Deep learning-based automated classification of endometrial lesions in IVF patients using hysteroscopic images.

Endometrial health is a key determinant of female fertility and successful pregnancy outcomes, making the accurate diagnosis of endometrial lesions essential for the success of assisted reproductive technology. While hysteroscopy remains the gold standard for uterine cavity evaluation, interpretation can vary based on clinical expertise. To address this, we developed a deep learning-based clinical decision support system to classify hysteroscopic images from high-resolution (4 K) videos into three categories: normal endometrium, endometrial polyps, and endometritis. Endometritis cases were classified based on hysteroscopic features suggestive of inflammation; no histopathological confirmation was obtained. Utilizing a dataset of 1500 expert-annotated images from 200 clinical videos, we applied transfer learning across four architectures: VGG-16, VGG-19, DenseNet-121, and EfficientNet-B0. Our results show that the models achieved classification accuracies between 85 and 89%, with DenseNet-121 demonstrating superior performance, specifically achieving a sensitivity of 93% and an AUC of 98.8% for polyp detection, alongside a precision of 90% for endometritis. Furthermore, Grad-CAM visualization confirmed that the networks focused on clinically relevant morphological features, enhancing model interpretability. These findings suggest that deep learning may serve as a supportive tool to assist clinicians in hysteroscopic analysis, pending validation on external datasets and with pathologically confirmed labels.