Applications of Machine Learning and Deep Learning in Benign Gynecology: A Scoping Review

INTRODUCTION: Artificial intelligence (AI) has been widely adopted into medical practice since the 1950s. Initially designed as non-adaptive rule-based expert systems, AI has advanced in recent decades to self-optimizing models trained to learn and predict patterns with increasing accuracy over time. These new machine and deep learning methods have been increasingly used in women’s healthcare subspecialties, including obstetrics, reproductive endocrinology, and obstetric ultrasound. However, its use in gynecology only constitutes a small portion of published literature. Although tens of millions of women receive general gynecologic care annually, gynecologic oncology and reproductive medicine are overrepresented in these studies. Existing systematic reviews on AI in this field are narrow in scope, restricted to the diagnosis of polycystic ovarian syndrome (PCOS) and sonographic assessment of adnexal masses. By characterizing the current use of data-driven predictive models in the broader field of benign gynecology, we can better identify future areas for research opportunities. OBJECTIVE: To summarize the current applications of machine and deep learning in benign gynecology. METHODS: PubMed, Embase, Scopus, and Web of Science databases were broadly searched for any publication related to AI in gynecology from their inception until August 2024. Publications were included if they contained machine or deep learning models, and excluded if they contained only rule-based systems. Animal studies, secondary research articles, and studies specific to pregnant populations, gynecologic oncology, breast conditions, and reproductive endocrinology were excluded. All relevant publications were screened by two independent reviewers with conflicts resolved by a third reviewer. Full-text extraction was performed by one reviewer. RESULTS: Our search across the four databases identified 10,687 studies. Of these, 1,641 full-text studies were assessed for eligibility, and 984 studies were included. Most frequently represented were cervical dysplasia (n=264, 27.0%) and PCOS (n=156, 15.9%), followed by endometriosis (n=84, 8.5%), uterine smooth muscle tumors (n=68, 6.9%), adnexal masses (n=60, 6.1%), and voiding dysfunction (n=49, 5.0%); also included were endometrial pathologies, gynecologic infections, and menstrual-related disorders, among others (all <5%) (Figure 1). Approximately half of the studies utilized deep-learning (n=440, 44.7%) compared to machine-learning (n=554, 55.3%) models. The primary application of intelligent models was for screening and diagnosis (n=484, 49.2%), although there were broad but less commonly included implementations, including biomarker identification, disease prognosis/treatment outcomes, and risk stratification, among others (Figure 2). CONCLUSIONS: Data-driven predictive models in benign gynecology are commonly used for screening/diagnosis of cervical dysplasia and polycystic ovarian syndrome, but also demonstrate potential for use in biomarker identification, treatment outcomes, and risk stratification for various gynecologic diseases. However, while promising, many of these models have not yet been adopted in practice. Future research is needed to optimize existing intelligent systems for healthcare implementation and to expand their use to topics that are currently underexplored in gynecologic care.Figure 1Figure 2