ABSTRACT
Background
Endometrial cancer is a significant gynecological malignancy with rising incidence among women of reproductive age, necessitating effective fertility‐sparing treatments. Machine learning offers potential in enhancing prognostic assessments through radiomics.
Objective
To develop and validate a machine learning‐based model integrating radiomics and clinical features to predict the prognosis of fertility‐sparing treatments in endometrial cancer patients.
Methods
This retrospective study included 102 endometrial cancer patients who received fertility‐sparing treatment at West China Second University Hospital from November 2017 to December 2023. Patients were randomly divided into training (n = 81) and testing (n = 21) cohorts. The primary outcome was 6 month treatment response evaluated by endometrial sampling. MRI‐based radiomic features and clinical data were analyzed using logistic regression and other machine learning algorithms. A combined model was constructed by incorporating radiomics and clinical features. Model performances were assessed by the area under the receiver operating characteristic curve (AUC) analysis, and decision curve analysis (DCA) was used to estimate the models' clinical values.
Results
Among 102 patients, 60 (58.8%) achieved complete remission (CR), whereas 42 (41.2%) did not (non‐CR). In total, 1197 radiomic features were extracted from MRI images. Thirteen radiomic features were deemed valuable by dimensionality reduction and selection. Among radiomic models, the logistic regression model was the most effective, showing high stability and accuracy, with AUCs of 0.875 in the training cohort and 0.852 in the test cohort. Ultimate clinical features selected by univariate and multivariable logistic regression constructed a clinical model. The combined model demonstrated superior performance with an AUC of 0.941 in the training cohort and 0.907 in the testing cohort. Combined model DCA revealed optimal clinical efficacy.
Conclusion
The integrated model effectively predicts fertility preservation outcomes, offering a reliable noninvasive tool for clinicians. This approach may enhance personalized treatment strategies for young endometrial cancer patients.
Keywords: endometrial cancer, fertility‐sparing treatment, machine learning, magnetic resonance imaging, prognostic model, radiomics
1. Introduction
Endometrial cancer (EC) is a prevalent gynecological malignancy, ranking sixth among all cancers globally [1]. In recent years, the incidence of EC in women of reproductive age has steadily increased, with approximately 7% of cases occurring in women under the age of 45 [2, 3].
In recent years, an increasing number of studies have validated the safety and efficacy of high‐dose progestin conservative treatment for patients with early‐stage endometrioid endometrial carcinoma [4, 5, 6, 7, 8, 9, 10]. The primary objective of conservative treatment is to achieve complete remission (CR); however, the factors influencing response rates remain inadequately defined. These factors include the molecular characteristics of the disease, patient weight (with improved response rates observed in those with a body mass index (BMI) < 25 kg/m2), histological grade, and polycystic ovary syndrome (PCOS), among others, though evidence supporting their clinical utility remains limited [11, 12, 13, 14, 15]. While guidelines recommend that the absence of myometrial invasion be confirmed by magnetic resonance imaging (MRI) before opting for a fertility‐sparing approach [16], retrospective case series suggest that patients with endometrial cancer and myometrial invasion may also be candidates for fertility‐sparing treatment [13, 17]. Thus, accurately identifying patients suitable for fertility‐sparing treatment is of paramount importance.
In recent years, machine learning technology has advanced rapidly, particularly with significant breakthroughs in image processing and analysis, demonstrating considerable advantages in disease prediction and prognosis. It has shown substantial potential in detecting brain hemorrhages on cranial CT scans [18], diagnosing skin cancer [19], and identifying diabetic retinopathy [20]. Radiomics, a subfield of machine learning, is an emerging and promising discipline that quantitatively characterizes tumors by extracting and analyzing numerous image features using high‐throughput techniques [21]. In the past few years, a growing body of research has explored the application of radiomics in endometrial cancer [22, 23, 24]. These studies suggest that machine learning can assist in addressing clinical challenges and optimizing treatment for endometrial cancer patients. However, to our knowledge, no studies have yet examined the application of different machine learning models in predicting the prognosis of fertility‐sparing treatments for endometrial cancer patients.
In this study, we employed various machine learning approaches to develop and validate an MRI‐based combined model integrating radiomics and clinical features to predict the prognosis (remission status) of endometrial cancer patients undergoing fertility‐sparing treatment after 6 months.
2. Materials and Methods
2.1. Study Design and Study Population
This retrospective study included patients who received fertility‐sparing treatment at West China Second University Hospital between November 2017 and December 2023. The study was reviewed and approved by the institutional review board of the West China Second Hospital of Sichuan University (IRB No. 2023[024]). Written informed consent was obtained from all the participants prior to the enrollment of this study. Medical records were utilized to collect data on each patient's clinical characteristics and oncological outcomes.
The inclusion criteria were as follows: (1) age ≤ 45 years; (2) a strong desire for fertility preservation; (3) histological diagnosis of endometrioid adenocarcinoma (grade 1 or grade 2) confirmed by designated gynecological pathologists; (4) positive progesterone receptor status in immunohistochemistry; (5) underwent enhanced abdomen‐pelvic MRI; (6) patient response evaluated through endometrial sampling via hysteroscopy 6 months posttreatment.
The exclusion criteria included: (1) the presence of malignancies at other sites or Lynch syndrome; (2) incomplete clinical data; (3) MRI quality that did not meet the analysis requirements.
All patients received fertility‐sparing therapy based on progestin regimens. The treatment options included: (1) oral progestin therapy with medroxyprogesterone acetate (MPA, 250–500 mg/day) or megestrol acetate (MA, 160–480 mg/day); (2) levonorgestrel‐releasing intrauterine system (LNG‐IUS, Mirena 52 mg); (3) gonadotropin‐releasing hormone agonist (GnRH‐a, 3.75 mg every 4 weeks) combined with the above regimens. Treatment continued for 6 months, and endometrial biopsy via hysteroscopy was performed to evaluate response.
Patients were randomly assigned to a training cohort and a test cohort in an 8:2 ratio (see Figure 1 for details).
FIGURE 1.
Flowchart for selecting the study population.
