Deep learning-based automated detection of endometrioid endometrial carcinoma in histopathology

Background: Rapid advances in artificial intelligence (AI) have enabled automated tumor identification. To overcome challenges in traditional pathology, including complex sampling and limited physician resources, accessible tools for automated diagnosis are urgently needed. Methods: We developed a deep learning system based on an improved ResNet-18 to automatically identify endometrioid endometrial carcinoma (EEC) from H&E-stained endometrial hyperplastic lesions and normal tissues. Results: The model demonstrated strong performance in detecting endometrioid endometrial carcinoma. The positive predictive value (PPV), defined as the proportion of true disease cases among all positive diagnostic results, reached 95.13%, and the F1-score, defined as the harmonic mean of precision and recall, reached 0.95. The model achieved a PPV of 87.15% and an F1-score of 0.87 for typical hyperplasia, as well as a PPV of 79.88% and an F1-score of 0.74 for atypical hyperplasia, both meeting clinically acceptable thresholds. For normal endometrial physiological states, the PPVs were 91.75% (proliferative phase), 80.94% (secretory phase), and 80.88% (menopausal phase). Conclusion: This multi-task deep learning system provides stable and efficient support for automated EEC identification and effectively classifies endometrial pathological and physiological states, demonstrating strong potential for clinical translation.