TP53 is often called the 'guardian of the genome' - it is the most commonly mutated gene across all human cancers. The TP53 protein normally acts as a tumor suppressor, detecting DNA damage and either repairing it or triggering cell death to prevent cancer from developing. When TP53 is mutated, this protective mechanism fails.
In endometrial cancer, TP53 mutations are strongly associated with greater tumor aggressiveness, drug resistance, and poorer patient prognosis. TP53-mutant EC belongs to the high-risk molecular subtype (also called the p53-abnormal or p53abn class in the current TCGA/ProMisE classification), which carries the worst outcomes among all EC molecular subtypes.
Knowing a patient's TP53 mutation status is clinically important because it directly influences treatment planning - patients with TP53-mutant tumors typically need more aggressive adjuvant therapy (chemotherapy, radiotherapy, or both) after surgery, whereas TP53-wild-type patients may be candidates for less intensive treatment.
Currently, TP53 status is assessed by immunohistochemistry (IHC) - a laboratory technique applied to tumor tissue samples that detects protein expression using antibodies. While effective, IHC requires adequate tissue from a biopsy or surgical specimen, is costly, and is typically only performed after surgery, limiting its usefulness for preoperative planning.
Diffusion-Weighted MRI (DWI) is a specialized MRI technique that measures how freely water molecules move (diffuse) within tissue. In densely packed cancer cells, water movement is restricted more than in normal tissue, producing measurable differences in a parameter called the Apparent Diffusion Coefficient (ADC). Lower ADC values typically indicate more cellular, more aggressive tumors.
Radiomics is the computational extraction of hundreds or thousands of quantitative features from medical images - features describing texture, shape, intensity patterns, and spatial relationships that are imperceptible to the human eye. These features can reveal biologically meaningful patterns in how tumors appear on MRI that correlate with molecular characteristics like mutation status.
Deep learning (DL) using Convolutional Neural Networks (CNNs) takes this further by automatically learning which visual patterns in images are predictive of a given outcome, without needing the features to be manually defined. CNNs are trained on large image datasets and can detect subtle differences in tumor appearance that would be missed by traditional radiomics or visual inspection.
Previous studies have validated radiomics-based MRI models for predicting TP53 mutation status in rectal and breast cancers, and deep learning has shown promise in EC for assessing lymph node metastasis and staging. However, no prior study had specifically combined DL and radiomics from DWI to predict TP53 mutation status in EC.
The study included 155 EC patients from two medical centers, divided into a training set (80 patients from Center I), a test set (35 patients from Center I), and an external validation set (40 patients from Center II, a different institution). The two-center design is critical for real-world applicability because it tests whether the model generalizes to patients scanned on different machines with different protocols.
TP53 mutation status was established by pathologists using immunohistochemistry on surgically removed tissue specimens. All patients had pre-operative pelvic MRI scans, and 3D tumor volumes were manually delineated on DWI images by experienced radiologists. Clinical variables including ADC value, age, tumor size, blood tumor markers (CEA, CA153), and FIGO stage were recorded for each patient.
From each patient's DWI scan, the team extracted two types of image features: 2,264 radiomics features (statistical, textural, and shape-based measurements computed using the PyRadiomics software with 12 different image filters applied), and 128 deep learning features (learned automatically by a custom multi-scale CNN architecture designed to capture both large-scale tumor structure and fine-grained textural patterns).
To select the most informative features and avoid overfitting, the researchers applied a multi-stage selection pipeline: statistical testing (Mann-Whitney U), a variance-based selection algorithm (SelectKBest), and LASSO regression (Least Absolute Shrinkage and Selection Operator) - a regularization technique that automatically eliminates redundant or weakly predictive features by penalizing model complexity.
Among the clinical variables tested, only two showed a significant relationship with TP53 mutation status: ADC value and FIGO stage. TP53-mutant tumors had consistently and significantly lower ADC values than TP53-wild-type tumors across all three data cohorts (training, test, and validation), confirming that TP53-mutant EC is more cellular and more restricted in water diffusion.
TP53-mutant tumors were also more frequently found at advanced FIGO stages compared to TP53-wild-type tumors. In the training set, 12.96% of TP53-mutant patients had Stage III or IV disease, while all TP53-wild-type patients had early-stage (I-II) disease. This confirms that TP53 mutations drive more aggressive, invasive tumor behavior that leads to higher staging at diagnosis.
