Detection of Polycystic Ovary Syndrome Using Clinical Data and Ultrasound Images

Polycystic Ovary Syndrome (PCOS) is a hormonal issue that occurs in women of adulthood. The World Health Organization (WHO) has identified PCOS as a prevalent endocrine illness that affects around 10% of women globally. It can result in various health complications, which includes infertility, metabolic complications such as insulin resistance, obesity, as well as cardiovascular problems, sleep apnea, endometrial cancer, and psychological disorders like anxiety and depression. Hence early diagnosis of PCOS is crucial. One of the diagnosis methods used for its detection is the Rotterdam criteria or Consensus. This diagnostic approach includes three criteria: Oligovulation or anovulation, presence of hyperandrogenism, and the identification of polycystic ovaries through ultrasound examination. Patients who meet two or more of these criteria can be diagnosed with PCOS. Cysts may indicate Polycystic Ovarian Disease (PCOD), a condition similar to PCOS. In PCOD, the ovaries release numerous immature or partially-mature eggs, which can develop into cysts over time. Among the numerous available techniques in the machine learning domain, only one criterion is typically assessed at a time, either clinical data or ultrasound, but not both simultaneously. The proposed system considers both and has two sections to aid in this process—one for detection through images and the other through clinical data. The dataset for the system includes 781 PCOS and 1143 Non-PCOS images, as well as clinical data from 541 patients, including 177 with PCOS and 43 features collected from open sources. Several models and techniques are used for the detection individually. A novel feature selection approach for CS-PCOS is employed, utilizing an optimized chi-squared mechanism. Additionally, overfitting is assessed using ten-fold cross-validation. Different pre-trained models are tried out for ultrasound images and the best is taken. Random forest is considered the best model for clinical data with an accuracy of 90.72% while Inception V3 is considered the best model with 99.64% for images. The both divisions yield high-accuracy results separately and can reduce false positives together. Both the outputs are compared and it is predicted whether the person has PCOS, has more likelihood for PCOS/PCOD or does not have PCOS. The system’s output can speed up diagnosis and aid in early detection of PCOS.