An ovarian cancer survival prediction method based on deep learning body component features

Abstract

The application discloses an ovarian cancer survival prediction method based on deep learning body component features, comprising the following steps: S1, collecting a plain scan CT image, and respectively segmenting a region of interest of a body component on a third lumbar vertebra layer CT image to extract deep learning features; S2, constructing a body component model and a body component prognosis scoring model, outputting a prognosis probability of a patient, and evaluating the prediction performance of the body component prognosis scoring model; S3, constructing a clinical model and a body component survival model, and establishing a joint model; and S4, evaluating the prediction performance of the clinical model, the body component survival model and the joint model. The application utilizes deep learning to automatically segment four body components on the third lumbar vertebra layer, including muscle tissue, visceral fat tissue, intermuscular fat tissue and subcutaneous fat tissue, and directly extracts deep features from the image for prognosis evaluation, thereby avoiding the low efficiency of traditional manual delineation, and solving the problem of non-uniform quantization indexes and grouping standards.

Description

Technical Field

[0001] This invention relates to the field of biotechnology, and in particular to a method for predicting ovarian cancer survival based on deep learning volumetric component features. Background Technology

[0002] Ovarian cancer is one of the deadliest types of gynecological malignancies. Due to its insidious early symptoms and rapid disease progression, more than two-thirds of patients are diagnosed at an advanced stage, with a 5-year survival rate of only about 30%. Cytoreductive surgery combined with platinum-based chemotherapy is currently the core of first-line treatment for ovarian cancer, and the application of targeted drugs has some value in improving patient prognosis. However, tumor heterogeneity and individual differences mean that ovarian cancer patients still face the risk of recurrence and death. Body composition is an important prognostic biomarker for ovarian cancer patients. In traditional assessment methods, body mass index (BMI) is widely used for body composition evaluation, but BMI cannot distinguish between different tissues such as fat and muscle, and patients with advanced ovarian cancer often have ascites, leading to significant predictive errors in BMI quantification. With the development of imaging technologies such as computed tomography (CT), body composition quantification based on patients' routine CT images has gradually become the mainstream approach. However, the body composition quantification indicators and grouping criteria used in different studies are not yet standardized, leading to discrepancies in the conclusions regarding the prognostic significance of body composition among studies.

[0003] The development of modern radiological image analysis techniques has provided new approaches to address the aforementioned issues of inconsistent body composition quantification and grouping standards. Radiomics methods have been used to extract body composition features and predict cancer patient prognosis. Deep learning, as a rapidly developing image analysis technology, can automatically learn image features and mine deeper information, showing great promise in medical image analysis. However, current applications of deep learning in body composition analysis are mostly limited to automatic segmentation; research on predicting ovarian cancer survival by directly extracting body composition features is still limited. Therefore, there is an urgent need for a standardized and automated body composition analysis method based on deep learning to objectively and reproducibly achieve risk stratification and survival prediction for ovarian cancer patients. Summary of the Invention

[0004] The purpose of this invention is to provide an ovarian cancer survival prediction method based on deep learning volume component features, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting ovarian cancer survival based on deep learning volume component features, comprising the following steps:

[0006] S1. Obtain ovarian cancer patient sample data, collect plain CT images, and segment the regions of interest of four body components on the CT image at the third lumbar vertebra level. The four body components include muscle tissue, visceral adipose tissue, intramuscular adipose tissue, and subcutaneous adipose tissue. Extract deep learning features of the four body components respectively.

[0007] S2. Based on the multi-mask Vision Transformer framework, four body composition models were constructed, and a network classifier fitted with tabular prior data was used as an ensemble framework to construct a body composition prognostic scoring model, output the prognostic probability of patients, and evaluate the predictive performance of the body composition prognostic scoring model.

[0008] S3. Based on Cox regression analysis, construct a clinical model and a body composition survival model respectively, and further combine the output results of the body composition prognostic scoring model with clinical characteristics to establish a joint model;

[0009] S4. Evaluate the predictive performance of the clinical model, body composition survival model, and combined model.

[0010] Preferably, the eligible selection criteria for ovarian cancer patients in step S1 are: surgically and pathologically confirmed malignant ovarian tumors, complete 5-year follow-up information, and abdominal CT scan within 3 months prior to standard ovarian cancer treatment.

[0011] Preferably, the ineligible selection criteria for ovarian cancer patients in step S1 are: coexisting tumors, poor image quality, or incomplete images.

