CNN Based Image Descriptor for Polycystic Ovarian Morphology from Transvaginal Ultrasound

Abstract

One in ten women of childbearing age experiences the health issue known as polycystic ovarian syndrome (PCOS). Hormonal dysregulation and metabolic issues in PCOS women might impact their general health and attractiveness. Infertility can also be caused by PCOS, which happens when the egg discharged each month as part of a normal menstrual cycle does not mature normally or may not be released during ovulation as it should be if PCOS is present. Hence detection of PCOS in its early stages is essential in many cases to help in ensuring swift treatment procedures. This detection may be tedious, especially if done by doctors and medical professionals using traditional ultrasound image analysis. Hence, automated ultrasound image detection techniques developed using deep learning methods like CNN architecture-based models are quite helpful. Studies in this area have yielded great detection results in recent years. This paper proposes a self-built CNN-based methodology for accurately detecting PCOS by classifying ultrasound ovary images into the PCO and non-PCO categories. The filters of the CNN are associated with the segmentation of the follicles while the fully connected layer of the CNN is responsible for the classification. A brief literature survey encapsulating previous works is also discussed. The findings substantiate our claim that segmentation of follicle blobs aids in isolating non-PCOS images. The CNN then proceeds to function as a confirmation test to classify the PCOS follicles with an accuracy of over 83%. The methodology and results are presented further in this study, and the discussion also involves the future scope & developments that this methodology can be improved.