An ovarian cancer survival prediction system based on multiple ultrasound examination image-based statistical quantities of characteristics

Abstract

The application discloses an ovarian cancer survival prediction system based on a plurality of ultrasonic examination image-based statistical quantities of features, comprising a data acquisition module, a feature extraction module, a data preprocessing module, a sample change stability feature screening module, a survival prediction feature screening module and a survival prediction model construction module. The application carries out statistical analysis on the multi-lesion multi-angle features collected by ultrasonic examination of patients, constructs patient ultrasonic image features by comprehensively considering the results of the lesions observed from different angles in the ultrasonic examination, carries out Cox univariate screening by repeated sampling, and counts the frequency of the features selected in the feature screening process to determine the features with stability to sample changes. Finally, the features with high correlation to survival prediction are screened in combination with prognosis information, and the ovarian cancer survival prediction system is established. The system has better effect than the model using only the features of the largest tumor region, and has important significance for improving the robustness and survival prediction effect of the ovarian cancer survival prediction system based on medical image-based features.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of medical imaging and cancer medicine, and relates to a cancer survival prediction system based on machine learning and multiple statistical methods, in particular to an ovarian cancer survival prediction system based on multiple ultrasound imaging feature statistics. BACKGROUND

[0002] Ovarian cancer is the fifth most common cancer in women and ranks fifth in cancer-related deaths. Its incidence and mortality rates rank second and first among female reproductive system cancers, respectively. Due to the lack of effective screening and early diagnosis tools, most ovarian cancer patients have already spread when diagnosed. The 5-year survival rate of ovarian cancer patients is only 49.1%. Therefore, timely diagnosis and effective prognosis evaluation of ovarian cancer are of great significance to improve the health status of ovarian cancer patients and prolong the survival time of ovarian cancer patients.

[0003] Ultrasound examination is the most common non-invasive diagnostic examination for ovarian cancer patients. Compared with magnetic resonance imaging (MRI) and computed tomography (CT), ultrasound examination is inexpensive, non-invasive, easy to obtain and has no contraindications, and is a common examination method for ovarian cancer screening.

[0004] Imaging-based features extracted from medical images have application value in the diagnosis and survival prediction of cancer patients. In a retrospective study, it is noted that the ultrasound images saved by patients often have two-dimensional images of tumor lesions observed from multiple angles at the same time, which is different from the three-dimensional form of CT and MRI, which is a stack of multiple consecutive two-dimensional images. Only one examination can more completely display the patient's lesion condition, so the results of multiple ultrasound examinations are of great significance for the survival prediction of ovarian cancer patients. SUMMARY

[0005] The present application aims to provide an ovarian cancer survival prediction system based on multiple ultrasound imaging feature statistics. The method is based on machine learning and multiple statistical methods, which can effectively utilize the information of the patient's multi-angle ultrasound images. Compared with the traditional imaging-based ovarian cancer survival prediction method based on the largest tumor ultrasound image, the present application has better performance, which is of great significance for improving the ovarian cancer survival prediction based on imaging, and can be applied to other studies where a single examination cannot fully display the tumor condition.

[0006] The purpose of the present application is achieved by the following technical solution: an ovarian cancer survival prediction system based on multiple ultrasound imaging feature statistics of patients, comprising the following modules:

[0007] (1) Data acquisition module: used to acquire survival follow-up data of ovarian cancer patients and transvaginal or transabdominal ultrasound images of patients before radical ovarian cancer surgery and comprehensive staging surgery. The ovarian cancer tumor area is manually marked in the ultrasound images to form an original dataset including ultrasound images and real follow-up data.

[0008] (2) Feature extraction module: The original ultrasound images are transformed, and radiomics features are extracted based on the manually labeled ovarian cancer tumor regions after the image transformation. Radiomics features are calculated based on the ovarian cancer tumor regions in each ultrasound image before and after the transformation of multiple ultrasound examinations of the same patient. At the same time, the patient's prognostic information is obtained based on the patient's follow-up data, and the ultrasound survival analysis dataset is constructed together.

[0009] (3) Data preprocessing module: Some features extracted after image transformation have extremely low discrimination. After removing low variance features, the standard deviation and mean of the image omics features are unified to form a standardized dataset.

[0010] (4) Sample change stability feature screening module: The standardized dataset is sampled multiple times and screened based on Cox univariate. The sample change stability feature is determined according to the frequency of feature selection. The sample change stability feature is still selected with a high probability when the training dataset changes.