[0012] Preferably, the nnU-Net self-configuration framework is used to perform automatic volume component segmentation on each CT image in step S1.

[0013] Preferably, the deep learning feature extraction of the four volume components in step S1 includes the following steps:

[0014] S101, based on the pre-segmented volume component region of interest mask, extract the corresponding volume component regions from the collected plain CT images respectively;

[0015] S102, Intensity normalization processing is performed on the extracted volume component regions;

[0016] S103, Perform logarithmic enhancement transformation on the normalized feature data;

[0017] S104 converts the single-channel grayscale image in the enhanced feature data into a three-channel image;

[0018] S105, the processed three-channel image is input into the Vision Transformer network, and deep learning features are extracted through forward propagation.

[0019] Preferably, the volume component prognostic scoring model is established by integrating a multi-mask Vision Transformer deep learning framework with a tabular prior data fitting network ensemble framework.

[0020] Preferably, the predictive performance of the body composition prognostic scoring model established in step S2 is evaluated using a confusion matrix, receiver operating characteristic (ROC) curve, decision curve, and calibration curve.

[0021] Preferably, the clinical features in step S3 are screened from the candidate clinical data through univariate and multivariate Cox regression analysis, and features with a p-value < 0.05 are selected.

[0022] Preferably, the performance of the clinical model, the body composition survival model, and the combined model in step S4 is evaluated using the C-index, AIC value, likelihood ratio test, and time-dependent receiver operating characteristic curve.

[0023] Preferably, after the clinical model, body composition survival model and combined model are evaluated in step S4, the feature data is input into any one of the clinical model, body composition survival model and combined model for training and then the risk score result is output. The feature data includes clinical data and body composition prognostic probability data.

[0024] The clinical data included age, tumor pathological type, tumor stage, chemotherapy regimen, surgical procedure, cytoreductive surgery, BMI, neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio, albumin, and cancer antigen 125.

[0025] The acquisition of the body component prognostic probability data first involves inputting deep learning features of four body component regions of interest at the third lumbar vertebra level into the body component prognostic scoring model, and then training the prognostic scoring model to output the body component prognostic probability.

[0026] The technical effects and advantages of this invention are as follows:

[0027] This invention utilizes deep learning to automatically segment four body components at the third lumbar vertebra level: muscle tissue, visceral adipose tissue, intramuscular adipose tissue, and subcutaneous adipose tissue. Deep features are directly extracted from the image for prognostic assessment, avoiding the inefficiency of traditional manual delineation and resolving issues related to inconsistent quantitative indicators and grouping standards. By combining clinical data, objective and standardized risk stratification and survival prediction are achieved. Verification shows that the combined model constructed in this invention exhibits superior predictive performance compared to individual clinical models and body component survival models. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the operation of the prediction method of the present invention.

[0029] Figure 2 This is a flowchart illustrating the deep learning feature extraction process of the present invention.

[0030] Figure 3 This is a flowchart illustrating the evaluation process of the prediction model of this invention.

[0031] Figure 4 This is a flowchart of the model construction and evaluation process of the present invention.

[0032] Figure 5 This is a schematic diagram of patient grouping according to the present invention.

[0033] Figure 6 The confusion matrix and ROC curve of the body component prognostic scoring model of this invention are shown.

[0034] Figure 7 The calibration curve and decision curve are for the body component prognostic scoring model of this invention.

[0035] Figure 8 This is a comparison of the time-dependent ROC curves of the clinical model, body component survival model, and combined model of this invention. Detailed Implementation

[0036] 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.

[0037] This invention provides, for example Figure 1 The method for predicting ovarian cancer survival based on deep learning volume component features, as shown, includes the following steps:

[0038] S1. Obtain ovarian cancer patient sample data, collect plain CT images, and segment the regions of interest of four body components on the CT images at the third lumbar vertebra level. The four body components include muscle tissue, visceral adipose tissue, intramuscular adipose tissue, and subcutaneous adipose tissue; extract the deep learning features of the four body components respectively.

[0039] The data from ovarian cancer patients included age, tumor pathological type, tumor stage, chemotherapy regimen, surgical procedure, surgical debulking status, BMI (body mass index), neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), albumin, and cancer antigen 125 (CA125); the preferred clinical data included age, tumor stage, and surgical debulking status.

