based on clinical features, CT imaging features and radiomic features is adopted. Through feature processing, feature extraction and feature selection, combined with a penalty logistic regression algorithm, a comprehensive model is established to optimize features and hyperparameters, and to improve diagnostic performance.
It significantly improves the accuracy of preoperative diagnosis of epithelial ovarian cancer, provides a scientific basis for formulating personalized treatment plans and predicting prognosis, and has important clinical application value.
Smart Images

Figure CN118942598A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of medical imaging and relates to a comprehensive diagnostic model and a training method for epithelial ovarian cancer. Background Art
[0002] Ovarian cancer (OC) is one of the most lethal malignant tumors in the female reproductive system. Among them, epithelial ovarian cancer (EOC) accounts for the vast majority of ovarian malignancies (about 80%) and about 90% of all primary ovarian malignancies. Its high heterogeneity is significant in etiology, molecular biology and prognosis. There are at least five histological subtypes in EOC, and there are significant differences in growth rate, invasiveness, prognosis and recurrence risk between different histological subtypes. Therefore, accurate preoperative differentiation of EOC subtypes is crucial for developing personalized treatment plans and predicting prognosis.
[0003] Because the ovaries are located deep in the pelvic cavity, early lesions often lack specific symptoms, resulting in about 70% of patients being diagnosed in the late stage, making early diagnosis difficult. Secondly, the diagnostic methods also have limitations. Existing screening methods such as CAl25 testing and transvaginal ultrasound, alone or in combination, cannot achieve satisfactory early screening results. Although imaging diagnostic techniques such as CT and MRI can clearly define the morphology and extent of invasion of the tumor, they have limitations in distinguishing EOC subtypes. With the deepening of our understanding of the tumor microenvironment and molecular biology, tailored personalized treatment plans have become a trend, but existing technical solutions are difficult to meet this demand.
[0004] Radiomics is an emerging technology that extracts quantitative image features from medical images and is used to improve diagnostic performance, predict prognosis, and assist clinical decision-making. In recent years, radiomics analysis has been increasingly used in tumor diagnosis, especially in differentiating benign and malignant tumors, predicting tumor risk levels, and assessing the risk of lymph node metastasis and tumor recurrence. Radiomics analysis based on CT and MRI has been shown to be useful for differentiating EOC subtypes and has good diagnostic performance.
[0005] In order to overcome the shortcomings of existing technologies and improve the accuracy of preoperative diagnosis of EOC, it is necessary to develop a comprehensive diagnostic model and establish and train a program that combines clinical features, CT imaging features, and radiomics features to achieve more accurate diagnosis and prognosis prediction. Summary of the invention
[0006] In order to solve some problems existing in the above-mentioned prior art, the present invention provides a comprehensive model of epithelial ovarian cancer based on the combination of clinical features, CT imaging features and radiomics features and a training method. The establishment and training of the comprehensive model of epithelial ovarian cancer is achieved by integrating clinical features, CT imaging features and radiomics features.
[0007] To achieve the above object, the present invention provides the following technical solutions:
[0008] A training method for a comprehensive diagnosis model of epithelial ovarian cancer (EOC) based on the combination of clinical features, CT imaging features and radiomics features. The specific training steps of the training method for the comprehensive diagnosis model of epithelial ovarian cancer (EOC) are as follows:
[0009] Step 1: The clinical data of EOC patients were processed. The inclusion criteria were as follows: (1) EOC confirmed by pathology; (2) enhanced CT examination within 1 month before surgery; (3) complete medical history and surgical pathology report. Exclusion criteria included: (1) treatment before CT examination (such as neoadjuvant chemotherapy); (2) the presence of obvious ascites, which obscured the lesion and made it impossible to segment the lesion; (3) obvious artifacts in the CT image.
[0010] Step 2: Obtain the clinical data and imaging data of the subjects, wherein the clinical data include continuous variables (age, preoperative CA-125 level) and categorical variables (history of endometriosis, history of venous thromboembolism); the imaging data are CT portal venous phase images.
[0011] Step 3: Each patient's CT images were independently reviewed using GE Healthcare's commercial viewing software and ten typical imaging features were evaluated.
[0012] Step 4: Using Dr.Wise research platform from Shenrui, a junior radiologist manually segmented the ovarian lesions in each CT image. After the segmentation was completed, another junior radiologist randomly selected a specified number of patient CT images from the training set to calculate the inter-class correlation coefficient (ICC).