2.2. Clinical Data Analysis
The patients' imaging and clinical data were obtained from routine clinical records and the Picture Archiving and Communication System (PACS) at our hospital. We conducted a retrospective analysis of various clinical parameters, including age, body mass index (BMI), parity, and comorbidities such as polycystic ovary syndrome (PCOS), thyroid disease, diabetes, hypertension, and insulin resistance (IR), as well as menstrual status. Laboratory data were collected and analyzed for hemoglobin (Hb), platelet count (PLT), neutrophil count (NEUT), lymphocyte count (LYMPH), monocyte count (MONO), tumor markers (including carbohydrate antigen 125 (CA125), carbohydrate antigen 199 (CA199), alpha‐fetoprotein (AFP), and HE4), and liver and kidney function indicators, including albumin (ALB), alanine aminotransferase (ALT), aspartate aminotransferase (AST), total bilirubin (TB), blood urea (URE), and serum creatinine (CRE). Pathological immunohistochemical indicators included p53 status, Ki‐67, myometrial invasion, and histological grade; imaging characteristics included the maximum tumor diameter. The primary outcome is the patient's response, assessed through endometrial sampling performed via hysteroscopy, 6 months after treatment.
2.3. Image Acquisition and Image Segmentation
All MR examinations were conducted using 1.5 T and 3.0 T scanners (GE Signa HDXt, Siemens Prisma, and Siemens Aera) equipped with eight‐channel phased‐array abdominal coils. Each patient underwent MR scanning according to a standardized protocol. In this study, uterus‐axial T2‐weighted images (T2WI) were acquired for image segmentation.
All images were saved using standard soft tissue settings (window width, 1200 HU; window level, 600 HU) and stored in Digital Imaging and Communications in Medicine (DICOM) format. Two radiologists, with 10 and 20 years of diagnostic experience, respectively, manually segmented the regions of interest (ROIs) using ITK‐SNAP software (version 3.8.0, http://www.itksnap.org/). On axial uterine T2WI, ROIs were manually contoured layer by layer to encompass the primary tumor, intratumoral components (hemorrhage, necrosis, cystic change), and immediately adjacent invaded myometrium. A complete schematic is provided in Figure 2.
FIGURE 2.
Workflow of the study.
Images and data were preprocessed through resampling and standardization to ensure the repeatability of the results. Intra‐ and interobserver reproducibility was assessed using the intraclass correlation coefficient (ICC). Thirty‐five patients (21 complete remission and 14 noncomplete remission) were randomly selected, and radiologist 1 re‐segmented the ROIs after 2 months. An ICC greater than 0.8 indicated good consistency.
2.4. Feature Extraction and Selection
Before feature extraction, to mitigate imaging variability between different MRI scanners, images are normalized to ensure that all gray‐level values fall within the range of 0 to 600. The normalized MR images are subsequently processed using various built‐in filters, including gradient, exponential, logarithmic, square, square root, wavelet, and Laplacian of Gaussian (LOG) filters, to generate derived images.
Radiomic features were extracted from radiologist‐defined ROIs using the open‐source Python package Pyradiomics (https://pypi.org/project/pyradiomics/). The extracted features were categorized into three groups: (1) geometry, (2) intensity, and (3) texture. Geometry features describe the three‐dimensional shape characteristics of the tumor. Intensity features represent the first‐order statistical distribution of voxel intensities within the tumor. Texture features capture patterns or the second‐ and higher‐order spatial distributions of intensities. Texture features were extracted using several methods, including gray‐level co‐occurrence matrix (GLCM), gray‐level dependence matrix (GLDM), gray‐level run length matrix (GLRLM), gray‐level size zone matrix (GLSZM), and neighborhood gray‐tone difference matrix (NGTDM) [25]. All features were standardized using Z‐scores.
Dimensionality reduction and feature selection methods for the radiomic features in the training cohort were conducted as follows: First, the 1197 extracted features underwent analysis of variance (ANOVA), with features exhibiting an ICC score greater than 0.8 and statistical significance being selected (628 features remain). Second, for features with high reproducibility, the Spearman rank correlation coefficient was used to assess the correlation between features, retaining only one of any two features with a correlation coefficient greater than 0.9 (176 features remain). Next, the least absolute shrinkage and selection operator (LASSO) regression model was employed to select the most representative features, followed by 10‐fold cross‐validation [26] (13 features remain) (Figure 3). All feature selection processes were performed on the training cohort and subsequently applied to the testing cohort.
FIGURE 3.
Radiomic feature selection based on LASSO regression model. (a, b) Ten‐fold cross‐validated coefficients and 10‐fold cross‐validated MSE. (c) Bar chart of feature weight for the LASSO regression model.
2.5. Development of Radiomics Signature, Clinical and Combined Models
To construct the clinical model, univariate logistic regression analysis was initially performed on each clinical predictor variable. Statistically significant features were then subjected to multivariate logistic regression analysis to identify the final predictor variables for model development. The odds ratios (ORs) and 95% confidence intervals (CIs) for each factor were calculated.
In our study, six widely used machine learning algorithms were employed to develop the radiomics model, including logistic regression (LR), support vector machine (SVM), K‐nearest neighbors (KNN), random forest (RF), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM). The predictive performance of these models was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The best‐performing radiomics models were then selected to construct a combined model incorporating both radiomics and clinical features [27].
The model's performance was assessed and validated using both the training and testing cohorts, focusing on metrics such as AUC, accuracy, sensitivity, specificity, PPV, and NPV. Decision curve analysis (DCA) was used to quantify the net benefits across different threshold probabilities in both cohorts, thereby evaluating the model's clinical utility.
2.6. Statistical Analysis
Statistical analyses were conducted using SPSS software (version 26.0, IBM), R software (version 4.1.2; https://www.r‐project.org), and Python software (version 3.9; http://www.python.org). The Shapiro–Wilk test was used to evaluate whether the continuous variables were normally distributed. A Student's t‐test was used to analyze continuous variables conforming to a normal distribution. The Mann–Whitney U test was used to analyze the continuous variables with non‐normal distributions. Continuous variables were expressed as mean ± standard deviation (x ± s). Categorical variables were analyzed using the chi‐square test or Fisher's exact test and expressed as ratios. The forward stepwise selection method was applied in multivariable logistic regression analysis using the Akaike information criterion (AIC) as the selection criterion. A two‐tailed p < 0.05 was considered statistically significant.