[0011] (5) Survival prediction feature screening module: For the sample change stability features, further screen out the features related to the survival prediction task according to the task requirements; and calculate statistical features based on the features related to the survival prediction task in multiple images of the same patient, and use multivariate Cox screening on the statistical features;

[0012] (6) Survival prediction model construction module: Based on the features selected in the survival prediction feature screening module, construct the ovarian cancer ultrasound image survival prediction model.

[0013] Furthermore, in the data acquisition module, the ultrasound image is a transvaginal ultrasound or transabdominal ultrasound, which does not include Doppler blood flow signals; the artificially marked ovarian cancer tumor area is clearly visible in the ultrasound image and the lesion can be completely displayed; the grayscale of the ultrasound image is normalized to [0, 255].

[0014] Further, the feature extraction module, the image transformation includes calculating the square value, square root value, exponential value, logarithmic value, gradient value, local binary pattern and four wavelet transforms (LL, LH, HL, HH) of pixels; the feature extraction is performed on each ultrasound image separately, the feature extraction module acts on all ultrasound images obtained by the data acquisition module, and the imageomics features are combined with the prognosis information in the patient follow-up data to jointly construct an ultrasound survival analysis data set. The calculated features include shape features (all image transformations are the same and are calculated only once), first-order statistical features, and texture features; the texture features include gray level co-occurrence matrix features, gray level run length matrix features, gray level zone size matrix features, domain gray difference matrix features, and gray level dependency matrix features.

[0015] Further, the ultrasound survival analysis data set is represented as (F, e, t), wherein F is an ultrasound examination imageomics feature, e and t are prognosis information extracted from patient follow-up data;

[0016] F = [f1, f2,... f m , f square1 , f square2 ,... f squarem , f waveletHH1 ,..., f waveletHHm ] ∈ R n×p

[0017] wherein R is a real number field, n is a sample size, p is an imageomics feature number of an original image and all image transformations, f represents an imageomics feature extracted after the picture is transformed, and the subscript indicates that the feature is extracted from an original image or a tumor region after image transformation, f1, f2,... f m respectively represent features extracted from an original ultrasound image, f square1 , f square2 ,... f squarem respectively represent imageomics features extracted after square image transformation, f waveletHH1 ,..., f waveletHHm represent imageomics features extracted after wavelet transform high-high component image transformation, wherein m represents an imageomics feature number extracted from an original image or a single image transformation;

[0018] e = [e1, e2,..., e n ] ∈ {0, 1} n is an event label of n samples, in a total survival prediction task, e indicates whether the sample is alive; and in a recurrence-free survival prediction task, e indicates whether the sample relapses;

[0019] t = [t1, t2,..., t n ] ∈ R nFor n samples of event time, in the total survival prediction task, if the patient dies t, it indicates that the sample survives the time, if the patient survives t, it indicates the follow-up time; in the recurrence-free survival prediction task, if the patient relapses t, it indicates the recurrence-free survival time of the sample, if the patient does not relapse t, it indicates the follow-up time.

[0020] Further, in the data preprocessing module, the features without discrimination (variance < 0.0001) are first removed, and all extracted image features F are subjected to z-score standardization, the standard deviation becomes 1, and the mean becomes 0, the standardized matrix is recorded as F' = [f'1, f'2,... f' n], and the prognosis information is used to form a standardized data set (F', e, t). m , f' square1 , f' square2 ,... f' squarem ,... f' waveletHHm ].

[0021] Further, in the sample change stability feature screening module, a part of the standardized data set samples are randomly extracted and repeated multiple times, cox univariate screening is performed on the extracted samples each time, and the frequency of being selected by different samples is counted, when the frequency of being selected is greater than 80%, the sample change stability feature F1 is considered.

[0022] Further, in the survival prediction feature screening module, according to the prognosis information of the patient and the screened sample change stability feature, the 2-year survival condition is used as a label, a part of the sample change stability feature samples are randomly extracted and repeated multiple times, the least absolute selection and shrinkage operator algorithm is used to screen the extracted samples each time, and the frequency of being selected by different features is counted, when the frequency of being selected is greater than 80%, the feature F2 related to the survival prediction task is considered; after screening the features related to the survival prediction task, a plurality of statistical features (including maximum value, minimum value, mean value, median) are calculated according to the features related to the survival prediction task in the multiple images of the same patient, and cox multivariate screening is used on the plurality of statistical features, and the statistical feature with the smallest p value is reserved as the final selected feature F3.