[0013] Step 5: Use professional medical image processing software (such as 3D Slicer, PyRadiomics, etc.) to preprocess the CT images, including denoising, standardization, and resampling.
[0014] Step 6: Wavelet decomposition and LoG (Laplacian of Gaussian) enhancement methods were used to identify features of the original image. Then, the radiomics package (Pyradioomics) in Python was used to extract radiomics features from each tumor ROI segmented on the portal vein phase CT image.
[0015] Step 7: Use the hot encoder to encode the categorical variables in the clinical characteristics, and use the z-score normalization method to standardize the continuous variables. For the radiomics features, ICC (inter-class correlation coefficient) is selected.
[0016] Features with a p ≥ 0.8 were included in the subsequent analysis, and the included radiomic features were standardized using the z-score normalization method.
[0017] Step 8: Use the joint mutual information maximization (JMIM) technique combined with the maximum relevance minimum redundancy (MRMR) technique to select the most relevant features and perform feature dimensionality reduction.
[0018] Step 9: The penalized logistic regression algorithm was used to establish the traditional model (TM), the radiomics model (RS), and the integrated model (IM) based on the combination of clinical features with CT imaging features and radiomics features. During the 10-fold cross-validation process, a grid search was performed to optimize the number of selected features and hyperparameters, and penalized logistic regression analysis was performed to maximize the AUC (area under the curve) of the cross-validation results. After determining the optimal number of features and hyperparameters, the model was retrained using the entire training set.
[0019] Step 10: Based on the principle of maximizing sensitivity, the cutoff value of the ROC (receiver operating characteristic) curve was calculated by 10-fold cross validation to evaluate the accuracy, sensitivity, and specificity of the model. The DeLong test's method was used to compare the statistical differences in the ROC curves between the traditional model, the radiomics model, and the comprehensive model.
[0020] The imaging features in step three include (1) laterality (unilateral or bilateral), (2) mass margin (clear or unclear boundaries), (3) mass shape (oval, lobed or irregular), (4) mass characteristics (simple cystic, unilocular cystic solid, multilocular cystic solid or solid), (5) maximum mass diameter on axial CT, (6) mural nodules (presence or absence), (7) growth pattern of mural nodules (broad-based nodular structure or polypoid structure), (8) height, width and aspect ratio (HWR) of the largest mural nodule, (9) peritoneal seeding (presence or absence), and (10) ascites (presence or absence).
[0021] In step 4, there is usually a senior doctor who verifies the segmentation results of the junior doctor.
[0022] The beneficial effects of the present invention are:
[0023] The training method of the comprehensive diagnostic model for epithelial ovarian cancer based on the combination of clinical characteristics with CT imaging characteristics and radiomics characteristics proposed in the present invention will significantly improve the accuracy of preoperative diagnosis of EOC, provide a scientific basis for formulating personalized treatment plans and predicting prognosis, and has important clinical application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] In order to facilitate the understanding of those skilled in the art, the present invention is further described below with reference to the accompanying drawings:
[0025] Figure 1 A flowchart of the training method of the comprehensive model of epithelial ovarian cancer based on the combination of clinical features with CT imaging features and radiomics features;
[0026] Figure 2 This is a patient screening flowchart of the OCCC subtype classification model of the embodiment;
[0027] Figure 3 This is an example diagram of the wall nodule growth pattern of the OCCC subtype classification model of the embodiment;
[0028] Figure 4 This is an example diagram of the definition of mural nodules in the OCCC subtype classification model of the embodiment;
[0029] Figure 5 is a ROC curve diagram of the OCCC subtype classification model in the embodiment; DETAILED DESCRIPTION
[0030] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The following examples are only used to more clearly illustrate the technical solutions of the present invention, and cannot be used to limit the scope of protection of the present invention.