3. Results
3.1. Patients' Population and Clinical Data
A total of 102 patients were enrolled in our study, with 81 in the training cohort and 21 in the testing cohort. The results revealed no significant differences between the training and testing groups, with the clinical characteristics of the patients summarized in Table 1. The primary outcome was 6 month complete remission (CR) vs. non‐CR on hysteroscopic biopsy. In total, CR occurred in 60/102 (58.8%) patients (training: 51/81, 63.0%; testing: 9/21, 42.9%). Among the training cohort, significant differences were identified between CR and non‐CR patients in HE4 levels, maximum tumor diameter, total bilirubin, PCOS, insulin resistance, Ki‐67 levels, and myometrial invasion. In the testing cohort, significant differences were found in PCOS, maximum tumor diameter, AFP, ALB, and HE4 levels. Univariate and multivariate logistic regression analyses demonstrated that HE4 levels, myometrial invasion, PCOS, insulin resistance, and maximum tumor diameter were independent predictors of successful fertility preservation (Table 2). A clinical model was developed based on these five factors. Additionally, patients who did not achieve complete remission after 6 months of treatment were more likely to present with PCOS (OR, 0.469; 95% CI, 0.286–0.770), insulin resistance (OR, 0.540; 95% CI, 0.329–0.886), myometrial invasion (OR, 0.506; 95% CI, 0.299–0.855), larger tumor diameters (OR, 0.194; 95% CI, 0.062–0.608), and elevated HE4 levels (OR, 0.431; 95% CI, 0.232–0.799).
TABLE 1.
Clinical characteristics in the training and test cohorts.
| Variable | Training cohort (n = 81) | Testing cohort (n = 21) | |||||
|---|---|---|---|---|---|---|---|
| Non‐CR (n = 30) | CR (n = 51) | p | Non‐CR (n = 12) | CR (n = 9) | p | ||
| Age (years) | 29.83 ± 4.43 | 30.59 ± 4.55 | 0.468 | 29.17 ± 4.61 | 30.44 ± 3.84 | 0.508 | |
| BMI (kg/m2) | 26.35 ± 6.24 | 25.03 ± 5.16 | 0.457 | 23.96 ± 5.31 | 28.09 ± 4.43 | 0.074 | |
| HB (g/L) | 119.17 ± 14.87 | 121.31 ± 21.64 | 0.357 | 124.58 ± 26.88 | 137.44 ± 10.11 | 0.433 | |
| PLT (109/L) | 283.30 ± 81.70 | 281.18 ± 85.38 | 0.912 | 288.42 ± 66.01 | 279.33 ± 47.64 | 0.730 | |
| NEUT (109/L) | 5.11 ± 2.41 | 4.21 ± 1.44 | 0.160 | 3.88 ± 0.65 | 5.35 ± 2.17 | 0.069 | |
| LYMPH (109/L) | 1.94 ± 0.65 | 2.02 ± 0.81 | 0.925 | 2.00 ± 0.58 | 2.56 ± 0.81 | 0.109 | |
| MONO (109/L) | 0.45 ± 0.17 | 0.37 ± 0.14 | 0.072 | 0.39 ± 0.13 | 0.47 ± 0.11 | 0.153 | |
| CA125 (U/ml) | 15.99 ± 11.62 | 13.78 ± 11.96 | 0.163 | 12.57 ± 5.13 | 15.48 ± 11.97 | 0.943 | |
| CA199 (U/ml) | 16.24 ± 10.74 | 17.18 ± 13.20 | 0.814 | 13.75 ± 6.87 | 14.83 ± 10.60 | 0.779 | |
| AFP (ng/ml) | 2.29 ± 1.29 | 2.82 ± 1.92 | 0.313 | 2.95 ± 1.57 | 1.96 ± 1.17 | 0.032* | |
| HE4 (pmol/L) | 43.63 ± 14.17 | 30.92 ± 9.68 | < 0.001* | 42.41 ± 13.77 | 28.06 ± 7.28 | 0.017* | |
| ALB (g/L) | 46.59 ± 5.37 | 46.86 ± 3.00 | 0.454 | 48.75 ± 1.90 | 46.60 ± 2.24 | 0.027* | |
| ALT (U/L) | 31.43 ± 21.35 | 28.43 ± 21.75 | 0.389 | 32.00 ± 23.31 | 34.33 ± 33.61 | 0.914 | |
| AST (U/L) | 25.80 ± 11.77 | 25.45 ± 11.64 | 0.720 | 26.17 ± 10.65 | 28.78 ± 20.31 | 1 | |
| TB (umol/L) | 10.68 ± 3.92 | 12.95 ± 4.72 | 0.015* | 13.89 ± 4.26 | 10.46 ± 3.99 | 0.075 | |
| URE (mmol/L) | 4.17 ± 1.36 | 4.35 ± 0.96 | 0.496 | 4.19 ± 1.14 | 5.00 ± 1.46 | 0.168 | |
| CRE (umol/L) | 53.07 ± 8.82 | 55.04 ± 8.29 | 0.315 | 52.67 ± 9.99 | 52.33 ± 6.87 | 0.932 | |
| Max‐diameter (cm) | 2.55 ± 1.58 | 1.14 ± 0.82 | < 0.001* | 2.52 ± 1.37 | 0.96 ± 0.59 | 0.007* | |
| Parity | 0 | 28 (93.33%) | 48 (94.12%) | 1 | 12 (100.00) | 9 (100.00) | 1 |
| ≥ 1 | 2 (6.67%) | 3 (5.88%) | Null | Null | |||
| PCOS | With | 11 (36.67%) | 5 (9.80%) | 0.008* | 6 (50.00%) | Null | 0.043* |
| Without | 19 (63.33%) | 46 (90.20%) | 6 (50.00%) | 9 (100.00%) | |||
| Thyroid disease | With | 1 (3.33%) | 1 (1.96%) | 1 | 1 (8.33%) | Null | 1 |
| Without | 29 (96.67%) | 50 (98.04%) | 11 (91.67%) | 9 (100.00%) | |||
| Diabetes | With | 4 (13.33%) | 4 (7.84%) | 0.678 | 1 (8.33%) | 3 (33.33%) | 0.377 |
| Without | 26 (86.67%) | 47 (92.16%) | 11 (91.67%) | 6 (66.67%) | |||