[0040] Specifically, the eligibility criteria for ovarian cancer patients are: surgically and pathologically confirmed malignant ovarian tumors, complete 5-year follow-up information, and abdominal CT scan within 3 months prior to standard ovarian cancer treatment; the ineligible criteria for ovarian cancer patients in step S1 are: coexisting tumors, poor image quality, or incomplete images; and the nnU-Net self-configured framework is used to automatically segment volume components for each CT image in step S1.

[0041] Further reference Figure 2 As shown, the deep learning feature extraction of the four volume components in step S1 includes the following steps:

[0042] S101, based on the pre-segmented volume component region of interest mask, extract the corresponding volume component regions from the collected plain CT images respectively;

[0043] S102, Intensity normalization processing is performed on the extracted volume component regions;

[0044] S103, Perform logarithmic enhancement transformation on the normalized feature data;

[0045] S104 converts the single-channel grayscale image in the enhanced feature data into a three-channel image;

[0046] S105, the processed three-channel image is input into the Vision Transformer network, and deep learning features are extracted through forward propagation.

[0047] S2. Based on the multi-mask Vision Transformer framework, four body composition models are constructed. A body composition prognostic scoring model is built using a table-based prior data fitting network classifier as an ensemble framework. The model outputs the patient's prognostic probability and evaluates its predictive performance. The body composition prognostic scoring model is established by integrating the multi-mask Vision Transformer deep learning framework with the table-based prior data fitting network ensemble framework. Specifically, the predictive performance of the body composition prognostic scoring model established in step S2 is evaluated using the confusion matrix, receiver operating characteristic (ROC) curve, decision curve, and calibration curve.

[0048] S3. Based on Cox regression analysis, construct a clinical model and a body composition survival model respectively, and further combine the output results of the body composition prognostic scoring model with clinical features to establish a joint model; the clinical features in step S3 are screened from candidate clinical data through univariate and multivariate Cox regression analysis, and features with P value < 0.05 are selected.

[0049] S4. Evaluate the predictive performance of the clinical model, body composition survival model, and combined model. Use the C-index, AIC value, likelihood ratio test, and time-dependent receiver operating characteristic curve to evaluate the performance of the clinical model, body composition survival model, and combined model in step S4.

[0050] Specifically, after the clinical model, body composition survival model and combined model are evaluated in step S4, the feature data is input into any one of the clinical model, body composition survival model and combined model for training and the risk score result is output. The feature data includes clinical data and body composition prognostic probability data.

[0051] Clinical data include age, tumor pathological type, tumor stage, chemotherapy regimen, surgical procedure, surgical cytoreductive effect, BMI, neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio, albumin, and cancer antigen 125.

[0052] To obtain prognostic probability data for body components, the deep learning features of four body component regions of interest at the third lumbar vertebra level are first input into the body component prognostic scoring model. The prognostic scoring model is then trained to output the prognostic probability of body components.

[0053] After training and screening, the joint model showed the best prediction performance. Therefore, this application employs a joint model as the optimal approach for ovarian cancer survival prediction based on deep learning body composition features. Specifically, the input data for the joint model includes age, tumor stage, surgical debulking status, and body composition prognostic probability. In detail, the data, including age, tumor stage, surgical debulking status, and body composition prognostic probability, are input into the joint model to ultimately obtain a risk score for each patient. Figure 3 and Figure 4 As shown, the verification method for the prediction model of this application includes the following steps:

[0054] Step 1: 157 patients who underwent ovarian cancer surgery at two centers were selected and divided into three groups: training, validation, and testing.

[0055] Specifically, such as Figure 5 As shown, 20% of the 137 patients in Center 1 were randomly selected for the internal testing group, and the remaining 80% were randomly assigned to the training and validation groups at an 8:2 ratio; 19 patients from Center 2 formed the external testing group. The internal and external testing groups were then combined into a joint testing group. All patients included in the study were screened according to inclusion and exclusion criteria to improve the accuracy of the experiment.

[0056] Inclusion criteria included: a1: malignant ovarian tumor confirmed by surgical pathology; a2: complete 5-year follow-up information; a3: abdominal CT scan performed within 3 months prior to standard ovarian cancer treatment. Exclusion criteria included: b1: coexisting tumors; b2: poor image quality or incomplete images. A total of 157 ovarian cancer patients from the two centers were ultimately included after screening.

