Non-Invasive and AI-Enabled Diagnostic Approaches for Polycystic Ovary Syndrome: A PRISMA-Guided Narrative Review Authors: Aman Golani, Soumi Ray Published: 2026 ABSTRACT BackgroundPolycystic ovary syndrome (PCOS) is one of the most common endocrine disorders in women of reproductive age and is associated with infertility, obesity, insulin resistance, and elevated long-term to cardiovascular risk. Traditional diagnostics rely on biochemical examinations and pelvic ultrasonography imaging. These are resource-intensive, operator-dependent, and frequently inaccessible in low-resource settings. Given the pronounced heterogeneity of PCOS phenotypes, traditional single-modality diagnostics frequently fail, necessitating a shift toward multimodal diagnostic approaches.ObjectiveTo critically evaluate the diagnostic efficacy and clinical applicability of non-invasive and minimally invasive diagnostic approaches for PCOS, encompassing artificial intelligence (AI) and machine learning (ML) methods, ultrasonographic imaging, hormonal profiling, clinical and biochemical markers, and insulin resistance indicators.MethodsA structured narrative review was conducted using Publish or Perish (Version 8.17.4863.9118) across PubMed and Google Scholar. Studies published between 2014 and 2025 in English-language peer-reviewed journals were considered eligible if they focused on PCOS detection or diagnosis and reported at least one non-invasive or minimally invasive diagnostic modality. After deduplication, 62 papers were selected from 1,110 retrieved. A PRISMA flow diagram is included to document the selection process transparently; this review does not claim formal systematic review compliance.ResultsAI-based diagnostic models have reported high accuracy in experimental settings for both image-based and data-driven PCOS detection. While specific architectures exceed 90% accuracy offering viable pathways for low-resource screening, real-world deployment remains constrained by a lack of dataset standardization. Automatic segmentation of ultrasound images has the potential to reduce inter-operator variability in antral follicle counting. Hormonal profiling, centred on LH/FSH ratio, anti-Müllerian hormone, and androgen panels, remains foundational, while emerging molecular, metabolomic, and metabolic biomarkers offer complementary phenotypic precision.ConclusionThe evidence projects a diagnostic shift toward reproducible, multimodal, and phenotype-sensitive PCOS detection. Integration of AI-based pipelines when integrated with imaging morphology, hormonal data, and metabolic markers offers significant opportunities for equitable, early-stage identification across diverse healthcare settings. These findings support the development of accessible, scalable screening tools to reduce diagnostic delay and improve women’s reproductive and metabolic health outcomes globally, though standardised external validation and equity-centered design remain critical priorities. Introduction Polycystic ovary syndrome (PCOS) is one of the most common endocrine disorders in women of reproductive age. It affects hormonal regulation, metabolic function, and fertility. The global burden of PCOS has increased considerably over time. Incident cases increase from roughly 1.48 million in 1990 to 2.3 million by 2021. Regions with a higher socio-demographic index (SDI) report a greater burden mainly because of greater diagnostic awareness and healthcare access within reach1,2 on the other hand, underdiagnosis remains a serious concern in low-SDI regions, where limited healthcare infrastructure increases existing health gaps. Figures 1a and b illustrate these regional trends visually. Globally, PCOS is estimated to affect 6–13% of women of reproductive age, yet around 70% of those affected are never diagnosed.3 Figure 1Global burden of PCOS based on GBD 2021 data. (a) Percentage change in age-standardised rate (ASR) per 100,000 across regions. (b) Age-standardised prevalence rate (ASPR) per 100,000 stratified by Social Demographic Index (SDI). Data derived from Wang Y et al.The image A showing a horizontal bar chart titled, PCOS burden change by region (GBD 2021). The x axis label is, Percent change in age-standardized prevalence rate (ASPR), 1990 to 2021, with range 0 to 80 percent and ticks at 0, 20, 40, 60, 80. The y axis lists regions: Southeast Asia, East Asia, South Asia, Southern Latin America, Central Sub-Saharan Africa, Western Sub-Saharan Africa, Central Asia, Oceania, Andean Latin America, North Africa and Middle East, Eastern Sub-Saharan Africa, Eastern Europe, Southern Sub-Saharan Africa, High-income North America, Caribbean, Central Europe, Australasia, Western Europe, High-income Asia Pacific, Central Latin America, Tropical Latin America. Bar endpoints are approximately: Southeast Asia 87 percent; East Asia 83 percent; South Asia 77 percent; Southern Latin America 60 percent; Central Sub-Saharan Africa 51 percent; Western Sub-Saharan Africa 45 percent; Central Asia 38 percent; Oceania 38 percent; Andean Latin America 37 percent; North Africa and Middle East 34 percent; Eastern Sub-Saharan Africa 29 percent; Eastern Europe 29 percent; Southern Sub-Saharan Africa 26 