| Hypertension | With | null | 1 (1.96%) | 1 | 1 (8.33%) | 1 (11.11%) | 1 |
| Without | 30 (100.00%) | 50 (98.04%) | 11 (91.67%) | 8 (88.89%) | |||
| Insulin resistance | With | 19 (63.33%) | 12 (23.53%) | < 0.001* | 7 (58.33%) | 1 (11.11%) | 0.079 |
| Without | 11 (36.67%) | 39 (76.47%) | 5 (41.67%) | 8 (88.89%) | |||
| Menstrual status | Regular | 13 (43.33%) | 20 (39.22%) | 0.896 | 6 (50.00%) | 5 (55.56%) | 1 |
| Irregular | 17 (56.67%) | 31 (60.78%) | 6 (50.00%) | 4 (44.44%) | |||
| P53 | Wild‐type | 27 (90.00%) | 44 (86.27%) | 0.886 | 11 (91.67%) | 7 (77.78%) | 0.787 |
| Mutant | 3 (10.00%) | 7 (13.73%) | 1 (8.33%) | 2 (22.22%) | |||
| Ki‐67 | < 20 | 16 (53.33%) | 4 (80.39%) | 0.020* | 10 (83.33%) | 5 (55.56%) | 0.364 |
| ≥ 20 | 14 (46.67%) | 10 (19.61%) | 2 (16.67%) | 4 (44.44%) | |||
| Myometrial invasion | With | 19 (63.33%) | 7 (13.73%) | < 0.001* | 6 (50.00%) | 1 (11.11%) | 0.160 |
| Without | 11 (36.67%) | 44 (86.27%) | 6 (50.00%) | 8 (88.89%) | |||
| Histological grade | G1 | 27 (90.00%) | 48 (94.12%) | 0.807 | 10 (83.33%) | 9 (100.00%) | 0.591 |
| G2 | 3 (10.00%) | 3 (5.88%) | 2 (16.67%) | null | |||
Represents p < 0.05. Numerical data are presented as mean ± standard deviation. Categorical data are presented as numbers (n%).
TABLE 2.
Univariate and multivariable logistic regression analyses conducted to select clinical features.
| Variable | Univariate analysis | Multivariate analysis | ||
|---|---|---|---|---|
| OR (95% CI) | p | OR (95% CI) | p | |
| Age | 1.229 (0.882, 1.713) | 0.307 | ||
| BMI | 0.968 (0.691, 1.331) | 0.871 | ||
| HB | 1.159 (0.833, 1.611) | 0.463 | ||
| PLT | 0.953 (0.687, 1.323) | 0.809 | ||
| NEUT | 0.817 (0.583, 1.143) | 0.323 | ||
| LYMPH | 1.218 (0.872, 1.701) | 0.332 | ||
| MONO | 0.759 (0.540, 1.067) | 0.183 | ||
| CA125 | 0.918 (0.660, 1.276) | 0.669 | ||
| CA199 | 1.117 (0.801, 1.559) | 0.584 | ||
| AFP | 1.136 (0.815, 1.584) | 0.529 | ||
| HE4 | 0.275 (0.167, 0.451) | < 0.001* | 0.431 (0.232, 0.799) | 0.025* |
| ALB | 0.903 (0.647, 1.260) | 0.614 | ||
| ALT | 0.906 (0.652, 1.260) | 0.623 | ||
| AST | 1.004 (0.723, 1.392) | 0.986 | ||
| TB | 1.243 (0.886, 1.744) | 0.291 | ||
| URE | 1.260 (0.901, 1.759) | 0.256 | ||
| CRE | 1.216 (0.872, 1.697) | 0.333 | ||
| Max‐diameter | 0.228 (0.135, 0.387) | < 0.001* | 0.194 (0.062, 0.608) | 0.018* |
| Parity | 0.989 (0.713, 1.373) | 0.957 | ||
| PCOS | 0.408 (0.269, 0.620) | < 0.001* | 0.469 (0.286, 0.770) | 0.012* |
| Thyroid disease | 0.822 (0.563, 1.200) | 0.394 | ||
| Diabetes | 0.993 (0.715, 1.377) | 0.971 | ||
| Hypertension | 1.057 (0.759, 1.473) | 0.784 | ||
| Insulin resistance | 0.421 (0.292, 0.606) | < 0.001* | 0.540 (0.329, 0.886) | 0.041* |
| Menstrual status | 1.073 (0.773, 1.489) | 0.724 | ||
| P53 | 1.176 (0.842, 1.642) | 0.425 | ||
| Ki‐67 | 0.726 (0.519, 1.015) | 0.116 | ||
| Myometrial invasion | 0.332 (0.222, 0.497) | < 0.001* | 0.506 (0.299, 0.855) | 0.033* |
| Histological grade | 0.767 (0.536, 1.097) | 0.223 | ||
Abbreviations: CI, confidence interval; OR, odds ratio.
Represents p < 0.05.
3.2. Radiomic Signature Models
A total of 1197 radiological features were extracted from MRI images. Following feature scaling and selection, 13 features with nonzero coefficients were retained. Six machine learning methods were employed in this study to construct radiomics models, with the ROC curves for both the training and testing cohorts shown in Figure 4. In the training cohort, all six models achieved AUCs above 0.85, with XGBoost performing best, yielding an AUC of 0.998, an accuracy of 0.975, a sensitivity of 0.961, a specificity of 1.000, a PPV of 1.000, and an NPV of 0.937. However, in the testing cohort, only one radiomics model, the LR model, achieved an AUC above 0.85. The LR model recorded an AUC of 0.852, an accuracy of 0.810, a sensitivity of 0.667, a specificity of 0.917, a PPV of 0.857, and an NPV of 0.786 (Table 3). The results suggest that while the XGBoost model performed well in the training cohort, it showed signs of overfitting. To ensure model stability and generalizability, we ultimately selected the LR model as the optimal radiomics model.