[0023] Further, in the survival prediction model construction module, the survival prediction model adopts a cox model.

[0024] The beneficial effects of the present application are:

[0025] This invention provides an ovarian cancer survival prediction system based on radiomics features from multiple ultrasound examinations of a patient. Addressing the issue that a single ultrasound examination cannot fully reveal the patient's tumor lesion information, this invention employs machine learning and various statistical methods, combined with survival analysis, to comprehensively construct patient ultrasound image features from lesion results observed from different angles during ultrasound examinations. The ovarian cancer survival prediction system is built based on the statistical values ​​of radiomics features from multiple ultrasound examinations, which can comprehensively utilize the features of multiple ultrasounds, resulting in better performance than ovarian cancer survival prediction systems built solely based on ultrasound images of the largest observed tumor. The feature selection method proposed in this invention, which integrates individual samples and patient feature statistics, performs better in ovarian cancer survival prediction systems than feature selection methods that only consider feature statistics. This invention is of great significance for improving the robustness and performance of radiomics-based ovarian cancer survival prediction systems based on multiple examinations. This system can fully utilize information from multiple angles and multiple lesions in a patient's ultrasound examination, achieving better results than models that only use features of the largest tumor region. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the ovarian cancer survival prediction system constructed based on the statistical characteristics of multiple ultrasound images of patients according to the present invention.

[0027] Figure 2 This is a schematic diagram of the feature extraction module of the ovarian cancer survival prediction system constructed based on the statistical characteristics of multiple ultrasound images of patients in this invention.

[0028] Figure 3 This is a flowchart of the sample change stability feature screening module and the survival prediction feature screening module of the ovarian cancer survival prediction system constructed based on the statistical characteristics of multiple ultrasound images of patients in this invention.

[0029] Figure 4Performance comparison of the ovarian cancer survival prediction system based on patient multi-portion ultrasonic examination imaging feature statistics constructed by the present application and other methods. (A) Comparison results of different feature screening methods in overall survival prediction, the method used in the present application is Frequency Appearance in Multiple Univariate pre-Screening (FAMUS) combined with Least absolute shrinkage and selection operator (LASSO) screening, abbreviated as FAMUS_LASSO, and the remaining methods are feature screening based on full feature statistics calculation. (B) Comparison results of different feature screening methods in recurrence-free survival prediction, FAMUS_LASSO is the method proposed in the present application, and the remaining methods are feature screening based on full feature statistics calculation. (C) Comparison results of different survival analysis models in overall survival prediction, the same feature screening method is used. (D) Comparison results of different survival analysis models in recurrence-free survival prediction, the same feature screening method is used. DETAILED DESCRIPTION

[0030] The present application will be further described in detail below in combination with the accompanying drawings and examples, but the present application is not limited thereto.

[0031] 1. Experimental materials:

[0032] Experimental sample data: ovarian cancer transvaginal ultrasound and abdominal ultrasound images, without Doppler blood flow signals; the survival prediction tasks corresponding to the data are overall survival (OS) prediction tasks and recurrence-free survival (RFS) prediction tasks; the ovarian cancer ultrasound data collected from a hospital is divided into a training cohort and an internal validation cohort according to the scanning time, and the data collected from another hospital is used as an external validation cohort.

[0033] Operating system: CentOS 7

[0034] Software: Python

[0035] 2. Experimental method, as shown in Figure 1

[0036] ​(1) Data acquisition module, used to acquire ultrasound images, patient prognosis information and artificially marked ovarian cancer tumor areas in ultrasound images; the artificially marked ovarian cancer tumor areas are clearly visible in ultrasound images and the lesions can be fully displayed; the grayscale of ultrasound images is normalized to [0, 255].