percent; High-income North America 25 percent; Caribbean 25 percent; Central Europe 23 percent; Australasia 16 percent; Western Europe 11 percent; High-income Asia Pacific 10 percent; Central Latin America 10 percent; Tropical Latin America 10 percent. The image B showing a grouped vertical bar chart titled, PCOS burden by SDI (1990 vs 2021), with legend entries 1990 and 2021. The x axis categories are Global, High SDI, Middle SDI, Low-middle SDI, Low SDI. The y axis label is, Age-standardized prevalence rate (per 100,000), with range 0 to 3500. Values are approximately: Global 1990 equals 1350 and 2021 equals 1750; High SDI 1990 equals 3000 and 2021 equals 3550; Middle SDI 1990 equals 1180 and 2021 equals 1980; Low-middle SDI 1990 equals 750 and 2021 equals 1180; Low SDI 1990 equals 500 and 2021 equals 700.Two bar charts showing polycystic ovary syndrome burden change by region and by social demographic index. Global burden of PCOS based on GBD 2021 data. (a) Percentage change in age-standardised rate (ASR) per 100,000 across regions. (b) Age-standardised prevalence rate (ASPR) per 100,000 stratified by Social Demographic Index (SDI). Data derived from Wang Y et al. PCOS is also a heavy burden on health eco-systems. The disease burden linked to metabolic, cardiovascular, and psychological complications is substantial and continues to grow.4 Standard diagnostic methods rely on biochemical testing and pelvic imaging. These approaches are mostly expensive, operator-dependent, and produce variable results, especially where specialist services are scarce. Diagnostic delay is also common. Many affected women attribute symptoms like irregular periods, acne, and excess hair growth to normal variation. This leads to disease that goes unrecognised or undertreated. Moreover, due to the pronounced heterogeneity of PCOS phenotypes, traditional single-modality diagnostics frequently fail, necessitating a shift toward multimodal diagnostic approaches. Artificial intelligence (AI), including machine learning (ML) and deep learning (DL), has emerged as a practical tool for improving PCOS diagnosis. AI models can process hormonal profiles, ultrasound images, and clinical data to support earlier and more accurate detection. Screening tools based on scleral imaging have already shown diagnostic performance that exceeds conventional approaches.5 Publication trends across these diagnostic modalities reflect the field’s fragmented development. Annual PubMed records from 2000 to 2025, retrieved using modality-specific search strings, show three distinct phases (Figure 2). Figure 2a shows that before 2003, all modalities remained at near-baseline activity. From 2003 to 2012, insulin resistance and hormonal research co-dominated, shaped by the Rotterdam diagnostic criteria. The Rotterdam criteria were introduced in 2003 and open the scope of PCOS diagnosis beyond hyperandrogenism by including polycystic ovarian morphology and ovulatory dysfunction. It in turn increased the number of research on the hormonal and metabolic features of the syndrome. Figure 2Annual PubMed publication trends across major PCOS diagnostic modalities. (a) Observed publication trends from 2000 to 2025 and projected trends from 2026 to 2033. (b) Enlarged view of publication trends from 2015 to 2025 with projections to 2033. Solid lines represent observed data; dashed lines indicate projections; shaded regions denote uncertainty. The vertical dashed line marks the transition from observed to projected data..The image A showing a multi series line graph with a legend listing, Ultrasound, Hormonal, Insulin resistance, ML slash AI and Genetic. The x axis label is 2000 to 2033 in years. The y axis label is Annual publications with range 0 to 700. A vertical dashed line at 2025 separates observed and projected. Observed lines rise gradually from near 0 in 2000 to roughly 40 to 140 by 2025 depending on modality. Projected dashed lines extend from 2026 to 2033 with shaded uncertainty bands widening toward 2033. The image B showing a multi series line graph with the same legend categories. The x axis label is 2015 to 2030 in years. The y axis label is Annual publications with range 0 to 700. A vertical dashed line at 2025 separates observed and projected. Observed values from 2015 to 2025 increase from roughly 10 to 70 up to roughly 40 to 140 across modalities. Projected dashed lines from 2026 to 2030 continue upward with widening shaded uncertainty bands toward 2030.Two line graphs showing annual PubMed publications and projections across PCOS diagnostic modalities. Annual PubMed publication trends across major PCOS diagnostic modalities. (a) Observed publication trends from 2000 to 2025 and projected trends from 2026 to 2033. (b) Enlarged view of publication trends from 2015 to 2025 with projections to 2033. Solid lines represent observed data; dashed lines indicate projections; shaded regions denote uncertainty. The vertical dashed line marks the transition from observed to projected data.. From 2012 onward, established modalities accelerated together, while ML/AI remained negligible until 2019 before showing the steepest growth of any category. Speculative projections up to 2033 were