FIGURE 4.
The receiver operating characteristic (ROC) curves of six models in the training (a) and testing (b) cohorts. LR logistic regression, SVM support vector machine, KNN K‐nearest neighbors, RF random forest, XGBoost extreme gradient boosting, LightGBM Light Gradient Boosting Machine.
TABLE 3.
Different models for predicting the effect of fertility preservation in training and test cohorts.
| Model | AUC (95% CI) | Accuracy | Sensitivity | Specificity | PPV | NPV | |
|---|---|---|---|---|---|---|---|
| LR | Training | 0.875 (0.798, 0.951) | 0.878 | 0.706 | 0.900 | 0.923 | 0.643 |
| Test | 0.852 (0.701, 0.960) | 0.810 | 0.667 | 0.917 | 0.857 | 0.786 | |
| SVM | Training | 0.957 (0.914, 0.999) | 0.914 | 0.922 | 0.900 | 0.940 | 0.871 |
| Test | 0.704 (0.464, 0.943) | 0.667 | 0.667 | 0.667 | 0.600 | 0.727 | |
| KNN | Training | 0.855 (0.779, 0.930) | 0.531 | 0.255 | 1.000 | 1.000 | 0.441 |
| Test | 0.708 (0.473, 0.943) | 0.571 | 0.222 | 0.833 | 0.500 | 0.588 | |
| RF | Training | 0.936 (0.886, 0.986) | 0.864 | 0.902 | 0.800 | 0.885 | 0.828 |
| Test | 0.759 (0.548, 0.970) | 0.667 | 0.889 | 0.500 | 0.571 | 0.857 | |
| XGBoost | Training | 0.998 (0.993, 1.000) | 0.975 | 0.961 | 1.000 | 1.000 | 0.937 |
| Test | 0.782 (0.572, 0.992) | 0.762 | 0.556 | 0.917 | 0.833 | 0.733 | |
| LightGBM | Training | 0.905 (0.842, 0.966) | 0.802 | 0.725 | 0.933 | 0.949 | 0.667 |
| Test | 0.657 (0.413, 0.900) | 0.571 | 0.111 | 0.917 | 0.500 | 0.579 |
Abbreviations: 95% CI, 95% confidence interval; AUC, area under the curve; KNN, K‐nearest neighbors; LightGBM, light gradient boosting machine; LR, logistic regression; NPV, negative predictive value; PPV, positive prediction value; RF, random forest; SVM, support vector machine; XGBoost, extreme gradient boosting.
3.3. Combined Model Construction and Validation
We developed a combined model by integrating five clinical independent predictors and 13 radiomic features. Clinical, radiomic, and combined models were constructed using logistic regression (LR), and the performance of each model was systematically evaluated. The ROC curves for the three models in both the training and testing cohorts are displayed in Figure 5. In both cohorts, the combined model demonstrated significantly superior discriminative ability compared to the clinical and radiomic models (Table 4), with an AUC of 0.941 in the training cohort and 0.907 in the testing cohort. The DCA curves for the three models are shown in Figure 6. The results suggest that the combined model offered the greatest net benefit in predicting fertility preservation success across both cohorts.
FIGURE 5.
Receiver operating characteristic (ROC) curves for clinical model, radiomics model and combined model in training cohort (a) and testing cohort (b).
TABLE 4.
Performance of the clinical model, radiomics model, and combined model in the training and testing cohorts.
| Model | AUC (95% CI) | Accuracy | Sensitivity | Specificity | PPV | NPV | |
|---|---|---|---|---|---|---|---|
| Clinical Model* | Training | 0.927 (0.874, 0.980) | 0.827 | 0.765 | 0.933 | 0.951 | 0.700 |
| Test | 0.769 (0.552, 0.985) | 0.762 | 0.889 | 0.667 | 0.667 | 0.889 | |
| Radiomics Model* | Training | 0.875 (0.798, 0.951) | 0.878 | 0.706 | 0.900 | 0.923 | 0.643 |
| Test | 0.852 (0.701, 0.960) | 0.810 | 0.667 | 0.917 | 0.857 | 0.786 | |
| Combined Model* | Training | 0.941 (0.890, 0.991) | 0.901 | 0.941 | 0.833 | 0.906 | 0.893 |
| Test | 0.907 (0.778, 1.000) | 0.810 | 0.889 | 0.750 | 0.727 | 0.900 |
Abbreviations: 95% CI, 95% confidence interval; AUC, area under the curve; NPV, negative predictive value; PPV, positive prediction value.
Represents models that were developed using LR.
FIGURE 6.
Decision curve analysis (DCA) for three models in training (a) and testing (b) cohorts. Net Benefit = TP/N − FP/N × pT/(1 − pT), where pT denotes the threshold probability (0.1–0.8). The clinical endpoint was 6 month CR vs. non‐CR. We plotted decision curves comparing the “Fertility‐sparing treatment” and “No fertility‐sparing treatment” strategies.
4. Discussion
4.1. Principal Findings
We created a machine‐learning‐based fusion model that integrates clinical predictors and radiomic features using a logistic regression (LR) approach to predict fertility preservation prognosis effectively. Our results demonstrate the model's efficacy, providing clinicians with a feasible and reliable noninvasive tool for assessing the suitability of fertility‐sparing treatments in young patients with endometrial cancer. In recent years, studies examining factors that influence the response rates of fertility‐sparing treatments in endometrial cancer patients have primarily focused on isolated clinical characteristics. To date, no prognostic model aimed at predicting fertility preservation outcomes in endometrial cancer patients has been developed.
4.2. Results in the Context of What Is Known
Our findings indicate that among patients who did not achieve remission after 6 months of treatment, the proportion with PCOS and insulin resistance is significantly higher compared to those who achieved remission, corroborating previous research findings [4, 28]. This disparity may be attributed to damage to progesterone receptors in PCOS patients, leading to progesterone resistance [29, 30]. Moreover, the excess endogenous estrogen produced by adipose tissue in insulin‐resistant patients can counteract progesterone therapy [8]. Li et al.'s study further demonstrates that PCOS patients are more susceptible to relapse [12].
