Gynecological ultrasound is the first-line, non-invasive tool for detecting endometrial cancer in postmenopausal women presenting with abnormal bleeding. However, effective ultrasound diagnosis requires significant skill and experience. In many primary care settings - particularly in economically disadvantaged regions - physicians lack the training to reliably interpret gynecological ultrasound findings, creating diagnostic gaps that lead to delayed referrals and later-stage diagnoses.
The standard clinical parameter used to screen for endometrial pathology is endometrial thickness: a measurement above 5mm in postmenopausal women is considered suspicious. But this single measurement has important limitations - it misses many cancers at lower thresholds and produces many false positives at higher thresholds, driving unnecessary biopsies. Its diagnostic AUC in this study's dataset was 0.734 - barely adequate.
This study aims to develop a deep learning model based on YOLOv8 - a state-of-the-art object detection framework - that can analyze ultrasound images more comprehensively than endometrial thickness alone. The goal is to create a tool that functions like a skilled ultrasound specialist, accessible to primary care physicians regardless of their own ultrasound experience level.
The study enrolled 877 postmenopausal women with histologically confirmed endometrial cancer from a single Chinese hospital (2020-2024). Their ultrasound images were divided 70:20:10 into training (614), validation (175), and testing (88) sets. Expert ultrasound specialists annotated tumor regions in each image with bounding rectangles, providing ground truth for the object detection model.
The core framework was YOLOv8 (You Only Look Once, version 8), a real-time object detection architecture that simultaneously localizes tumors in an image and classifies them as benign or malignant in a single forward pass. The team made two important architectural modifications. First, deformable convolutional layers were added - unlike standard convolutions with fixed sampling grids, deformable convolutions learn to shift their sampling positions to better fit the irregular, non-rigid shapes of tumors in ultrasound images.
Second, a custom dynamic prototype loss function was added to the training objective. Standard training loss can struggle to separate tumor tissue from anatomically similar structures like uterine muscle and benign endometrial thickening. The dynamic prototype loss forces the model to learn tighter clustering of tumor features while maximizing separation from non-tumor classes - directly addressing the most common source of false positives in ultrasound-based EC detection.
The modified YOLOv8 model achieved consistent performance across training (AUC 0.844), validation (AUC 0.811), and testing (AUC 0.858) sets - with no statistically significant differences between sets, demonstrating genuine generalization rather than overfitting. The test set AUC confidence interval of 0.800-0.905 confirms reliable performance.
Most importantly, the DL model significantly outperformed the standard clinical benchmark of endometrial thickness: AUC 0.858 vs. 0.734 (p=0.004). Additional comparisons confirmed superiority across all diagnostic metrics - accuracy (85.8% vs. 75.0%), sensitivity (77.3% vs. 66.7%), specificity (90.1% vs. 78.6%), and positive predictive value (83.6% vs. 57.1%). The only metric where the models performed similarly was negative predictive value (86.7% vs. 84.6%).
An ablation study validated that both architectural modifications contributed independently. Baseline YOLOv8 achieved AP@0.5 of 74.7%. Adding deformable convolutions alone improved this to 76.2%. Adding the dynamic prototype loss alone improved it to 78.6%. Combining both components achieved the best performance at AP@0.5 of 82.0% - a synergistic improvement demonstrating that each component addresses a distinct and complementary limitation of the baseline model.
The model learns to integrate the same diagnostic features that expert sonographers use. These include: endometrial thickness (thickness above 1.49 cm suggests malignancy), echo heterogeneity (irregular, non-uniform internal patterns within the endometrium), and blood flow characteristics assessed by Doppler imaging - including the blood flow area (BFA), resistance index (RI), end-diastolic velocity (EDV), and peak systolic velocity (PSV).
Particularly important is the resistance index (RI) from Doppler imaging: an RI below 0.4 indicates neoangiogenesis - the formation of new blood vessels that tumors create to sustain rapid growth. This vascular signature is a key distinguishing feature of malignant endometrial tissue. Unlike simple endometrial thickness, incorporating blood flow features allows the AI to detect functional changes in tumor physiology, not just anatomical size.
Grad-CAM (Gradient-weighted Class Activation Mapping) heatmaps were used to verify that the model was attending to clinically meaningful regions. Overlay visualizations confirmed that the AI focused on the endometrial lining and abnormal echogenic regions - the same areas a trained sonographer would examine - rather than irrelevant background structures.
The study's most compelling clinical implication is its target deployment context: primary care settings where expert sonographers are unavailable. The AI system could function as an always-available second opinion - analyzing ultrasound images captured even by less-trained operators and flagging cases that require urgent specialist referral.
The authors envision the AI serving two complementary clinical roles: first, as a real-time diagnostic support tool that suggests interpretations during the scan, helping less experienced physicians identify suspicious findings they might otherwise miss; and second, as a training tool that demonstrates expert-level analysis on diverse cases, accelerating the learning curve for novice ultrasound physicians in resource-limited settings.
The primary study limitation is that the dataset contains only confirmed endometrial cancer cases - no benign controls were included. This means the current model cannot differentiate EC from benign endometrial conditions in a general screening population. Future work should include benign endometrial hyperplasia, polyps, and normal postmenopausal endometrium to develop a true screening tool.