generated using ARIMA modelling for established modalities. Linear regression was applied to ML/AI systems due to their relatively short observable history, suggesting that these diverging trends are likely to persist. The widening forecast bands, more clearly shown in Figure 2b, reflect genuine uncertainty at longer horizons, which is expected rather than indicative of modelling instability. Together, the observed and projected trends affirm the relevance of an integrated review at this stage and highlight the growing need to understand the direction of high-quality research across diagnostic modalities. While multiple reviews have provided highly valuable insights into specialized domains of PCOS diagnosis, these foundational efforts remain intentionally focused on specific areas. Consequently, there remains a clear opportunity for a clinically integrated, modality-spanning synthesis which is aligned with current diagnostic practice. Deswal et al provided a comprehensive review of circulating microRNAs as diagnostic markers. They had a specific focus on molecular biomarkers rather than imaging or AI-based tools.6 Studies conducted by Tiwari et al and Bharati et al evaluated ML classifiers, where integration of hormonal, metabolic, or imaging data could be considered.7,8 Javid-Naderi et al examined nanotechnology-based strategies comparison of which with conventional or AI-based approaches could bring more inslights.9 Fahs et al addressed disease mechanisms rather than a comparative evaluation of diagnostic tools.10 Jacob and Cedars reviewed clinical diagnostic guidelines where incorporating emerging modalities such as AI, molecular biomarkers, or multimodal frameworks could broaden the scope.11 This review addresses that gap by providing a unified, clinically oriented synthesis of non-invasive PCOS diagnostic methods. It evaluates their comparative roles, translational readiness, and use across different healthcare settings. It integrates AI-based methods with established clinical, imaging, and biochemical approaches to reflect real-world diagnostic pathways. Review Methodology This review aimed to evaluate current diagnostic techniques for PCOS. The scope covered clinical, imaging, biochemical, and AI-based approaches. Search Strategy and Inclusion Criteria A systematic and well-documented literature search was conducted using Publish or Perish (Version 8.17.4863.9118; Harzing, 2024).12 This citation analysis tool retrieves indexed records primarily from PubMed and Google Scholar. Google Scholar was included to capture relevant articles which are not indexed in PubMed. Two complementary search strategies were employed across these two databases. For PubMed, search strategies were developed using a combination of controlled vocabulary (Medical Subject Headings, MeSH) and free-text terms. These were restricted to titles and abstracts ([tiab]). Also the search was organized into domain-specific query blocks. These blocks covered key aspects of PCOS detection, including clinical, biochemical, imaging, metabolic, genetic, and artificial intelligence–based approaches. All retrieved records were manually verified for bibliographic accuracy. Google Scholar does not support MeSH terms or field-tag qualifiers. To ensure proper retrieval of relevant literature, Boolean keyword combinations were applied in the free-text search strategy. The search string used was: (“PCOS” OR “Polycystic Ovary Syndrome”) AND (“detection” OR “PCOS detection” OR “AI-based detection” OR “non-invasive”) AND (“hormonal markers” OR “ultrasound” OR “insulin resistance” OR “Personalized Healthcare”). This approach was designed to capture a wide range of relevant literature across the same diagnostic domains. It compensates for the absence of controlled vocabulary support. Studies were considered eligible for inclusion if they satisfied all of the following criteria: (i) published within 2014 and 2025 time-window; (ii) published in English in peer-reviewed journals or conference proceedings; (iii) primarily focused on the detection or diagnosis of PCOS rather than only therapeutic management; and (iv) reported at least one non-invasive or minimally invasive diagnostic modality. These included ultrasonography, hormonal or biochemical profiling, clinical assessment, genetic or molecular biomarker identification, or artificial intelligence/machine learning–based predictive models. Studies involving human participants, irrespective of age group, were included. The final literature search was conducted in March 2025 and was not updated prior to submission. This cut-off was maintained intentionally to preserve consistency and reproducibility of the dataset used for analysis. Inclusion of newly emerging studies during manuscript preparation could introduce variability. The selected time window remains sufficient to capture recent diagnostic developments, particularly in AI-based and omics-based approaches. Exclusion Criteria All duplicate records were removed before screening. Studies lacking sufficient bibliographic information (eg, absence of an abstract) were excluded to enable initial eligibility assessment. Articles not included if full-text access could not be