Notably, our study identified serum human HE4 level as an independent predictor of fertility‐sparing treatment outcomes. Patients responsive to progestogen therapy exhibited significantly lower baseline serum HE4 levels compared to nonresponders, aligning with Behrouzi et al.'s findings [15]. Recent research has demonstrated a correlation between elevated serum HE4 levels and advanced tumor stage and grade [31]. Furthermore, HE4 has been shown to enhance the proliferation of endometrial cancer cell lines [32, 33]. These findings suggest that HE4 may play a crucial role in promoting tumor invasiveness, proliferation, and resistance to progestogen therapy [34, 35].
4.3. Research Implications
Radiomics is an approach that involves extracting and analyzing large quantities of data through the examination of features in medical imaging modalities, including CT and MRI. The fundamental principle of radiomics is to uncover potentially valuable information within imaging characteristics, which can be utilized for disease diagnosis, prognostic assessment, and prediction of treatment outcomes. In our study, the three most important radiomic features are wavelet_LHL_glrlm_ShortRunEmphasis, wavelet_HHL_glcm_Contrast, and log_sigma_30mm_3D_glszm_SmallAreaLowGrayLevelEmphasis. ShortRunEmphasis quantifies the distribution of short run lengths, with higher values indicating shorter run lengths and finer texture. Contrast assesses local intensity variation, with larger values associated with greater intensity differences between adjacent voxels. SmallAreaLowGrayLevelEmphasis evaluates the proportion of smaller areas with lower gray‐level values in the image. These variables suggest that endometrial cancer tumors exhibiting coarser texture, higher gray‐level values, and lower contrast are more prone to developing progestogen tolerance. Similar findings have been observed in studies on myometrial invasion in endometrial cancer [36], suggesting that tumors with poor prognosis are often accompanied by varying degrees of necrosis, hemorrhage, and differences in tissue microstructure between different regions [37] heterogeneities that are typically imperceptible to radiologists through visual inspection alone.
4.4. Clinical Implications
Our study demonstrates that the linear LR model outperforms all other models, possibly due to the limited generalization capability and risk of overfitting associated with more complex models when applied to small sample cohorts, coupled with the relatively low number of endometrial cancer patients eligible for fertility preservation [38]. Nevertheless, the LR model achieved impressive performance with fewer training samples while maintaining satisfactory stability and efficiency. Our combined model demonstrated superior predictive performance, followed by the standalone radiomics model, and lastly, the clinical feature model. This finding suggests that radiomic features can provide valuable information on tumor phenotype and microenvironment, complementing other data sources, including clinical characteristics [39]. An integrated model incorporating all features has the potential to yield an accurate and robust evidence‐based clinical decision support system [40].
4.5. Strengths and Limitations
This is currently the largest single‐center fertility‐sparing study in endometrial cancer. Furthermore, our study develops the first prognostic model integrating clinical and radiomic features for fertility‐sparing treatment in endometrial cancer, which demonstrates the potential of machine learning in optimizing patient selection for fertility‐preserving approaches and provides a noninvasive tool to assist clinicians in treatment decision‐making for young endometrial cancer patients. This study has several limitations. First, as a retrospective single‐center study, there is a potential for selection bias. Second, manual lesion segmentation is prone to subjectivity and experiential influence [41]; we aim to achieve full automation through deep learning techniques in future studies. Third, we did not extensively extract deep learning‐based features; in future studies, upon collecting a sufficient sample size, we intend to conduct deep learning research in conjunction with traditional radiomics.
5. Conclusion
In summary, our research proposes and validates a novel prognostic model for fertility‐preserving treatment in endometrial cancer patients, integrating both clinical and radiomic features derived from MRI. Our study demonstrates that the combined model using LR exhibited superior performance, potentially paving the way for the development of precision medicine approaches and the enhancement of clinical treatment strategies in this field.
Author Contributions
Wenhan Yuan: conceptualization; data curation; visualization; writing – original draft; writing – review and editing; formal analysis; validation. Haidan Yin: software; visualization. Jianhong Liu: data curation. Ying Zheng: conceptualization; methodology; writing – review and editing; funding acquisition.
Disclosure
During the preparation of this work, the author used ChatGPT in order to improve the readability and language of the manuscript. After using this tool, the author reviewed and edited the content as needed and takes full responsibility for the content of the publication.
Ethics Statement
The study was reviewed and approved by the institutional review board of the West China Second Hospital of Sichuan University (IRB No. 2023[024]).
Consent
Written informed consent was obtained from all the participants prior to the enrollment (or for the publication) of this study.
Conflicts of Interest
The authors declare no conflicts of interest.
Acknowledgments
This work was supported by the National Key Research and Development Program of China (grant number 2022YFC2704103).
Yuan W., Yin H., Liu J., and Zheng Y., “Using Machine Learning to Predict the Prognosis in Endometrial Cancer Patients Undergoing Fertility‐Sparing Treatment,” Cancer Medicine 14, no. 22 (2025): e71355, 10.1002/cam4.71355.
Funding: This work was supported by the National Key Research and Development Program of China (2022YFC2704103).
Data Availability Statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.