[0057] It should be noted that the retrospective study was approved by the institution's ethics review committee. Because this study was retrospective and all participants were anonymous, the ethics review committee waived the need for informed consent from the patients.

[0058] Step 2: Acquire plain CT images, process the images and extract deep learning volume component features, and establish a volume component prognostic scoring model based on the deep learning features;

[0059] Specifically, all operations in this application followed the standards of the Imaging Biomarker Standardization Initiative (IBSI). Image acquisition for all imaging data was performed using spiral CT scanners from two manufacturers: Scanner A (Newsford) and Scanner B (United Imaging), with the scan range covering the third lumbar vertebra (L3). Specific acquisition parameters were as follows: Scanner A used a tube voltage of 120 kV, automatic tube current adjustment, a rotation time of 0.6 seconds, a pitch of 0.8, and images were reconstructed using an F20 filter with a slice thickness and interslice spacing of 5 mm; Scanner B used the same tube voltage of 120 kV, automatic tube current adjustment, a rotation time of 0.5 seconds, a pitch of 0.9875, and images were also reconstructed using a 5 mm slice thickness and interslice spacing with an F20 filter.

[0060] The nnU-Net self-configurable framework was used to automate the segmentation of L3 layer volume components in abdominal plain CT images of ovarian cancer patients. Given the limited size of the fine-annotated dataset (N=30), a transfer learning strategy was introduced within the nnU-Net framework: network weights were pre-loaded with parameters trained on a large-scale public dataset to provide the encoder with prior knowledge of common anatomical structures and CT texture patterns, reducing dependence on the size of the target data. During training, end-to-end full parameter fine-tuning was performed on 30 target data cases using a stochastic gradient descent optimizer with Nesterov momentum and a weight decay of 3×10⁻⁶. -5 The loss function is a weighted sum of the Dice loss and the cross-entropy loss, and the initial learning rate is set to 1×10⁻⁶. -2The model is then adjusted using polynomial attenuation. During the inference phase, the fine-tuned model employs a sliding window strategy to predict full-volume CT scans. A high-precision final segmentation mask is generated by aggregating Gaussian-weighted probability maps within overlapping regions, outputting segmentation results for four body components: muscle tissue, visceral adipose tissue, intramuscular adipose tissue, and subcutaneous adipose tissue. After automatic image segmentation, the Region of Interest (ROI) is adjusted by a radiologist with 5 years of experience and a radiologist with 10 years of experience (neither radiologist knew the patient's information).

[0061] A volume component prognostic scoring model was constructed by integrating the multi-mask Vision Transformer (ViT) deep learning framework with the TabPFN ensemble framework. First, ViT base models were constructed for four volume components: based on pre-segmented masks of mixed regions of skeletal muscle, subcutaneous fat, visceral fat, and intramuscular fat, CT images were subjected to mask extraction, intensity normalization, logarithmic enhancement, and three-channel transformation, and then input into independent ViT networks. A two-stage training strategy was adopted, in which the backbone was frozen in the first stage and only the classification head was trained; in the second stage, all parameters were unfrozen for end-to-end fine-tuning. Focus loss function, CutMix data augmentation, and exponential moving average techniques were applied to generate the prediction probabilities of the four base models. Subsequently, the TabPFN framework was used to achieve multi-model fusion: the probabilities output by the four basic models on the training group were used as features to construct a training feature matrix, which was then input into the pre-trained TabPFN classifier to obtain the predicted probabilities of the validation group; an adaptive threshold optimization strategy was used on the validation group to find the optimal classification threshold to balance sensitivity and specificity; finally, the optimized threshold was applied to the test group to output the final comprehensive component prognostic score, which was used for prognostic risk assessment of ovarian cancer patients.

[0062] To verify the predictive ability of this prognostic scoring model for body components, this application conducted a preliminary performance evaluation. Table 1 below shows the performance metrics of the prognostic scoring model. In the combined test group, the model's accuracy was 76.6% (95% CI: 63.8%–89.4%). True positive, true negative, false positive, and false negative values ​​are shown in Table 1. Figure 6 a; The area under the receiver operating characteristic (AUC) was 0.808 (95% CI: 0.650–0.938), and the ROC curve is shown below. Figure 6 b; Calibration curve is shown in... Figure 7 a) This shows that the model's predictions are in good agreement with the actual observations; the decision curve is shown below. Figure 7 b indicates that the model has guiding value for clinical decision-making in most threshold ranges. These results demonstrate that the body component prognostic scoring model possesses good predictive performance. Subsequent steps will involve further integrating clinical characteristics to construct a survival model and conducting comparative evaluations.