[0031] The present invention is not limited to the following embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0032] The present invention is described in detail below according to specific embodiments and drawings:
[0033] The present invention uses the histological subtype OCCC (ovarian clear cell carcinoma) in EOC (epithelial ovarian cancer) as an example to use the training method of the present invention to establish and train the model of OCCC subtype, which specifically includes the following steps:
[0034] 1. Patient screening: Fushun Central Hospital retrospectively collected patients with epithelial ovarian cancer (EOC) who underwent surgical treatment in the center from January 2016 to December 2022. Inclusion criteria were as follows: (1) pathologically confirmed EOC; (2) enhanced CT examination within 1 month before surgery; (3) complete medical history and surgical pathology report. Exclusion criteria included: (1) treatment before CT examination (such as neoadjuvant chemotherapy); (2) the presence of obvious ascites, which obscured the lesion and made it impossible to segment the lesion; (3) obvious artifacts in the CT images. Finally, 349 EOC patients were consecutively enrolled in Fushun Central Hospital. Patients were divided into a training set (n = 280) and an internal test set (n = 69) according to the time of surgery. In addition, two external centers registered using the same criteria, and patients who met the criteria formed two independent external test sets (ETS1 = 74, ETS2 = 34). Figure 1 The patient selection flow chart for this example. All patients underwent unilateral / bilateral salpingo-oophorectomy or hysterectomy, and the histological subtype of the tumor was determined by postoperative pathological section microscopy and immunohistochemistry results. In this example, the patients were divided into OCCC group and Non-OCCC group based on the surgical pathology report of the patients as the gold standard.
[0035] 2. Clinical data collection: Clinical characteristics that were reported in previous studies and were helpful for differential diagnosis of OCCC subtypes and were easily obtained preoperatively were collected, including age, preoperative CA-125 level, history of endometriosis, and history of venous thromboembolism.
[0036] 3. Evaluation of CT morphological features: The CT images of each patient were independently reviewed by a senior physician with 12 years of experience and a junior physician with 5 years of experience in gynecological tumor imaging diagnosis using commercial viewing software (American Central Radiology RA1000, GE Healthcare, USA). The following 10 features of the CT images of each patient were evaluated without knowing the pathological results: (1) laterality (unilateral or bilateral), (2) mass margin (well-defined or ill-defined), (3) mass shape (oval, lobulated, or irregular), (4) mass characteristics (simple cystic, unilocular cystic solid, multilocular cystic solid, or solid), (5) maximum mass diameter on axial CT, (6) mural nodules (presence or absence), (7) growth pattern of mural nodules (broad-based nodular structure or polypoid structure), (8) height, width, and height-to-width ratio (HWR) of the largest mural nodule, (9) peritoneal seeding (presence or absence), and (10) ascites (presence or absence). Lesions were classified according to their composition: (1) "simple cystic," meaning lesions without a solid component, (2) "solid," meaning lesions with more than two-thirds of a solid component, and (3) the remainder, called "cystic-solid." The evaluation also considered the growth pattern of the mural nodules, with a "broad-based nodular architecture" indicating multiple nodules growing toward the center of the cyst and a "polypoid architecture" indicating a few nodules growing eccentrically on the inner surface of the cyst ( Figure 3 The height of the mural nodules was defined as the maximum vertical length from the cyst wall to the top of the nodule, and the width was defined as the maximum vertical length perpendicular to the height ( Figure 4 ). In case of discordance between the two radiologists, a consensus was reached on the lesion assessment, and the agreement of their assessment results was calculated.
[0037] 4. Lesion segmentation: CT data are uploaded to the Dr.Wise research platform (Shenrui Company). A radiologist with 5 years of CT reading experience manually segments the ovarian lesions on each CT image. Thereafter, each segmentation is verified by a senior radiologist with 19 years of gynecological CT reading experience. In addition, another junior radiologist randomly selects 30 patient CT images from the training set, and calculates the reproducibility (interclass correlation coefficient, ICC) of the radiomics features extracted by different operators, aiming to select radiomics features with high stability. The two radiologists can refer to the patient's pathology report to ensure the consistency of each segmented lesion with the reported primary ovarian lesion. If there are metastatic lesions next to the primary tumor, the diagnostician can refer to the surgical records to accurately distinguish the boundaries between the primary tumor and the metastasis to ensure the accuracy of the segmentation results.