References
- 1. Sung H., Ferlay J., Siegel R. L., et al., “Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries,” CA: A Cancer Journal for Clinicians 71 (2021): 209–249. [DOI] [PubMed] [Google Scholar]
- 2. Pellerin G. P. and Finan M. A., “Endometrial Cancer in Women 45 Years of Age or Younger: A Clinicopathological Analysis,” American Journal of Obstetrics and Gynecology 193, no. 5 (2005): 1640–1644. [DOI] [PubMed] [Google Scholar]
- 3. Soliman P. T., Oh J. C., Schmeler K. M., et al., “Risk Factors for Young Premenopausal Women With Endometrial Cancer,” Obstetrics and Gynecology 105, no. 3 (2005): 575–580. [DOI] [PubMed] [Google Scholar]
- 4. Zhou R., Yang Y., Lu Q., et al., “Prognostic Factors of Oncological and Reproductive Outcomes in Fertility‐Sparing Treatment of Complex Atypical Hyperplasia and Low‐Grade Endometrial Cancer Using Oral Progestin in Chinese Patients,” Gynecologic Oncology 139, no. 3 (2015): 424–428. [DOI] [PubMed] [Google Scholar]
- 5. Gotlieb W. H., Beiner M. E., Shalmon B., et al., “Outcome of Fertility‐Sparing Treatment With Progestins in Young Patients With Endometrial Cancer,” Obstetrics and Gynecology 102, no. 4 (2003): 718–725. [DOI] [PubMed] [Google Scholar]
- 6. Koskas M., Uzan J., Luton D., Rouzier R., and Daraï E., “Prognostic Factors of Oncologic and Reproductive Outcomes in Fertility‐Sparing Management of Endometrial Atypical Hyperplasia and Adenocarcinoma: Systematic Review and Meta‐Analysis,” Fertility and Sterility 101, no. 3 (2014): 785–794. [DOI] [PubMed] [Google Scholar]
- 7. Silva J. L., Paulino E., Dias M. F., and Melo A. C., “Endometrial Cancer: Redefining the Molecular‐Targeted Approach,” Cancer Chemotherapy and Pharmacology 76, no. 1 (2015): 1–11. [DOI] [PubMed] [Google Scholar]
- 8. Park J.‐Y., Kim D.‐Y., Kim J.‐H., et al., “Long‐Term Oncologic Outcomes After Fertility‐Sparing Management Using Oral Progestin for Young Women With Endometrial Cancer (KGOG 2002),” European Journal of Cancer 49, no. 4 (2013): 868–874. [DOI] [PubMed] [Google Scholar]
- 9. Ramirez P. T., Frumovitz M., Bodurka D. C., Sun C. C., and Levenback C., “Hormonal Therapy for the Management of Grade 1 Endometrial Adenocarcinoma: A Literature Review,” Gynecologic Oncology 95, no. 1 (2004): 133–138. [DOI] [PubMed] [Google Scholar]
- 10. Niwa K., Tagami K., Lian Z., Onogi K., Mori H., and Tamaya T., “Outcome of Fertility‐Preserving Treatment in Young Women With Endometrial Carcinomas,” BJOG: An International Journal of Obstetrics and Gynaecology 112, no. 3 (2005): 317–320. [DOI] [PubMed] [Google Scholar]
- 11. Fukui Y., Taguchi A., Adachi K., et al., “Polycystic Ovarian Morphology May Be a Positive Prognostic Factor in Patients With Endometrial Cancer Who Achieved Complete Remission After Fertility‐Sparing Therapy With Progestin,” Asian Pacific Journal of Cancer Prevention 18 (2017): 3111–3116. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Li M., Guo T., Cui R., Feng Y., Bai H., and Zhang Z., “Weight Control Is Vital for Patients With Early‐Stage Endometrial Cancer or Complex Atypical Hyperplasia Who Have Received Progestin Therapy to Spare Fertility: A Systematic Review and Meta‐Analysis,” Cancer Management and Research 11 (2019): 4005–4021. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Chung Y. S., Woo H. Y., Lee J.‐Y., et al., “Mismatch Repair Status Influences Response to Fertility‐Sparing Treatment of Endometrial Cancer,” American Journal of Obstetrics and Gynecology 224 (2021): e313–e371. [DOI] [PubMed] [Google Scholar]
- 14. Baxter E., Brennan D. J., McAlpine J. N., et al., “Improving Response to Progestin Treatment of Low‐Grade Endometrial Cancer,” International Journal of Gynecological Cancer 30 (2020): 1811–1823. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Behrouzi R., Ryan N. A. J., Barr C. E., et al., “Baseline Serum HE4 but Not Tissue HE4 Expression Predicts Response to the Levonorgestrel‐Releasing Intrauterine System in Atypical Hyperplasia and Early Stage Endometrial Cancer,” Cancers 12, no. 2 (2020): 276, 10.3390/cancers12020276. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Rodolakis A., Scambia G., Planchamp F., et al., “ESGO/ESHRE/ESGE Guidelines for the Fertility‐Sparing Treatment of Patients With Endometrial Carcinoma,” Human Reproduction Open 2023, no. 1 (2023): hoac057. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Lee A. J., Yang E. J., Kim N. K., et al., “Fertility‐Sparing Hormonal Treatment in Patients With Stage I Endometrial Cancer of Grade 2 Without Myometrial Invasion and Grade 1‐2 With Superficial Myometrial Invasion: Gynecologic Oncology Research Investigators coLLaborAtion Study (GORILLA‐2001),” Gynecologic Oncology 174 (2023): 106–113. [DOI] [PubMed] [Google Scholar]
- 18. Gulshan V., Peng L., and Coram M., “Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs,” JAMA 316, no. 22 (2016): 2402–2410. [DOI] [PubMed] [Google Scholar]
- 19. Chilamkurthy S., Ghosh R., and Tanamala S., “Deep Learning Algorithms for Detection of Critical Findings in Head CT Scans: A Retrospective Study,” Lancet 392, no. 10162 (2018): 2388–2396. [DOI] [PubMed] [Google Scholar]
- 20. Poplin R., Varadarajan A. V., and Blumer K., “Prediction of Cardiovascular Risk Factors From Retinal Fundus Photographs via Deep Learning,” Nature Biomedical Engineering 2, no. 3 (2018): 158–164. [DOI] [PubMed] [Google Scholar]
