Automated Classification of Endometrial Pathologies Using Artificial Intelligence

Hysteroscopy provides direct visualization of the uterine cavity; however, diagnostic outcomes remain highly operator-dependent and subjective. This study aimed to develop an automated deep learning framework based on the ConvNeXt-XLarge architecture, in comparison with an EfficientNetV2 baseline, to objectively classify common uterine pathologies and thereby reduce diagnostic variability. A prospective dataset of 1262 histopathologically confirmed hysteroscopic images was analyzed using an 80:20 training to validation split and a two-stage transfer learning strategy. In addition to the primary multiclass framework, three binary classification models were developed to evaluate pairwise differentiation between the pathologies. The multiclass model achieved an accuracy of 80.5%, a precision of 82.3%, a weighted Area Under the Receiver Operating Characteristic Curve (AUC) of 94.8%, and a weighted average precision (AP) of 91.6%. In the binary classification tasks, model 1 (fibroids vs. functional polyps) achieved an accuracy of 96.7% and an AUC of 99.8%; model 2 (fibroids vs. hyperplastic polyps) achieved 87.2% accuracy and an AUC of 87.4%; and model 3 (functional vs. hyperplastic polyps) achieved 92.2% accuracy and an AUC of 99.0%. ConvNeXt-XLarge statistically significantly outperformed EfficientNetV2 in the multiclass task in terms of accuracy and AUC, and in model 2 in terms of accuracy, underscoring the added value of the deeper backbone for this hysteroscopic image classification. Overall, these findings demonstrate that deep learning frameworks can accurately automate the classification of hysteroscopic findings. By providing high-precision analysis, these models offer a robust tool to mitigate diagnostic subjectivity and optimize clinical decision-making.