Diagnostic Accuracy of the DEERS Endometrial Imaging Test

Background

Endometrial cancer requires early and accurate diagnosis, but the gold standard, histopathological examination (HPE) of an endometrial biopsy, is invasive and uncomfortable. The DEERS imaging test (Diseases of Endometrium – Evaluation and Risk Scoring) is a novel technology offering real-time visualization and classification of endometrial tissue. This study rigorously evaluates its diagnostic performance against HPE.

Approach

This retrospective analytical study was conducted at a tertiary care hospital in Telangana among 309 women with abnormal uterine bleeding. It compared 309 paired samples, matching DEERS classifications (Normal, Benign, Complex Hyperplasia, Malignancy) against the HPE gold standard. A four-class categorical agreement was assessed via confusion matrix, overall accuracy, and Cohen’s kappa, alongside three clinically relevant binary scenarios (Normal vs Abnormal, Malignancy vs all others, Benign vs malignant/pre-malignant). Analysis used R with tidyverse, yardstick, pROC, and janitor, reporting sensitivity, specificity, kappa, and ROC/AUC.

What we found

• Overall four-class accuracy was 73.5% with moderate agreement (Cohen’s kappa 0.503).
• DEERS identified normal tissue well (sensitivity ~81%) and distinguished normal from abnormal with fair discrimination (AUC 0.774, accuracy 77.7%).
• Sensitivity for malignancy was extremely poor at 9.1%, missing 10 of 11 malignant cases, with several misclassified as complex hyperplasia.
• For benign versus malignant/pre-malignant lesions, AUC was 0.775 but sensitivity only 28.2%, indicating poor categorical separation of higher-risk pathology.

Outputs & impact

The study concludes that DEERS is a reasonable tool for confirming normal endometrium but is unreliable as a standalone cancer-screening test because of a high risk of false negatives for malignancy. The findings caution against using it as a primary diagnostic tool and identify where its categorical thresholds need refinement. Outputs include a reproducible R diagnostic-accuracy pipeline, confusion matrices, and ROC analyses.