Authors: Mrs. Rajeswari M, Dr. Monika G
Abstract: Ultrasound is the first-line imaging modality for assessing the ovary, both for monitoring follicles in fertility care and for characterising adnexal and ovarian tumours. Interpretation, however, depends on operator skill and is affected by speckle noise, low contrast, and ambiguous boundaries. Deep learning, and in particular the U-Net family of encoder-decoder networks, has become the dominant approach for automating segmentation of the ovary, follicles, and masses. This paper reviews the literature on automated follicular and ovarian tumour analysis from ultrasound using U-Net variants. It summarises the clinical background, the architectural variants (plain, attention, nested, self-configuring, recurrent, and transformer-hybrid U-Nets), their application to follicle segmentation and adnexal mass segmentation, and the downstream tasks of quantification and classification. The reviewed studies report high overlap with expert outlines, for example Dice coefficients of 0.912 for the ovary and 0.858 for follicles with a composite U-Net, and 0.91 for adnexal masses using a bounding-box-guided U-Net. Evidence nevertheless remains constrained by small, single-centre, and retrospective datasets, limited external validation, device variability, and the scarcity of Indian data. The paper presents a conceptual analysis pipeline, compares key studies, and outlines research directions toward robust, clinically integrated tools.
DOI: https://doi.org/10.5281/zenodo.23411233
