Comparing Different Models for Polycystic Ovary Syndrome Diagnosis: An Empirical Investigation on a Large Clinical Dataset

Poly-cystic Ovary Syndrome (PCOS) is a lifestyle ailment in women and is distinguished by hyperandrogenism, menstrual irregularities, and the formation of fluid-filled sacs in the ovary. The manual detection of PCOS using ultrasound frames is presently challenging and a tedious task for young gynecologists. Automated artificial intelligence-based pipelines can expedite the detection of PCOS symptoms in ultrasound frames and reduce false positive rates. The present work presents an empirical investigation for automatically detecting PCOS and non-PCOS ultrasound frames using a large clinical dataset. Four Transfer Learning (TL) models namely InceptionResnetV2, EfficientNetV2M, EfficientNetV2B3, and NASNetMobile have been utilized. A comparative analysis of the TL models has been done through various evaluation metrics like accuracy, loss, Area-Under-Curve (AUC) and F1-score, test set analysis, and interpretability plots like gradient-weighted class activation mapping, Shapley additive explanations, and local interpretable model-agnostic explanations. Out of the four TL models, the EfficientNetV2B3 model proved to be the most precise with a training accuracy, validation accuracy, loss, AUC, and F1-score up to 95.89%, 99.41%, 0.00027, 99.60%, and 99.41% respectively on the validation data.