[0037] (2) In the feature extraction module, features are extracted from the original ultrasound image and its series of image transformations. The image transformations used include calculating the square value, square root value, exponential value, logarithm value, gradient value, local binary mode (LBP), and four wavelet transforms (wavelet LL, wavelet LH, wavelet HL, wavelet HH). The calculated features include shape features, first-order statistical features, and texture features. The texture features include gray-level co-occurrence matrix features, gray-level run-length matrix features, gray-level region size matrix features, neighborhood gray-level difference matrix features, and gray-level dependency matrix features. Further, in the feature extraction module, the original dataset (X, e, t) is converted into ([X, X... square X squareroot , ..., X waveletHH The dataset is generated by extracting radiomics features from the original image and various transformed images (f1, f2, ..., ft) and then converting them into an ultrasound survival analysis dataset (F, e, t). Where F = [f1, f2, ..., ft] m f square1 f square2 , ...f squarem , ...f waveletHHm ]∈R n×p Where R is the real number field, n is the sample size, p is the number of radiomics features of the original image and all image transformations, X represents the ultrasound image dataset, f represents the radiomics features extracted from the transformed images, m represents the number of radiomics features extracted from the original image or a single image transformation, and e = [e1, e2, ..., e] n ]∈{0,1} n Let t be the event label for n samples. In a total survival prediction task, e represents whether the sample is alive; while in a relapse-free survival prediction task, e represents whether the sample has relapsed. t = [t1, t2, ..., t...] n ]∈R n For n samples, t represents the time of survival of the sample (if the patient dies) or the follow-up time (if the patient survives) in the overall survival prediction task; and t represents the time of relapse-free survival of the sample (if the patient relapses) or the follow-up time (if the patient does not relapse) in the relapse-free survival prediction task.

[0038] Specifically, the feature extraction module is configured to extract features from the image data as shown in Figure 2 Figure 1 There are 11 images in total, which are the original image, the pixel value square processing image, the pixel value square root processing image, the pixel value exponential processing image, the pixel value logarithmic processing image, the pixel value gradient calculation image, the pixel value local binary pattern image, and four two-dimensional image wavelet transforms (LL, LH, HL, and HH). When calculating the features, there are 9 two-dimensional shape features, and these features are the same in all image transformations, so they only need to be calculated once. The remaining features need to be recalculated for each image transformation, including 18 first-order statistical features, 24 Gray-level co-occurrence matrix (GLCM) features, 16 Gray-level size zone matrix (GLSZM) features, 16 Gray-level run-length matrix (GLRLM) features, 5 Neighbourhood gray-tone difference matrix (NGTDM) features, and 14 Gray-level dependence matrix features (GLDM). In total, there are 9 + 11 × (18 + 24 + 16 + 16 + 5 + 14) = 1032 features.

[0039] (3) In the data preprocessing module, low-variance features (variance < 0.0001) are first removed, and all extracted radiomics features F are subjected to z-score standardization, with a standard deviation of 1 and a mean of 0. The standardized matrix is denoted as the standardized feature set F' = [f'1, f'2,... f'N], and the standardized data set is (F', e, t). m square1 square2 squarem waveletHHm

[0040] (4) In the sample variation stability feature screening module, a portion of the training set samples are randomly selected and repeated multiple times. Cox univariate screening is performed on the selected portion of the training set samples each time, and the frequency of selection of different variables is counted. When the frequency of selection is greater than a certain value (80%), it is considered to be a sample variation stability feature.

[0041] ​​​​​​(5) In the survival prediction feature screening module, according to the prognosis information of the patient and the screened sample change stability features, the 2-year survival condition is used as a label, and the least absolute selection and shrinkage operator algorithm is combined with a method similar to the sample change stability feature screening module. The method is repeated multiple times and the features with sample stability and high correlation with prognosis are retained to obtain the features related to the survival prediction task. After screening the features related to the survival prediction task, the statistical features (such as maximum value, minimum value, mean value, median) of the features related to the survival prediction task in the same patient's multiple ultrasound examination results (i.e. in multiple images) are calculated, and the cox multivariate screening is used for the statistical features. The statistical feature with the smallest p value is retained as the final selected feature.

[0042] Specifically, the sample change stability feature module and the survival prediction feature screening module are as shown in Figure 3 The training group data is randomly sampled 1000 times at the sample level, and the features with p value less than 0.05 are retained based on cox univariate screening, and the frequency of each feature being selected is counted. The features with a selection frequency greater than 80% (i.e. selected more than 800 times) are selected as sample change stability features, denoted as F1. Then, according to the prognosis information of the patient, for the OS task, it is whether the patient survives for 2 years, and for the RFS task, it is whether the patient relapses within 2 years. Combined with the LASSO algorithm, the training group data is randomly sampled 1000 times, and the features highly correlated with 2-year survival are selected using the LASSO algorithm. The features with a selection frequency greater than 80% are retained, denoted as F2. F1 and F2 are both selected from F' at the ultrasound image level. On the basis of F2, the statistical quantities of the image features in different ultrasound examinations of each patient are calculated, including maximum value, minimum value, mean value and median value. The cox multivariate screening is used for the four statistical quantities calculated for each feature in F2. The statistical quantity feature with the smallest p value is retained as the patient's feature, denoted as F3. Finally, the survival prediction model is constructed based on F3 and the clinical feature patient age.