obtained through institutional or alternative academic sources. Articles focused exclusively on therapeutic interventions or management strategies without addressing diagnostic or detection-related outcomes were not considered in this review. Studies investigating PCOS-associated complications, including metabolic, cardiovascular, or psychological sequelae, were also excluded unless they prominently contributed to diagnostic characterization or early detection frameworks. In the next step, the studies employing diagnostic modalities solely for confirmatory or descriptive purposes without quantitative analysis, feature extraction, or predictive modelling were excluded from modality-specific synthesis. For example, pharmacological or interventional studies (eg, those evaluating dapagliflozin or bariatric surgery outcomes) that utilized transvaginal ultrasonography only to confirm polycystic ovarian morphology (PCOM), without performing image-based analysis, were not considered within the ultrasound diagnostic domain and were instead contextualized under biochemical or therapeutic evidence where relevant. Screening and Study Selection The initial search retrieved 1,110 records where 986 articles were from Google Scholar and 124 from PubMed. We found 57 duplicate records and one non-English article which were all removed. The 1,052 unique records were retained for title and abstract screening. 990 records were excluded based on predefined eligibility criteria like lack of relevance to PCOS diagnosis or detection, absence of accessible full text, or non-PCOS focus. A total of 62 studies were included in this review. Other articles listed in references are used for different supportive purposes. These studies were selected based on their relevance to key areas of PCOS diagnosis like epidemiology, clinical frameworks, imaging and AI-based methods, hormonal and biochemical markers, insulin resistance, molecular studies, and emerging diagnostic approaches. Details of all included studies, such as study design, population, diagnostic focus, methods, and key findings, are provided in the Supplementary Table 1 which is supplied separately. This review adopts a structured narrative approach. It aims to provide conceptually comprehensive coverage of diagnostic modalities rather than comprehensive record of all available studies. Screening and study selection were conducted by a single reviewer using standardised inclusion criteria and predefined domain classification to minimise bias and ensure consistency. The overall search and selection process is summarized in Figure 3.13 Figure 3PRISMA 2020 flow diagram illustrating the literature search, screening, and selection process for this structured narrative review.A rectangular, box-and-arrow flowchart summarizes how research articles were narrowed down for a review. It begins with records gathered from two literature databases (PubMed and Google Scholar) and notes that none came from registers. A side box indicates that some items were removed prior to screening due to duplication and one language-related exclusion. The central pathway shows the number of records screened, followed by a large set excluded for relevance and access-related reasons. The remaining reports were retrieved and evaluated for eligibility, with none removed at that stage, ending with 62 studies included.PRISMA 2020 flow diagram showing the study selection process: 1,052 records identified, 57 duplicates and 1 non-English article removed, 994 records screened, 990 reports excluded, 62 reports assessed for eligibility, and 62 studies included in the review. PRISMA 2020 flow diagram illustrating the literature search, screening, and selection process for this structured narrative review. The included articles were categorized according to their primary focus of research - general PCOS epidemiology and diagnostic context (n = 4); AI-based diagnostic models (n = 13); ultrasound imaging and segmentation approaches (n = 13); hormonal biomarkers (n = 5); clinical biochemical markers (n = 9); and insulin resistance and obesity-related mechanisms (n = 22). The remaining references correspond to methodological tools and data sources utilized in the study. Formal risk-of-bias assessment tools (eg, QUADAS-2, ROBINS-I) were not applied, as the objective of this review was conceptual cross-domain synthesis not the quantitative evaluation of diagnostic performance. Diagnostic Approaches for Pcos Diagnosis of PCOS relies on a combination of clinical, biochemical, and imaging-based criteria. Each of these captures a different aspect of the syndrome. Recent advances have expanded the opportunity to include AI-driven models, molecular biomarkers, and metabolic indicators which enables more nuanced and potentially earlier detection. This section reviews the major diagnostic domains to inspect their individual roles and clinical relevance in different healthcare settings. Detection Using Artificial Intelligence: Machine Learning and Deep Learning Artificial intelligence has been a powerful non-invasive tool for detecting PCOS, complementing the conventional clinical, biochemical, and imaging-based approaches. AI-based diagnostic works can