Polycystic Ovary Syndrome (PCOS) is one of the most prevalent endocrine and metabolic diseases affecting women of reproductive age and is a significant contributor to reproductive dysfunction, metabolic dysfunction, and long-term cardiovascular disease risk. Standard diagnostic pathways are typically driven by clinical, biochemical, and imaging criteria e.g., Rotterdam, National Institutes of Health (NIH) all of which are limited by inter-observer variability, delayed diagnosis, and limited sensitivity in assessing different PCOS phenotypes. The study covered has included defined clinical data focused on hormone and metabolic profiling, and imaging data, specifically, ultrasound and magnetic resonance imaging. The study architected a number of models with numbers of layers/dimensions, as well as convolutional neural networks, recurrent networks, autoencoders, and transformers. The review covered a comparative study of the performance across several of the deep learning studies, as well as questions related to model strengths and limitations and study trends. Significant barriers, including small datasets, imbalanced data, interpretability, and ethical issues, have been thoroughly examined. By synthesizing existing literature, this work has illustrated the feasibility of utilizing deep learning to revolutionize PCOS diagnosis while identifying future avenues of research such as explainable-AI, federated learning, and multimodal fusion together towards clinical implementation.