AI-powered diagnosis of ovarian conditions: insights from a newly introduced ultrasound dataset.

Introduction: Ovarian diseases, including Polycystic Ovary Syndrome (PCO) and Dominant Follicle irregularities, present significant diagnostic challenges in clinical practice. Traditional diagnostic methods, reliant on subjective ultrasound interpretation, often lead to variability in accuracy. Recent advancements in artificial intelligence (AI) and transfer learning offer promising opportunities to improve diagnostic consistency and accuracy in ovarian disease detection. Methods: We introduced a new, publicly available dataset of ultrasound images representing three ovarian conditions: Normal, PCO, and Dominant Follicle. Using transfer learning, we applied four CNN models-AlexNet, DenseNet121, ResNet18, and ResNet34-to evaluate their performance in multiclass classification of these conditions. The models were assessed using macro and micro metrics, including accuracy, F1 score, precision, and recall, to determine their effectiveness in classifying ovarian conditions. Results: The results showed that ResNet18 demonstrated the highest performance across all metrics, particularly excelling in the classification of Normal and PCOS conditions. ResNet18 achieved the best performance, with an accuracy of 76.2% and a macro F1-score of 78.2%, demonstrating its effectiveness in distinguishing ovarian conditions. AlexNet also delivered strong results, achieving near-perfect precision in PCOS classification. However, DenseNet121 showed less competitive performance in classifying Dominant Follicle, although it still benefited from transfer learning. The overall results suggest that transfer learning enhances the classification accuracy of CNN models in ovarian disease diagnosis. Discussion: The application of transfer learning in this study significantly improved the performance of CNN models, especially for Normal and PCOS classifications. The introduction of a publicly available dataset serves as an important contribution to the field, facilitating further research in AI-driven diagnostics. These findings highlight the potential of AI to revolutionize ovarian disease diagnosis by providing more reliable and accurate results, reinforcing the importance of AI in early detection and diagnosis. Conclusion: This study demonstrates the significant potential of CNN models, enhanced by transfer learning, in improving the diagnostic accuracy of ovarian conditions. The publicly available dataset introduced here will serve as a valuable resource for future research, advancing AI-based medical diagnosis. Further work on refining model architectures and applying these methods in clinical practice is necessary to ensure their reliability and broader applicability.