be classified broadly into two domains: (1) image-based diagnosis and (2) clinical and physiological data-driven diagnosis. These approaches minimise operator dependency, reduce subjective interpretation of ultrasound scans, and support reproducible, early-stage PCOS identification. Though ultrasound imaging is still the primary modality for morphological assessment of PCOS, the recent advances in deep learning have significantly improved automated interpretation. Convolutional neural networks (CNNs) and transfer-learning models such as VGG16 and InceptionV3 have demonstrated promising results. Hybrid pipelines of CNN feature extractors with ensemble classifiers (eg, XGBoost) have shown strong performance in detecting PCOS-related ovarian morphology.14,15 Hosain et al developed PCONet, a custom CNN for PCOS detection from ultrasound images, achieving 98.12% accuracy compared with 96.56% for InceptionV3 on the same dataset.16 It highlights the value of domain-specific architectures for gynaecological imaging. Deep learning has also been applied to follicle segmentation for automated estimation of antral follicle count (AFC), mean follicle diameter, and ovarian volume, which are closely linked to the Rotterdam criteria.17–19 Approaches using particle swarm optimisation (PSO) combined with neural networks, U-Net variants, attention mechanisms, and region-growing segmentation have reduced inter-operator variability and improved consistency in follicle quantification. A new research direction explores extra-ovarian imaging biomarkers like ocular features. Lv et al developed a sclera-based screening model combining an enhanced U-Net with ResNet classification. The model achieved an AUC of 0.979 and accuracy of 0.929 on a dataset of 721 women (388 PCOS, 333 controls).5 It suggests that systemic physiological alterations in PCOS may be detectable through biomarkers beyond ovarian morphology. The second branch of AI-based PCOS detection uses clinical, demographic, and biochemical features, such as body mass index (BMI), waist-hip ratio, LH/FSH ratio, fasting insulin, menstrual irregularity, and hyperandrogenic symptoms. Machine-learning models, like Random Forest, XGBoost, CatBoost, LightGBM, SVM/LS-SVM, logistic regression, naïve Bayes, and ensemble hybrids, have been evaluated for this task. Faris and Miften developed a classifier by coupling k-means and LS-SVM which achieves approximately 99% accuracy using physiological and biochemical inputs. It shows that the adequate performance is achievable with a reduced feature set.20 Modi et al compared multiple classifiers and identified CatBoost as the strongest overall performer, with precision 95%, accuracy 93%, recall 96%, and F1-score 96%.21 These results reflect the ability of gradient-boosting models to capture non-linear hormonal and metabolic relationships. Tiwari et al established the viability of using exclusively non-invasive, symptom-based, and anthropometric features for PCOS detection, achieving 93.25% accuracy with a Random Forest algorithm. It underscores the potential of AI as a first-line screening tool in low-resource settings.7 Beyond classification, AI models have been used in comparative studies evaluating multiple architectures for their ability to generalise across imaging and clinical data.22,23 Although, these accuracy figures should be interpreted cautiously, because the majority of studies used small, single-centre datasets with internal validation only. This limitation is examined in the Discussion. Below, a comparative summary of model performance across modalities is presented in Table 1. Table 1Performance Comparison of AI-Based Models for PCOS Detection Across Imaging, Clinical, and Symptom-Based ModalitiesModel TypeStudy (Ref)ModalityReported Metric(s)PCONet (CNN)Hosain et al,16Ultrasound imagingAccuracy: 98.12%InceptionV3Hosain et al,16Ultrasound imagingAccuracy: 96.56%CNN-based segmentation(U-Net variants / Attention U-Net)Kodipalli et al,19Ultrasound (segmentation)Dice coefficient: 0.78Scleral imaging model(U-Net + ResNet)Lv et al,5Ocular imagingAccuracy: 92.9%; AUC: 0.979Random ForestTiwari et al,7Symptom-based / anthropometricAccuracy: 93.25%CatBoostModi et al,21Clinical / biochemicalAccuracy: 93%; F1-score: 96% Table 1Performance Comparison of AI-Based Models for PCOS Detection Across Imaging, Clinical, and Symptom-Based ModalitiesModel TypeStudy (Ref)ModalityReported Metric(s)PCONet (CNN)Hosain et al,16Ultrasound imagingAccuracy: 98.12%InceptionV3Hosain et al,16Ultrasound imagingAccuracy: 96.56%CNN-based segmentation(U-Net variants / Attention U-Net)Kodipalli et al,19Ultrasound (segmentation)Dice coefficient: 0.78Scleral imaging model(U-Net + ResNet)Lv et al,5Ocular imagingAccuracy: 92.9%; AUC: 0.979Random ForestTiwari et al,7Symptom-based / anthropometricAccuracy: 93.25%CatBoostModi et al,21Clinical / biochemicalAccuracy: 93%; F1-score: 96% Performance Comparison of AI-Based Models for PCOS Detection Across Imaging, Clinical, and Symptom-Based Modalities Ultrasound Imaging Ultrasound is one of the main supports of PCOS diagnosis. The original Rotterdam criteria defined polycystic ovarian morphology (PCOM) as the presence of ≥12 follicles (2–9 