Integrating deep learning and clinical characteristics for early prediction of endometrial cancer using multimodal...

Endometrial cancer (EC) is one of the most prevalent malignancies affecting the female reproductive system. It poses significant health risks to women and imposes a substantial economic burden on healthcare systems. Early and accurate diagnosis is critical for improving patient outcomes. While traditional diagnostic methods rely on clinical evaluation and imaging, there is growing interest in leveraging artificial intelligence, particularly deep learning (DL), to enhance diagnostic accuracy.

This study developed a DL-based predictive model integrating multimodal ultrasound features and clinical risk factors to improve early EC diagnosis. A retrospective, multicenter analysis was conducted using 1,443 multimodal ultrasound images-including two-dimensional (2D) and color Doppler images-from 611 patients, of whom 132 were diagnosed with EC and 479 were non-EC cases. Clinical risk factors such as body mass index (BMI), menopausal status, irregular vaginal bleeding, and hypertension were identified as significant predictors (P < 0.05) and incorporated into a clinical model. Separate DL models were trained on 2D and color Doppler ultrasound images, and their performance was evaluated individually and in combination with the clinical model.

The area under the receiver operating characteristic curve (AUC) for the clinical model was 0.772 (95% CI: 0.690-0.854). The 2D and color Doppler DL models achieved AUCs of 0.792 (95% CI: 0.719-0.864) and 0.813 (95% CI: 0.745-0.881), respectively. When combined with the clinical model, the merged model demonstrated superior predictive performance. In the external validation cohort, the merged model achieved an AUC of 0.892 (95% CI: 0.846-0.938), indicating high diagnostic accuracy.

The integration of multimodal ultrasound imaging and clinical risk factors using DL significantly enhances the accuracy of endometrial cancer diagnosis. The merged model demonstrated strong generalizability in external validation, underscoring its potential clinical utility. Future studies should focus on larger, prospective multicenter trials to further validate these findings and explore the implementation of this approach in personalized patient care.