[0063] Table 1. Performance of the body composition prognostic scoring model in the joint test group.

[0064]

[0065] In Table 1, AUC represents the area under the curve, and CI represents the confidence interval.

[0066] Step 3: Collect candidate clinical features and select features with a P value < 0.05 to build a clinical model.

[0067] Specifically, 11 candidate clinical data were included to build the clinical model, including age, tumor stage, pathological type, chemotherapy regimen, surgical regimen, surgical debulking status, BMI, NLR, PLR, albumin, and CA125. All clinical data were obtained from the hospital's electronic medical records.

[0068] This study included 156 patients, of whom 50 experienced positive events during the 5-year follow-up period, with a mortality rate of 32.1%. Detailed patient baseline characteristics are shown in Table 2 below. There were no statistically significant differences in baseline characteristics among the training, validation, and combined testing groups (P > 0.05).

[0069] Table 2. Clinicopathological characteristics of ovarian cancer patients in the training group, validation group, and combined test group.

[0070]

[0071] In Table 2, BMI represents body mass index, NLR represents the neutrophil-to-lymphocyte ratio, and PLR represents the platelet-to-lymphocyte ratio. CA125 was logarithmically transformed for analysis; the transformed data are expressed as mean ± standard deviation.

[0072] In the training group, univariate and multivariate Cox regression analysis were used to screen clinical features, and features with P values ​​<0.05 (i.e., age, tumor stage, and surgical debulking status) were selected to establish a clinical model.

[0073] All statistical tests in this application are two-tailed tests, with a significance level of α=0.05. Analysis of variance and χ² were used. 2 Use the Mann-Whitney U test (select according to data type) to compare differences in clinical characteristics among the training group, validation group, and combined test group.

[0074] Step four: Construct clinical models, body composition survival models, and combined models based on clinical characteristics and prognostic probabilities of body composition.

[0075] Specifically, the combined model includes age, tumor stage, surgical debulking status, and prognostic probability of body composition.

[0076] Time-dependent receiver operating characteristic (ROC) curves and their associated area under the curve (AUC) were used to quantify the performance of three survival models (body composition survival model, clinical model, and combined model), i.e., how to use these models to predict the risk of patient death at years 1, 2, 3, and 4. Model performance was assessed as follows: Figure 8 As shown, the combined model demonstrated the best predictive ability for the final outcome at multiple time points. Measured by AUC, the predictive performance of this model at years 1, 2, 3, and 4 was 0.733, 0.776, 0.732, and 0.771, respectively. Although the combined model did not show an advantage over the clinical model in year 1, the likelihood ratio test indicated that the combined model's fit was significantly better than that of the clinical model (χ²). 2 =18.30, P value <0.001; however, compared with the body component survival model, the goodness of fit of the joint model was not significantly different. 2 =6.66, P=0.155.

[0077] The C-index and AIC values ​​of each survival model are shown in Table 3 below. The C-indexes of the body composition survival model and the clinical model were 0.745 (95% CI: 0.606-0.872) and 0.661 (95% CI: 0.530-0.787), respectively, while the C-index of the combined model reached 0.767 (95% CI: 0.640-0.888). Furthermore, the body composition survival model had the lowest AIC value among all models.

[0078] Table 3 Performance Comparison of Three Survival Models

[0079]

[0080] In Table 3, C-index is the consistency index, CI is the confidence interval; AIC is the Akaike Information Criterion; LRχ 2 LRχ represents the likelihood ratio chi-square value (degrees of freedom are in parentheses). 2 The values ​​and P-values ​​are obtained by comparing the clinical model and the combined model; or LRχ². 2 The values ​​and P-values ​​are obtained by comparing the body component survival model and the joint model.

[0081] Following steps one through four of the analysis, an ovarian cancer survival prediction model was established and validated using deep learning body component features and key clinical variables. The innovation of this study lies in directly applying deep learning body component features to ovarian cancer survival prediction, overcoming the challenge of inconsistent body component quantification indicators and grouping standards in traditional methods. Analysis shows that clinical models, body component survival models, and combined models can all predict ovarian cancer survival, but the combined model has greater advantages.