mm) and/or an ovarian volume >10 mL in at least one ovary.11 With the availability of higher-resolution transvaginal probes, international guidelines now recommend a higher threshold of ≥20 follicles per ovary, while retaining ovarian volume ≥10 mL as a criterion (with exceptions for older equipment). In practice, clinicians rely on transvaginal ultrasound to assess AFC, ovarian volume, and stromal echogenicity. Manual follicle counting and visual assessment are operator-dependent and time-consuming, specially when ovaries contain numerous small follicles. This has driven extensive work on computer-assisted analysis and automated diagnostic systems.24 Image analysis-based methods isolate ovarian boundaries, stroma, and follicles from ultrasound images to compute morphological metrics such as follicle number, size distribution, and ovarian volume. Classical approaches use Otsu thresholding, active contours, and region-growing to separate cystic regions from background tissue in grayscale ultrasound images. Figure 4a shows the segmentation output of our pipeline on a representative frame containing a dominant ovarian cyst, the diameter of which is 5.52 cm as per the marking done by clinician. The method preprocesses the input via median filtering and CLAHE and restricts processing to the sonographic field-of-view. As the target is a single large anechoic structure rather than a cluster of small follicles, the cyst is localised from its darkest central seed and refines the boundary with a region-based active contour (Chan–Vese) which is followed by morphological cleanup and shape-based filtering on area and circularity. The entire process is summarised in Figure 4b. This recovers a single connected region of 81,637 px (equivalent diameter ≈ 322 px, ≈ 58 px/cm at the marked scale) with a circularity of 0.82, consistent with a simple cyst. Implementations combining Otsu method with level-set or Chan–Vese models further refine boundaries in low-contrast, noisy images.25,26 Figure 4Otsu and active-contour–based segmentation pipeline for ovarian ultrasound images. (a) Final segmentation output, with the cyst boundary in yellow colour, overlaid on the region of interest for a representative frame containing a dominant ovarian cyst. (b) Flowchart of the complete image-processing pipeline. The ultrasound image is from the Multi-Modality Ovarian Tumor Ultrasound (MMOTU) dataset, available at https://www.kaggle.com/datasets/orvile/mmotu-ovarian-ultrasound-images-dataset.27.The image A shows an ultrasound image with a highlighted boundary around a dominant ovarian cyst. The image B shows a flowchart detailing the segmentation process for ovarian ultrasound images. The process begins with the input of an ovarian ultrasound (transvaginal frame), followed by preprocessing using a median filter, CLAHE and cropping. Next is field-of-view masking to restrict to the sonographic fan. Anechoic seed detection identifies the darkest central low-intensity point. A region-based active contour (Chan-Vese refinement) is applied from the seed. Morphological cleanup involves open, close and hole fill operations. Shape-based filtering is performed based on area and circularity. The final step is determining the cyst boundary and reading out the area and equivalent diameter.Ultrasound image and flowchart of ovarian cyst segmentation process. Otsu and active-contour–based segmentation pipeline for ovarian ultrasound images. (a) Final segmentation output, with the cyst boundary in yellow colour, overlaid on the region of interest for a representative frame containing a dominant ovarian cyst. (b) Flowchart of the complete image-processing pipeline. The ultrasound image is from the Multi-Modality Ovarian Tumor Ultrasound (MMOTU) dataset, available at https://www.kaggle.com/datasets/orvile/mmotu-ovarian-ultrasound-images-dataset.27. Among deep learning architectures, U-Net and its derivatives, including Attention U-Net, Residual U-Net, and transformer-hybrid variants, have become the dominant frameworks for ovarian and follicle segmentation. Their encoder–decoder structure along with skip connections preserves spatial detail which are very important for follicle boundary delineation. In contrast, newer modifications that incorporate adaptive bilateral filtering and attention mechanisms further enhance robustness to ultrasound speckle and anatomical variation.28 These methods emphasise structural interpretation rather than direct classification, providing detailed morphological maps for more objective PCOM assessment. Table 2 presents a comparative analysis of classical and deep learning segmentation methods on the MMOTU (Multi-Modality Ovarian Tumor Ultrasound) dataset.27 Table 2Quantitative Comparison of Classical and Deep Learning–Based Ultrasound Segmentation Methods for PCOS-Related Follicle Detection (Dice, IoU, Precision, Recall)Dice scoreIoUPrecisionRecallOtsu0.16420.09910.17940.2095Chan–Vese0.29960.18390.20030.8016U-Net0.77950.67320.72470.9080Attention U-Net0.78460.68360.76650.8820 Table 2Quantitative Comparison of Classical and Deep Learning–Based Ultrasound Segmentation Methods for PCOS-Related Follicle Detection (Dice, IoU, Precision, Recall)Dice