- 21. Gillies R. J., Kinahan P. E., and Hricak H., “Radiomics: Images Are More Than Pictures, They Are Data,” Radiology 278 (2016): 563–577. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Chen X., Wang Y., Shen M., et al., “Deep Learning for the Determination of Myometrial Invasion Depth and Automatic Lesion Identification in Endometrial Cancer MR Imaging: A Preliminary Study in a Single Institution,” European Radiology 30, no. 9 (2020): 4985–4994, 10.1007/s00330-020-06870-1. [DOI] [PubMed] [Google Scholar]
- 23. Stanzione A., Cuocolo R., Del Grosso R., et al., “Deep Myometrial Infiltration of Endometrial Cancer on Mri: A Radiomics‐Powered Machine Learning Pilot Study,” Academic Radiology 28, no. 5 (2021): 737–744, 10.1016/j.acra.2020.02.028. [DOI] [PubMed] [Google Scholar]
- 24. Luo Y., Mei D., Gong J., et al., “Multiparametric MRI‐Based Radiomics Nomogram for Predicting Lymphovascular Space Invasion in Endometrial Carcinoma[J],” Journal of Magnetic Resonance Imaging 52, no. 4 (2020): 1257–1262, 10.1002/jmri.27142. [DOI] [PubMed] [Google Scholar]
- 25. van Griethuysen J., Fedorov A., Parmar C., et al., “Computational Radiomics System to Decode the Radiographic Phenotype,” Cancer Research 77, no. 21 (2017): e104–e107, 10.1158/0008-5472.CAN-17-0339. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Sauerbrei W., Royston P., and Binder H., “Selection of Important Variables and Determination of Functional Form for Continuous Predictors in Multivariable Model Building,” Statistics in Medicine 26, no. 30 (2007): 5512–5528, 10.1002/sim.3148. [DOI] [PubMed] [Google Scholar]
- 27. Parmar C., Grossmann P., Bussink J., Lambin P., and Aerts H. J. W. L., “Machine Learning Methods for Quantitative Radiomic Biomarkers,” Scientific Reports 5 (2015): 13087. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Shao R., Li X., and Billig H., “Promising Clinical Practices of Metformin in Women With PCOS and Early‐Stage Endometrial Cancer,” BBA Clinical 2 (2014): 7–9, 10.1016/j.bbacli.2014.07.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Li X., Feng Y., Lin J. F., Billig H., and Shao R., “Endometrial Progesterone Resistance and PCOS,” Journal of Biomedical Science 21, no. 1 (2014): 2, 10.1186/1423-0127-21-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30. Savaris R. F., Groll J. M., Young S. L., et al., “Progesterone Resistance in PCOS Endometrium: A Microarray Analysis in Clomiphene Citrate‐Treated and Artificial Menstrual Cycles,” Journal of Clinical Endocrinology and Metabolism 96, no. 6 (2011): 1737–1746, 10.1210/jc.2010-2600. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Karlsen N. S., Karlsen M. A., Høgdall C. K., and Høgdall E. V., “HE4 Tissue Expression and Serum HE4 Levels in Healthy Individuals and Patients With Benign or Malignant Tumors: A Systematic Review,” Cancer Epidemiology, Biomarkers & Prevention 23, no. 11 (2014): 2285–2295, 10.1158/1055-9965. [DOI] [PubMed] [Google Scholar]
- 32. Li J., Chen H., Mariani A., et al., “HE4 (WFDC2) Promotes Tumor Growth in Endometrial Cancer Cell Lines,” International Journal of Molecular Sciences 14, no. 3 (2013): 6026–6043, 10.3390/ijms14036026. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33. Lu Q., Chen H., Senkowski C., et al., “Recombinant HE4 Protein Promotes Proliferation of Pancreatic and Endometrial Cancer Cell Lines,” Oncology Reports 35, no. 1 (2016): 163–170, 10.3892/or.2015.4339. [DOI] [PubMed] [Google Scholar]
- 34. Mutz‐Dehbalaie I., Egle D., Fessler S., et al., “HE4 Is an Independent Prognostic Marker in Endometrial Cancer Patients,” Gynecologic Oncology 126, no. 2 (2012): 186–191, 10.1016/j.ygyno.2012.04.022. [DOI] [PubMed] [Google Scholar]
- 35. Brennan D. J., Hackethal A., Metcalf A. M., et al., “Serum HE4 as a Prognostic Marker in Endometrial Cancer‐A Population Based Study,” Gynecologic Oncology 132, no. 1 (2014): 159–165, 10.1016/j.ygyno.2013.10.036. [DOI] [PubMed] [Google Scholar]
- 36. Wang Y., Bi Q., Deng Y., et al., “Development and Validation of an MRI‐Based Radiomics Nomogram for Assessing Deep Myometrial Invasion in Early Stage Endometrial Adenocarcinoma,” Academic Radiology 30, no. 4 (2023): 668–679, 10.1016/j.acra.2022.05.017. [DOI] [PubMed] [Google Scholar]
- 37. Ghosh A., Singh T., Singla V., Bagga R., Srinivasan R., and Khandelwal N., “DTI Histogram Parameters Correlate With the Extent of Myoinvasion and Tumor Type in Endometrial Carcinoma: A Preliminary Analysis,” Acta Radiologica 61, no. 5 (2020): 675–684, 10.1177/0284185119875019. [DOI] [PubMed] [Google Scholar]
- 38. Mao B., Ma J., Duan S., Xia Y., Tao Y., and Zhang L., “Preoperative Classification of Primary and Metastatic Liver Cancer via Machine Learning‐Based Ultrasound Radiomics,” European Radiology 31, no. 7 (2021): 4576–4586, 10.1007/s00330-020-07562-6. [DOI] [PubMed] [Google Scholar]
- 39. Gatenby R. A., Grove O., and Gillies R. J., “Quantitative Imaging in Cancer Evolution and Ecology,” Radiology 269, no. 1 (2013): 8–15, 10.1148/radiol.13122697. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40. Lambin P., Leijenaar R. T. H., Deist T. M., et al., “Radiomics: The Bridge Between Medical Imaging and Personalized Medicine,” Nature Reviews. Clinical Oncology 14, no. 12 (2017): 749–762, 10.1038/nrclinonc.2017.141. [DOI] [PubMed] [Google Scholar]
- 41. Lohmann P., Bousabarah K., Hoevels M., and Treuer H., “Radiomics in Radiation Oncology‐Basics, Methods, and Limitations,” Strahlentherapie Und Onkologie 196, no. 10 (2020): 848–855, 10.1007/s00066-020-01663-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.