[0038] 5. Feature extraction: To mitigate the impact of image spatial resolution on radiomic features, the images were first normalized and resampled to obtain consistent image resolution. All original CT images were resampled to 1 × 1 × 1 mm 3 The pixel spacing was adjusted to reduce the potential impact of inconsistent spatial resolution. Wavelet decomposition and LoG (Laplacian of Gaussian) enhancement methods were used to identify features of the original images under different σ values. Subsequently, 1218 radiomic features were extracted from each tumor ROI segmented on the portal phase CT images using the radiomics package in Python (https: / / pyradiomics.readthedocs.io), including 252 first-order features, 14 shape features, 308 gray-level co-occurrence matrix (GLCM) features, 224 gray-level size zone matrix (GLSZM) features, 224 gray-level run length matrix (GLRLM) features, and 196 gray-level dependency matrix (GLDM) features.
[0039] 6. Feature selection: For the four clinical features collected, categorical variables (including endometriosis and history of venous thromboembolism) were encoded using a one-hot encoder, while the remaining continuous variables (including age and preoperative CA-125 level) were standardized using z-score normalization. For radiomic features, features with an ICC (interclass correlation coefficient) ≥ 0.8 were selected for subsequent analysis. Subsequently, the included radiomic features were standardized using the z-score normalization method. The feature selection process was based on the joint mutual information maximization (JMIM) technique combined with the maximum relevance minimum redundancy (MRMR) technique to select the most relevant features.
[0040] 7. Model construction (TM\IM\RS): This study established three OCCC subtype diagnostic models (IM-comprehensive model, TM-traditional model, RS-radiomics model). The traditional model was built using selected clinical and traditional imaging features to simulate the diagnostic performance in clinical scenarios; the traditional imaging features only include feature selection based on the evaluation of senior radiologists. The radiomics model was built using selected key radiomics features, and the output probability of the radiomics model was regarded as the radiomics feature (RS); the comprehensive model was built using radiomics features, clinical features, and conventional imaging features.
[0041] After feature selection using the JMIM method, six features, including age, tumor laterality, mural nodules, mural nodule HWR (height-to-width ratio), peritoneal seeding metastasis, and peritoneal effusion, were identified as the best feature subsets for building the traditional model (TM). For the radiomics model (RS), 386 RS features with an ICC ≥ 0.8 were initially screened. Five radiomics features (wavelet-HHLglcmCorrelation, wavelet-LLLfirstorder90Percentile, wavelet-LLHfirstorderVariance, originalfirstorder90Percentile, and log-sigma-4-0-mm-3DglcmContrast) were selected based on the JMIM method to build the radiomics model. In addition, the output scores of the selected five radiomics feature models (rad_score) were merged into one radiomics feature model (RS). The best feature subsets, including age, mural nodules, and RS, were selected by the JMIM method to build the integrated model (IM).
[0042] 8. Model performance evaluation: In the three test sets, the AUC (area under the curve) values of the traditional model (TM), radiomics model (RS) and integrated model (IM) were calculated respectively. The cutoff value of the ROC (receiver operating characteristic) curve was calculated by 10-fold cross-validation based on the principle of maximizing sensitivity to evaluate the accuracy, sensitivity and specificity of the model. The DeLong test's method was used to compare the statistical differences in the ROC curves between the traditional model, radiomics model and integrated model. In the internal test set, the diagnostic performance of the integrated model (IM) was better than that of the (TM) traditional model (AUC: 0.890 vs. 0.840) and the (RS) radiomics model (AUC: 0.890 vs. 0.811). Compared with the internal test set, the performance of the three prediction models in the external test set 1 was more stable, and the AUC values were comparable to those of the internal test set. The ROC curves of the three models are shown in Figure 2. Figure 5 shown.
[0043] In summary, the EOC classification comprehensive model and training method used in the present invention show good diagnostic sensitivity and accuracy in OCCC subtypes. Relevant studies show that the average diagnostic sensitivity of senior doctors is 39.3%, and the average diagnostic sensitivity of junior doctors is 42.9%. The diagnostic sensitivity of the embodiments used in the present invention is much higher than the diagnostic sensitivity in relevant studies. It is foreseeable that the training method of the comprehensive diagnostic model of epithelial ovarian cancer based on the combination of clinical characteristics, CT imaging characteristics and radiomics characteristics proposed in the present invention will significantly improve the preoperative diagnostic accuracy of EOC, provide a scientific basis for formulating personalized treatment plans and predicting prognosis, and have important clinical application value and promotion prospects.