scoreIoUPrecisionRecallOtsu0.16420.09910.17940.2095Chan–Vese0.29960.18390.20030.8016U-Net0.77950.67320.72470.9080Attention U-Net0.78460.68360.76650.8820 Quantitative Comparison of Classical and Deep Learning–Based Ultrasound Segmentation Methods for PCOS-Related Follicle Detection (Dice, IoU, Precision, Recall) As shown in Table 2, deep learning–based models have outperformed classical image-processing approaches across all evaluation metrics. Otsu’s thresholding exhibits very low Dice score (0.16) and IoU (0.09), reflecting poor overlap due to its dependence on global intensity assumptions. The Chan–Vese model demonstrates improved Dice score (0.30) and notably high recall (0.80), indicating effective detection of follicular regions. But, its low precision (0.20) suggests significant over-segmentation. In contrast, U-Net and Attention U-Net achieve Dice scores of approximately 0.78 and IoU values above 0.67, demonstrating strong agreement with ground truth segmentation. These models also maintain a favourable balance between precision and recall, which highlight their ability to capture both structural boundaries and textural variations in ultrasound data. Figure 5 shows the precision–recall distribution of the methods. Classical approaches fall in low-precision or imbalanced regions, while deep learning models cluster in the high-performance area. Dice and IoU follow the same trend, with IoU consistently lower due to its stricter definition. U-Net and Attention U-Net appear very close, indicating only minor performance differences. Figure 5Comparative performance of classical and deep learning–based segmentation methods for ovarian follicle detection, summarising Dice coefficient, IoU, precision, and recall across approaches.Scatter plot comparing precision and recall of Chan–Vese, Otsu, U-Net, and Attention U-Net, with overlapping bubbles representing Dice score and IoU; U-Net and Attention U-Net show higher precision and recall than the classical methods. Comparative performance of classical and deep learning–based segmentation methods for ovarian follicle detection, summarising Dice coefficient, IoU, precision, and recall across approaches. These improvements in segmentation accuracy directly support more reliable follicle quantification in clinical practice. Accurate AFC is one of the strongest ultrasonographic markers of PCOS.20,29 Several studies propose automated or semi-automated follicle detection using pattern recognition, blob detection, and region-based algorithms applied to transvaginal ultrasound scans.30–32 Systems designed specifically for PCOS often perform follicle recognition followed by rule-based or machine-learning based classification of PCOS versus non-PCOS ovaries. It relies on AFC and derived morphological indices, reducing subjectivity and inter-observer variability.30–32 More recent work applies deep learning directly to ultrasound images. Vision Transformer (ViT)-based models and CNN–U-Net hybrids have been developed to classify PCOS from B-mode ovarian images, sometimes after an explicit follicle segmentation step,22,28,33 achieving high diagnostic accuracy in experimental datasets. Other studies combine clinical metrics with ultrasonographic imaging features to build integrated PCOS prediction models using machine-learning classifiers which are trained on both morphology-related measurements and patient-level clinical data.33,34 There is also work on mining large collections of ultrasound reports from electronic medical records to automatically identify PCOM patterns using text- or data-mining approaches.35 Hormonal Profile-Based Diagnosis of PCOS Hormonal profiling remains one of the most widely used biochemical approaches for PCOS diagnosis. It supports identification of the condition and helps distinguish it from other endocrine disorders such as hypothyroidism, hyperprolactinaemia, and non-classical congenital adrenal hyperplasia.11 A common abnormality in women with PCOS is elevated luteinising hormone (LH). In PCOS Patients, LH-to-FSH ratios frequently exceeding 2:1, compared with an approximate 1:1 ratio in eumenorrheic women during the early follicular phase.20 This rise in LH stimulates theca cells to produce androgens, while relatively low follicle-stimulating hormone (FSH) impairs granulosa-cell conversion of androgens to oestrogens. This contributes to follicular arrest and anovulation. Genetic studies of LH receptor polymorphisms further support the role of LH signaling in PCOS susceptibility.36 Clinically, elevated LH, raised LH/FSH ratios, and biochemical hyperandrogenism, including raised total or free testosterone and sometimes elevated DHEAS, correlate with hirsutism, acne, and androgenic alopecia.20,37 AMH, secreted by granulosa cells of growing follicles, has become a reliable marker of PCOS. Women with PCOS have a greater number of antral follicles, which collectively secrete elevated AMH, resulting in markedly higher circulating concentrations.38–40 Several studies report AMH cut-off values that differentiate PCOS from eumenorrheic controls with good sensitivity and specificity.39 Work in Chinese and other ethnic populations confirms AMH’s diagnostic value, though population-AMH’s