Polycystic Ovary Syndrome (PCOS) Detection: Using Deep Learning Approaches for Images and Clinical Features

Abstract

The hormone disorder known as polycystic ovarian syndrome (PCOS) is most prevalent in women nowadays. The ovaries of women are directly impacted by this metabolic disorder, where numerous small sacs of fluid develop around the edges of the ovary, called cysts or follicles. Despite advancements in technology, the exact cause of PCOS still remains unknown. It takes a lot of work for the doctors to manually diagnose it from the ultrasound (US) images. For PCOS identification, a range of machine learning (ML), deep learning (DL) and image processing techniques were employed, that allowed for the identification of follicle counts and follicle size analysis. The current research on PCOS detection for image segmentation and classification, as well as clinical characteristics, was reviewed in this work. For this study, we utilised two different kinds of dataset: image data and clinical feature dataset. We proposed a transfer learning-based DL approach for classifying women with PCOS by using ultrasound images. InceptionV3 and ResNet50 models were used, which achieved accuracy rates of 99.68% and 97.5%, respectively. Numerous DL algorithms were applied on the clinical dataset using various feature selection techniques. The flower pollination algorithm (FPA), cuckoo search algorithm (CSA), and genetic algorithm (GA) are the nature-inspired feature selection algorithms that have been utilised in the study. The results of the proposed GA-based feature selection using an ensemble model were statistically tested with the existing algorithms using Analysis of variance (ANOVA) and the Tukey test, with a 94.58% accuracy, 97.48% precision, 83.76% recall, 90.08% F1-Score, and 99.09% specificity. The interpretability of the results is further ascertained by explainable artificial intelligence (XAI) techniques, i.e., LIME and SHAP. Overall, by using reliable and accurate prediction models, this work advances predictive analytics in PCOS diagnosis with the goal of assisting early diagnosis and well-informed decision-making processes.