[0043] (6) In the survival prediction model construction module, the survival prediction model uses the cox model. Independent feature screening and cox model are used to construct the survival prediction model in the training group data, and the model performance is evaluated in the internal validation group and the external validation group.

[0044] 3. Experimental results

[0045] To demonstrate the advantage of the method in the present application, first and only using a patient tumor area of the largest ultrasound image of the image set of the construction of the benchmark model for comparison, feature selection method similar to the last step of feature selection (ie F2 to F3 calculation and screening, also combined with clinical features patient age) comparison results as shown in table 1, can be seen in the present application than the C index of the benchmark model is better, indicating that in the survival prediction has a better effect. In addition, the different feature selection methods are compared, one is the method FAMUS_LASSO used in the present application, in addition to the combination of LASSO / Random forest (RF) / Gradient Boosting Decision Tree (GBDT) and Sequential Feature Selection (SFS) / Sequential Floating Forward Selection (SFFS) method, directly on the statistics of all image features (ie directly on the F' to calculate the statistics of 1032x4 = 4128 features, and combined with age characteristics) are calculated, comparison results as shown in Figure 4 (A, B) shown, the feature selection method proposed in the present application has better effect than the common feature selection method in the survival prediction of ovarian cancer. The cox model in the actual application has better generalization ability than other machine learning based survival prediction models. The several machine learning based survival prediction models compared include Random survival forest (abbreviated as RSF), conditional survival forest (abbreviated as CSF), extremely random survival forest (also known as extra survival trees, abbreviated as EST) and DeepSurv (abbreviated as DS). The results are shown in Figure 4 (C, D) shown, other algorithms are prone to overfitting, although better results can be obtained in the training set, but the effect is not good in internal and external validation. The prognosis analysis model used in the present application has better generalization effect than other machine learning based survival prediction models in the survival prediction of ovarian cancer.

[0046] The above embodiments are only the preferred embodiments of the present application, which are used to illustrate the technical solutions and specific details of the present application, and are not intended to limit the protection scope of the present application. Although the foregoing embodiments have been described in detail, the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced equivalently without departing from the spirit and scope of the claims and their equivalents. Therefore, these modifications or replacements are within the protection scope of the technical solutions.

[0047] Table 1. Multisample ultrasonic ovarian cancer survival prediction model in the present application and single sample using only the largest tumor area, expressed as mean and 95% confidence interval

[0048]