[0082] In summary, the combined model integrating clinical and deep learning body composition features demonstrates good survival prediction performance for ovarian cancer. By integrating multi-dimensional body composition information, it avoids the subjectivity of traditional indicator selection and grouping criteria, providing gynecologists with a new approach to early identification of patients with poor prognosis and timely intervention.

[0083] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting ovarian cancer survival based on deep learning volume component features, characterized in that, The following steps are included: S1. Obtain ovarian cancer patient sample data, collect plain CT images, and segment the regions of interest of four body components on the CT image at the third lumbar vertebra level. The four body components include muscle tissue, visceral adipose tissue, intramuscular adipose tissue, and subcutaneous adipose tissue. Deep learning features were extracted from the four volume components respectively; S2. Based on the multi-mask Vision Transformer framework, four body composition models were constructed, and a network classifier fitted with tabular prior data was used as an ensemble framework to construct a body composition prognostic scoring model, output the prognostic probability of patients, and evaluate the predictive performance of the body composition prognostic scoring model. S3. Based on Cox regression analysis, construct a clinical model and a body composition survival model respectively, and further combine the output results of the body composition prognostic scoring model with clinical characteristics to establish a joint model; S4. Evaluate the predictive performance of the clinical model, body composition survival model, and combined model.

2. The ovarian cancer survival prediction method based on deep learning volume component features according to claim 1, characterized in that, The eligibility criteria for ovarian cancer patients in step S1 are: surgically and pathologically confirmed malignant ovarian tumors, complete 5-year follow-up information, and abdominal CT scan within 3 months prior to standard ovarian cancer treatment.

3. The ovarian cancer survival prediction method based on deep learning volume component features according to claim 2, characterized in that, The ineligible selection criteria for ovarian cancer patients in step S1 are: coexisting tumors, poor image quality, or incomplete images.

4. The ovarian cancer survival prediction method based on deep learning volume component features according to claim 1, characterized in that, The nnU-Net self-configuration framework is used to automatically segment each CT image in step S1.

5. The ovarian cancer survival prediction method based on deep learning volume component features according to claim 1, characterized in that, The deep learning feature extraction of the four volume components in step S1 includes the following steps: S101, based on the pre-segmented volume component region of interest mask, extract the corresponding volume component regions from the collected plain CT images respectively; S102, Intensity normalization processing is performed on the extracted volume component regions; S103, Perform logarithmic enhancement transformation on the normalized feature data; S104 converts the single-channel grayscale image in the enhanced feature data into a three-channel image; S105, the processed three-channel image is input into the Vision Transformer network, and deep learning features are extracted through forward propagation.

6. The ovarian cancer survival prediction method based on deep learning volume component features according to claim 1, characterized in that, The volume component prognostic scoring model is established by integrating the multi-mask Vision Transformer deep learning framework with the tabular prior data fitting network ensemble framework.

7. The ovarian cancer survival prediction method based on deep learning volume component features according to claim 1, characterized in that, The predictive performance of the body composition prognostic scoring model established in step S2 was evaluated using the confusion matrix, receiver operating characteristic (ROC) curve, decision curve, and calibration curve.

8. The ovarian cancer survival prediction method based on deep learning volume component features according to claim 1, characterized in that, In step S3, clinical features are selected from candidate clinical data through univariate and multivariate Cox regression analysis, and features with a p-value < 0.05 are chosen.

9. The ovarian cancer survival prediction method based on deep learning volume component features according to claim 1, characterized in that, The performance of the clinical model, the body composition survival model, and the combined model described in step S4 was evaluated using the C-index, AIC value, likelihood ratio test, and time-dependent receiver operating characteristic curve.

10. The ovarian cancer survival prediction method based on deep learning volume component features according to claim 1, characterized in that, After the clinical model, body composition survival model and combined model are evaluated in step S4, the feature data is input into any one of the clinical model, body composition survival model and combined model for training and the risk score result is output. The feature data includes clinical data and body composition prognostic probability data. The clinical data included age, tumor pathological type, tumor stage, chemotherapy regimen, surgical procedure, cytoreductive surgery, BMI, neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio, albumin, and cancer antigen 125. The acquisition of the body component prognostic probability data first involves inputting deep learning features of four body component regions of interest at the third lumbar vertebra level into the body component prognostic scoring model, and then training the prognostic scoring model to output the body component prognostic probability.

Images