Ultrasound is the standard first-line imaging tool for evaluating postmenopausal women with vaginal bleeding, a common symptom of endometrial cancer (EC). Doctors measure the thickness of the endometrium (uterine lining) on ultrasound - a thicker lining raises concern for cancer - but thickness alone is an imprecise predictor that leads to many unnecessary biopsies.
Machine learning offers a way to extract more information from ultrasound images by analyzing textural features, shape, and intensity patterns invisible to the naked eye. These quantitative features may more accurately distinguish cancerous from benign endometrial changes compared to simple thickness measurements.
This study developed and validated a Deep Learning Radiomics (DLR) model that combines classical radiomics features with deep learning features extracted by a convolutional neural network, then uses a support vector machine (SVM) to classify EC risk. The model was tested in 1,861 postmenopausal women from two medical centers.
The research team developed a pipeline with several key steps. First, super-resolution (SRGAN) preprocessing was applied to ultrasound images to enhance image resolution before analysis. SRGAN (Super-Resolution Generative Adversarial Network) uses deep learning to reconstruct fine details that are lost in standard ultrasound image compression, effectively creating higher-quality inputs for the subsequent analysis.
Two region-of-interest (ROI) strategies were compared: one focused on the endometrium directly, the other on the broader uterine corpus. Features were extracted from both classical radiomics (texture statistics, shape descriptors) and a CNN that learned image patterns automatically. A support vector machine (SVM) classifier combined these feature sets to output a risk score.
Data came from 1,861 postmenopausal women across two centers: one served as the training/internal validation cohort, the other as a fully independent external validation set. Using two geographically separate centers helps confirm that the model performs consistently across different ultrasound equipment manufacturers and scanning protocols.
The full DLR-SVM model achieved an area under the ROC curve (AUROC) of 0.893 on the internal validation cohort and 0.871 on the external validation cohort. These values indicate strong discrimination between cancer and non-cancer cases, considerably outperforming endometrial thickness alone, which has an AUROC around 0.65-0.70 in most studies.
A key finding was that SRGAN preprocessing significantly improved performance. Models trained on standard-resolution images performed measurably worse than those trained on super-resolution enhanced images, demonstrating that image quality directly impacts AI diagnostic accuracy and that pre-processing steps are not trivial.
The endometrium-level ROI (focused specifically on the uterine lining) outperformed the broader uterine-corpus-level ROI on both internal and external validation. This suggests that pathological changes specific to EC are concentrated in the endometrial layer, and that zooming in on the relevant anatomy improves signal quality for the classifier.
Ultrasound images often have lower resolution than CT or MRI, and subtle textural changes in the endometrium that indicate early malignancy may fall below standard ultrasound image resolution. SRGAN enhancement recovers some of this detail, giving downstream feature extraction more meaningful information to work with.
The combination of classical radiomics features (which capture hand-crafted texture statistics) with deep learning features (which capture learned patterns the radiomics pipeline might miss) achieves complementary coverage. Neither approach alone performed as well as the combined DLR model, supporting the feature fusion approach.
An important limitation is that the model was developed specifically for postmenopausal women, the highest-risk group for EC. Performance in premenopausal women, who have thicker endometria normally due to menstrual cycling, would likely differ and would require separate validation and potentially different thresholds.
The practical value of this model is its potential to reduce unnecessary biopsies in postmenopausal women with abnormal uterine bleeding. Currently, many women with thickened endometria on ultrasound undergo invasive biopsy, only to receive a benign diagnosis. A more precise AI risk score could concentrate biopsy referrals on higher-risk cases.
Deploying this model in clinical settings would require standardization of ultrasound acquisition protocols to ensure consistent image quality for the SRGAN preprocessing step. Variability in probe frequency, image settings, and operator technique could affect model performance, necessitating site-specific calibration or quality controls.
The two-center validation is a promising step but remains limited in geographic and demographic scope. Larger multicenter trials including diverse populations - with varying rates of EC risk factors like obesity, diabetes, and hormone use - are needed to establish the model's robustness before routine clinical deployment.