[0044] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A comprehensive diagnostic model and training method for epithelial ovarian cancer based on the combination of clinical features, CT imaging features and radiomics features, characterized in that: The specific training steps of the comprehensive diagnosis model for epithelial ovarian cancer based on the combination of clinical features, CT imaging features and radiomics features are as follows: Step 1: The clinical data of EOC patients were processed. The inclusion criteria were as follows: (1) EOC confirmed by pathology; (2) enhanced CT examination within 1 month before surgery; (3) complete medical history and surgical pathology report. Exclusion criteria included: (1) treatment before CT examination (such as neoadjuvant chemotherapy); (2) the presence of obvious ascites, which obscured the lesion and made it impossible to segment the lesion; (3) obvious artifacts in the CT image. Step 2: Obtain the clinical data and imaging data of the subjects. Step 3: Examine each patient's CT images and evaluate ten typical imaging features. Step 4: A radiologist manually segmented the ovarian lesions in each CT image. After the segmentation was completed, another radiologist randomly selected a specified number of patient CT images from the training set and calculated the inter-class correlation coefficient (ICC). Step 5: Use professional medical image processing software (such as 3D Slicer, PyRadiomics, etc.) to preprocess the CT images, including denoising, standardization, and resampling. Step 6: Wavelet decomposition and LoG (Laplacian of Gaussian) enhancement methods were used to identify features of the original image. Then, the radiomics package (Pyradioomics) in Python was used to extract radiomics features from each tumor ROI segmented on the portal vein phase CT image. Step 7: The categorical variables in the clinical characteristics were encoded using a hot encoder, and the continuous variables were standardized using the z-score normalization method. For radiomic features, features with ICC (interclass correlation coefficient) ≥ 0.8 were selected for subsequent analysis, and the included radiomic features were standardized using the z-score normalization method. Step 8: Use the joint mutual information maximization (JMIM) technique combined with the maximum relevance minimum redundancy (MRMR) technique to select the most relevant features and perform feature dimensionality reduction. Step 9: The penalized logistic regression algorithm was used to establish the traditional model (TM), the radiomics model (RS), and the integrated model (IM) based on the combination of clinical features with CT imaging features and radiomics features. A grid search was performed during the 10-fold cross-validation process to optimize the number of selected features and hyperparameters, and penalized logistic regression analysis was performed to maximize the AUC (area under the curve) of the cross-validation results. After determining the optimal number of features and hyperparameters, the model was retrained using the entire training set. Step 10: Based on the principle of maximizing sensitivity, the cutoff value of the ROC (receiver operating characteristic) curve was calculated by 10-fold cross validation to evaluate the accuracy, sensitivity, and specificity of the model. The DeLong test's method was used to compare the statistical differences in the ROC curves between the traditional model, the radiomics model, and the comprehensive model.
2. The method according to claim 1, characterized in that: The clinical data in step 1 include continuous variables (age, preoperative CA-125 level) and categorical variables (history of endometriosis, history of venous thromboembolism); the imaging data are CT portal venous phase images.
3. The method according to claim 1, characterized in that The imaging features in step three include (1) laterality (unilateral or bilateral), (2) mass margin (clear or unclear boundaries), (3) mass shape (oval, lobed or irregular), (4) mass characteristics (simple cystic, unilocular cystic solid, multilocular cystic solid or solid), (5) maximum mass diameter on axial CT, (6) mural nodules (presence or absence), (7) growth pattern of mural nodules (broad-based nodular structure or polypoid structure), (8) height, width and aspect ratio (HWR) of the largest mural nodule, (9) peritoneal seeding (presence or absence), and (10) ascites (presence or absence).
4. The method according to claim 1, characterized in that The model trained in step nine when applied to EOC subtype OCCC (ovarian clear cell carcinoma) has the following model features: (1) The best feature subset for constructing the traditional model (TM) is age, tumor laterality, wall nodules, wall nodule HWR (height-to-width ratio), peritoneal seeding metastasis, and peritoneal effusion. (2) The features used to construct the radiomics model (RS) are wavelet-HHLglcmCorrelation, wavelet-LLLfirstorder90Percentile, wavelet-LLHfirstorderVariance, originalfirstorder90Percentile, and log-sigma-4-0-mm-3DglcmContrast. (3) The features used by the integrated model (IM) are age, wall nodules, and RS feature subset (the output scores of the selected five radiomics feature models (rad_score) are merged into one radiomics feature model (RS)).
Citation Information
Cited By
Intelligent auxiliary processing system for abdominal surgery images
CN119679525A