Claims

1. An ovarian cancer survival prediction system based on a plurality of ultrasound examination image-based statistical quantities, characterized in that, The method comprises the following modules: A data acquisition module is configured to acquire survival follow-up data of ovarian cancer patients, and transvaginal or transabdominal ultrasound images before radical surgery and comprehensive staging of the ovarian cancer patients, manually mark the ovarian cancer tumor regions in the ultrasound images, and obtain an original data set comprising the ultrasound images and the real follow-up data; A feature extraction module is configured to perform image transformation on the original ultrasound images in the original data set, and calculate radiomics features according to the manually marked ovarian cancer tumor regions before and after the image transformation, and obtain the prognosis information of the patients according to the follow-up data of the patients, to jointly construct an ultrasound survival analysis data set; A data preprocessing module is configured to remove low-variance radiomics features in the ultrasound survival analysis data set, and unify the standard deviation and mean value of the radiomics features, to obtain a standardized data set; A sample variation stability feature screening module is configured to perform multiple sampling on the standardized data set, and screen based on a cox univariate, to regard the features with a frequency of being selected greater than 80% as sample variation stability features; A survival prediction feature screening module is configured to further screen the sample variation stability features according to the task requirements to obtain features related to the survival prediction task, and calculate statistical features according to the features related to the survival prediction task in multiple images of the same patient, and screen the statistical features using a multivariate cox; A survival prediction model construction module is configured to construct an ovarian cancer ultrasound image survival prediction model based on the features selected by the survival prediction feature screening module. In the survival prediction feature screening module, the prognosis information of the patients and the sample variation stability features are used, the 2-year survival condition is taken as a label, a part of the sample variation stability feature samples are randomly extracted and repeated multiple times, the least absolute selection and shrinkage operator algorithm is used to screen the extracted samples each time, the frequency of being selected by different features is counted, the features related to the survival prediction task are regarded as features F2 when the frequency of being selected is greater than 80%, after the features related to the survival prediction task are screened out, multiple statistical features are calculated according to the features related to the survival prediction task in multiple images of the same patient, the cox multivariate is used to screen the multiple statistical features, and the statistical feature with the smallest p value is retained as the final selected feature F3; the statistical features include the maximum value max, the minimum value min, the mean value mean and the median value median. 2.The ovarian cancer survival prediction system based on multiple ultrasound examination image-based statistical quantities of radiomics features according to claim 1, characterized in that, In the data acquisition module, the ultrasound images do not contain Doppler blood flow signals; the manually marked ovarian cancer tumor regions are clearly visible in the ultrasound images and can completely display the lesions; the grayscale of the ultrasound images is standardized to [0, 255]. 3.The ovarian cancer survival prediction system based on multiple ultrasound examination image-based statistical quantities of radiomics features according to claim 1, characterized in that, The image transformation includes square value, square root value, exponential value, logarithm value, gradient value, local binary pattern (LBP) and four wavelet transforms (wavelet LL, wavelet LH, wavelet HL and wavelet HH) of the pixels. 4.The ovarian cancer survival prediction system based on multiple ultrasound examination image-based statistical quantities of radiomics features according to claim 1, characterized in that, The imageomics features include shape features, first-order statistical features and texture features. The texture features include gray level co-occurrence matrix features, gray level run-length matrix features, gray level size zone matrix features, neighborhood gray tone difference matrix features and gray level dependence matrix features. 5.The ovarian cancer survival prediction system based on multi-portion ultrasonic examination image group statistics constructed according to claim 1, characterized in that, The ultrasound survival analysis data set is represented as (F, e, t), wherein F is an ultrasound imageomics feature, e and t are prognosis information extracted from patient follow-up data, respectively. F = [f1, f2,... f m ,f square1 ,f square2 ,…f squarem ,f waveletHH1 ,...,f waveletHHm ] ∈ R n×p where R is the real number field, n is the sample size, p is the number of radiomics features of the original image and all image transformations, f represents the radiomics features extracted from the image after transformation, and the subscripts indicate that the features are extracted from the original image or the tumor region after image transformation. f1, f2, … f m respectively represent the features extracted from the original ultrasound image, f square1 square2 squarem respectively represent the radiomics features extracted from the square image transformation, f waveletHH1 waveletHHm represent the radiomics features extracted from the wavelet transform high-high component image transformation, where m represents the number of radiomics features extracted from the original image or single image transformation.​​​ e = [e1, e2, …, en] ∈ {0, 1} n ]∈{0,1} n e is the event label for n samples, e represents whether the sample is alive in the overall survival prediction task, and e represents whether the sample relapses in the relapse-free survival prediction task. t = [t1, t2,..., t n ] ∈ R n is the event time for n samples, in the overall survival prediction task, if the patient dies, t represents the time of sample survival, if the patient survives, t represents the follow-up time; in the recurrence-free survival prediction task, if the patient relapses, t represents the recurrence-free survival time of the sample, if the patient does not relapse, t represents the follow-up time. 6.The ovarian cancer survival prediction system based on multi-portion ultrasonic examination image group statistics constructed according to claim 1, characterized in that, In the data preprocessing module, low-variance imageomics features with a variance less than 0.0001 in the ultrasound survival analysis data set are first removed, and all extracted imageomics features F are subjected to z-score standardization, so that the standard deviation becomes 1 and the mean becomes 0, and the standardized feature set obtained is F ′ The standardized data set is jointly constituted with the prognosis information. 7.The ovarian cancer survival prediction system based on multi-portion ultrasonic examination image group statistics constructed according to claim 1, characterized in that, In the sample change stability feature screening module, a part of the standardized data set samples are randomly extracted and repeated multiple times, the extracted samples are subjected to cox univariate screening each time, and the frequency of being selected of different samples is counted; when the frequency of being selected is greater than 80%, the sample change stability feature F1 is considered. 8.The ovarian cancer survival prediction system based on multi-portion ultrasonic examination image group statistics constructed according to claim 1, characterized in that, In the survival prediction model construction module, the survival prediction model adopts a cox model.

Images