Importantly, standard blood biomarkers (CEA and CA153), patient age, and maximum tumor diameter did not significantly differ between TP53-mutant and TP53-wild-type patients. This underscores the limitations of conventional clinical variables for molecular subtype prediction and reinforces the need for imaging-based approaches.
The clinical-only model using FIGO stage and ADC achieved AUC values of 0.780 (training), 0.685 (test), and 0.695 (validation) - demonstrating some discriminatory power but insufficient accuracy for reliable clinical use, particularly in the test and validation cohorts where performance dropped considerably.
Each of the three single-modality models (clinical, radiomics, DL) achieved moderate performance on its own. The DL model outperformed the radiomics model in training and test sets, suggesting CNNs extract more diagnostically relevant information from DWI images than traditional feature engineering. However, combining all three modalities produced the best results by a significant margin.
The final combined Gaussian Process model - incorporating 2 clinical variables, 5 radiomics features, and 4 deep learning features - achieved AUC values of 0.949 (training), 0.877 (test), and 0.914 (external validation). For context, an AUC above 0.9 is generally considered excellent for a diagnostic model, and the external validation performance of 0.914 is particularly meaningful as it comes from a completely independent institution.
The combined model also dramatically outperformed the individual models in Net Reclassification Index (NRI) - a metric measuring how much the combined model improves risk classification relative to single models. NRI values in the external validation set reached 120.83% (vs. clinical), 87.50% (vs. radiomics), and 64.58% (vs. DL), meaning the combined model correctly reclassified a large proportion of patients who were misclassified by any single approach.
The model achieved 100% specificity in the external validation set - meaning it correctly identified all TP53-wild-type patients - while maintaining 70.83% sensitivity. This high specificity has practical value: avoiding false positives prevents TP53-wild-type patients from being unnecessarily subjected to aggressive high-risk treatment protocols.
The Gaussian Process (GP) algorithm is a probabilistic machine learning method that models the relationship between input features and output predictions as a continuous probability distribution rather than a fixed decision rule. Unlike simpler models, GP quantifies uncertainty in its predictions - outputting a probability estimate with confidence intervals rather than just a yes/no classification.
Across all models and data splits, the GP algorithm consistently outperformed the Decision Tree (DT) algorithm in terms of stability and generalization. The researchers attribute GP's advantage to its suitability for small-sample, nonlinear, noisy problems - exactly the characteristics of clinical imaging data, where sample sizes are often limited and the relationship between image features and molecular outcomes is complex and non-linear.
The Decision Tree algorithm, while interpretable and widely used, tends to overfit on small datasets - it can memorize training data patterns without generalizing well to new patients. This was reflected in the DT model's much larger gap between training set performance (high) and test set performance (lower), whereas the GP model maintained more consistent performance across all cohorts.
The study also highlights that choosing the right algorithm matters as much as feature engineering. The authors recommend that future studies test a wider range of algorithms when constructing predictive models, as the optimal algorithm is likely to vary depending on dataset size, feature dimensionality, and the specific clinical prediction task.
The key clinical contribution of this study is demonstrating that TP53 mutation status in EC can be reliably estimated from pre-operative MRI scans, before any surgery or biopsy is performed. This allows TP53 molecular risk classification to inform treatment planning from the outset, rather than only after surgery.
For patients with suspected high-grade EC where TP53 mutation is likely, early knowledge of molecular risk status could prompt more aggressive surgical planning (such as extended lymph node dissection), early referral to specialized oncology centers, or enrollment in clinical trials of novel targeted therapies before the primary operation.
Conversely, for patients where imaging analysis suggests TP53-wild-type status (low-risk molecular subtype), this information could support decisions to consider conservative surgery or fertility-sparing approaches - particularly important for younger patients who wish to preserve the possibility of future pregnancy.
The external validation across two imaging centers with different MRI scanners and protocols adds confidence that the model is practical in real-world clinical settings, though larger prospective multicenter studies are needed before clinical deployment. Future work will focus on expanding sample sizes, optimizing imaging protocols, and integrating additional machine learning algorithms to further improve model performance.