TongueNet-GYN: a multimodal deep learning framework for non-invasive gynecological disease screening in digital public health

Background: Gynecological diseases, such as polycystic ovary syndrome (PCOS) and endometriosis, are prevalent global health concerns. Conventional diagnostics often rely on invasive procedures or costly imaging, limiting accessibility in resource-constrained settings. This study proposes TongueNet-GYN, a novel, non-invasive screening framework that leverages tongue image analysis integrated with modern AI. Methods: We compiled a dataset of 3,167 tongue images. To address class imbalance, a hybrid strategy combining Borderline-SMOTE and clinically constrained data augmentation was employed. The framework integrates structured clinical priors with deep semantic features extracted via an enhanced Attention-CLIP model. Additionally, quantified morphological features were incorporated to mirror clinical diagnostic logic. Results: TongueNet-GYN was evaluated using a robust framework comprising 5-fold cross-validation on a discovery set (85%) and subsequent validation on an independent held-out test set (15%). The model achieved a high diagnostic Accuracy of 90.14% and an AUC of 89.74% on the unseen test data. Furthermore, the integration of patient age was identified as a critical factor, yielding measurable improvements in both diagnostic accuracy and framework robustness. Conclusion: These results demonstrate that TongueNet-GYN provides a precise, efficient, and scalable digital health solution, offering potential for improving early screening and health equity in women's chronic disease management.