Endometrial cancer (EC) is one of the most common gynecological malignancies worldwide, and its incidence is rising in East Asia due to dietary and lifestyle changes. While most patients present early with vaginal bleeding and have favorable outcomes, specific subtypes carry significantly worse prognoses.
In 2013, The Cancer Genome Atlas introduced a molecular classification system for EC that divides tumors into four subtypes based on genetic markers: POLE-mutated, microsatellite instability-high (MSI-H), copy-number-low (CNL), and copy-number-high with p53 mutation (p53abn).
The p53-abnormal (p53abn) subtype is classified as high-risk. These tumors tend to be non-endometrioid histological types (like serous carcinoma) and are associated with higher recurrence rates and worse overall survival compared to the other molecular subtypes.
Accurate identification of the p53abn subtype before surgery is clinically valuable because it changes treatment planning - these patients may need more aggressive surgical staging, adjuvant chemotherapy, or clinical trial enrollment. Currently, molecular subtyping requires tissue sampling, limiting preoperative availability.
Radiomics is a field that uses computer algorithms to extract hundreds of quantitative features from medical images - such as texture, shape, and intensity patterns - that are invisible to the human eye. These features can reflect underlying tumor biology and predict clinical outcomes.
Traditional radiomics treats the entire tumor as a single homogeneous region, extracting features from the whole mass. However, tumors are biologically heterogeneous, meaning different areas within the same tumor can have very different properties.
Habitat analysis addresses this by dividing tumors into biologically distinct sub-regions (habitats) based on imaging characteristics. Each habitat can then be analyzed separately, capturing the internal diversity of the tumor microenvironment in ways that whole-region analysis misses.
MRI is the gold standard imaging modality for EC staging and is routinely performed before surgery. Combining MRI-based habitat radiomics with machine learning offers a potential non-invasive pathway to predict molecular subtype - including p53abn status - before tissue is even obtained.
This multicentre study enrolled 494 patients from three hospitals: a training cohort of 270 patients (FOGH, January 2019 to August 2022), test cohort 1 of 70 patients (FSCC, January 2021 to January 2022), and test cohort 2 of 154 patients (BCYH, January 2022 to March 2023). This design tests whether the model generalizes across institutions and scanner types.
Of the 494 patients, 92 (18.6%) had p53abn EC. MRI was performed using both 1.5T and 3.0T scanners at different centres, capturing T1-weighted, T2-weighted, diffusion-weighted (DWI), and contrast-enhanced (CE) images for each patient.
Molecular subtyping was performed using genetic sequencing panels targeting key genes including TP53, POLE, and mismatch repair genes. A gynaecologic pathologist with extensive experience reviewed all samples alongside molecular genetics experts to establish definitive diagnoses.
A key challenge was that the three hospitals used different MRI scanners with different field strengths and protocols. To ensure comparability, all images underwent standardized preprocessing including N4 bias field correction and resampling to isotropic 1mm voxels.
Tumor subregions were created using the K-means clustering algorithm, which groups voxels (3D pixels) based on their DWI signal intensity and contrast-enhancement values. Unlike manual segmentation, this approach is data-driven and reproducible.
After testing cluster numbers from 2 to 10, the Calinski-Harabasz criterion identified four as the optimal number of clusters. This produced four biologically meaningful habitat sub-regions: low DWI/low CE (subregion 1), high DWI/low CE (subregion 2), low DWI/high CE (subregion 3), and high DWI/high CE (subregion 4).
From these four subregions across four MRI sequences (T1WI, T2WI, DWI, CE), a total of 1,920 habitat radiomic features were extracted using the PyRadiomics software package, covering intensity statistics, shape, and texture measures.
Feature selection used a rigorous three-step process: statistical t-tests to remove non-discriminative features, Spearman correlation filtering to remove redundant features, and LASSO regression to select the final compact set of 8 features for model construction.
Among three machine learning classifiers (logistic regression, support vector machine, and random forest), the support vector machine (SVM) model using habitat features achieved the highest AUC: 0.855 in the training cohort, 0.769 in test cohort 1, and 0.766 in test cohort 2.
The random forest model showed near-perfect performance in training (AUC 0.997) but collapsed to near-chance in testing (AUC 0.511 and 0.542), a classic sign of severe overfitting. The SVM model was selected as the final Habitat Radiomics Model (HRM) for its consistent cross-cohort generalization.
A Whole-Region Radiomics Model (WRM) using features from the entire tumor (not split into habitats) achieved lower AUCs: 0.707 training, 0.703 test 1, 0.738 test 2. DeLong's statistical test confirmed the HRM significantly outperformed the WRM in training and test cohort 2, demonstrating the added value of habitat analysis.
A clinical model using just five conventional clinical variables (age, BMI, CA125, tumor diameter, and ADC value) achieved AUCs of 0.709, 0.641, and 0.665 - significantly worse than the HRM across all three cohorts, confirming that imaging features carry predictive information beyond what clinical data alone provides.
To make the habitat radiomics model interpretable for clinicians, SHAP (SHapley Additive ExPlanations) analysis was applied to the final SVM model. SHAP values quantify how much each feature pushes a prediction toward or away from p53abn classification.
The SHAP summary plot revealed which of the eight selected features dominated the model's predictions. Four features were derived from DWI images and four from T2WI images, with four features from subregion 1 (the low DWI/low CE habitat) carrying the most predictive weight.
For individual patients, SHAP force plots visualize how specific feature values combine to produce the final probability estimate. This transparency allows clinicians to see not just whether a patient is predicted to have p53abn disease, but which imaging characteristics drove that prediction.
This explainability addresses a critical limitation of many radiomics studies, which produce accurate but opaque predictions. By showing feature contributions at the individual level, the model can support rather than replace clinical judgment, helping oncologists understand the basis for AI-generated risk estimates.
This study demonstrates the feasibility of predicting p53abn endometrial cancer non-invasively from preoperative MRI using habitat radiomics, achieving consistent AUCs around 0.77 in independent test cohorts from different hospitals and scanner types.
Identifying p53abn tumors before surgery enables more informed treatment planning - these high-risk patients may benefit from modified surgical staging protocols, earlier consideration of adjuvant chemotherapy, or enrollment in clinical trials for novel agents targeting this molecular subtype.
The habitat approach's advantage over whole-region radiomics suggests that intratumoral heterogeneity - the biological diversity within a single tumor - contains prognostically relevant information that single-region analyses discard. This validates habitat analysis as a meaningful refinement of standard radiomics methodology.
Future work should expand the patient cohort, incorporate additional MRI sequences, and explore integration with other molecular markers. Prospective validation in clinical practice will be needed before the model can influence treatment decisions in routine oncology care.