specific thresholds are consistently required.38,39 Strong correlations between serum AMH, intrafollicular AMH, and AFC further support AMH as a reliable indicator of follicular excess.38,40 AMH has also been proposed to inhibit aromatase and dampen FSH responsiveness in granulosa cells, thereby promoting anovulation.38 This dual role of a diagnostic marker as well as a pathophysiological mediator makes AMH especially relevant in clinical practice. Beyond LH/FSH and AMH, broader hormonal panels typically include total and free testosterone, sex hormone-binding globulin (SHBG), DHEAS, oestradiol, TSH and sometimes prolactin to exclude differential diagnoses.20,37 Studies in different populations, including Nigerian cohorts, show that though exact distributions of LH, FSH, AMH, and androgens may vary, the combination of elevated AMH, hyperandrogenism, and ovulatory dysfunction is consistently useful for PCOS diagnosis.37,40 Key biochemical markers and their diagnostic relevance across different studies are summarised in Table 3. Figure 6 illustrates the relationship between AMH levels and follicular morphology, highlighting how elevated AMH is associated with increased antral follicle count (AFC) and the development of PCOM. Figure 6Comparative diagnostic performance of key hormonal biomarkers in PCOS, illustrating sensitivity and specificity of AMH, LH/FSH ratio, testosterone, and intrafollicular AMH across reported studies.A diagram showing AMH′s role in PCOM development via AFC increase. Table 3Diagnostic Performance of Key Hormonal Biomarkers (AMH, LH/FSH Ratio, Testosterone) Across Different Populations, Including Sensitivity and SpecificityBiomarkerStudy (Ref)Key FindingDiagnostic RolePopulationAMHYue.39Elevated in PCOS; defined diagnostic cut-offHigh sensitivity and specificityChineseAMHSivanandy.38Strong correlation with AFCDiagnostic & pathophysiologicalReviewIntrafollicular AMHStracquadanio.40Correlates with serum AMHReflects follicular environmentItalianLH/FSH ratioEzeiruaku.37Higher in PCOS vs controlIndicates endocrine imbalanceNigerianTestosteroneEzeiruaku.37Elevated in PCOSMarker of hyperandrogenismNigerian Comparative diagnostic performance of key hormonal biomarkers in PCOS, illustrating sensitivity and specificity of AMH, LH/FSH ratio, testosterone, and intrafollicular AMH across reported studies. Table 3Diagnostic Performance of Key Hormonal Biomarkers (AMH, LH/FSH Ratio, Testosterone) Across Different Populations, Including Sensitivity and SpecificityBiomarkerStudy (Ref)Key FindingDiagnostic RolePopulationAMHYue.39Elevated in PCOS; defined diagnostic cut-offHigh sensitivity and specificityChineseAMHSivanandy.38Strong correlation with AFCDiagnostic & pathophysiologicalReviewIntrafollicular AMHStracquadanio.40Correlates with serum AMHReflects follicular environmentItalianLH/FSH ratioEzeiruaku.37Higher in PCOS vs controlIndicates endocrine imbalanceNigerianTestosteroneEzeiruaku.37Elevated in PCOSMarker of hyperandrogenismNigerian Diagnostic Performance of Key Hormonal Biomarkers (AMH, LH/FSH Ratio, Testosterone) Across Different Populations, Including Sensitivity and Specificity Clinical, Biochemical, and Molecular Diagnostic Approaches Beyond hormonal profiling, a wide range of clinical, biochemical, and molecular tests are explored to refine PCOS detection and understand its metabolic and genetic heterogeneity. These approaches extend conventional diagnostics by incorporating nanotechnology-enhanced assays, environmental biomarkers, miRNAs, and metabolomic signatures. Among these emerging approaches, nanotechnology-based platforms have been proposed to improve the sensitivity, specificity, and early-detection capability of conventional PCOS biomarkers. A comprehensive review described the application of nano-formulations, nano-biosensors, and nano-antioxidants in both PCOS diagnosis and therapy, leveraging properties such as high surface-area-to-volume ratio, tunable particle size, and enhanced molecular reactivity.9 From a diagnostic perspective, nano-biosensors have demonstrated the ability to detect classical PCOS markers, including LH, FSH, androgens, AMH, and SHBG, at lower concentrations and with greater speed than standard immunoassays.9 A separate spectroscopic approach, surface-enhanced Raman scattering (SERS), has also been applied to PCOS diagnosis. One study used SERS to analyse chemerin-related spectral signatures in follicular fluid and plasma, combining spectral data with multivariate statistical tools, including principal component analysis and partial least squares regression, to discriminate PCOS from non-PCOS samples.41 Although the causal role of chemerin in PCOS pathogenesis remains under investigation, this work illustrates how label-free spectroscopic techniques can function as non-traditional biochemical diagnostics. Along with advances in diagnostic technologies, environmental factors are also causing PCOS development. A systematic review and meta-analysis reported significantly higher circulating bisphenol A (BPA) levels in women with PCOS, with associations observed between BPA exposure, hyperandrogenism, and insulin resistance.42 While BPA cannot serve as a standalone diagnostic marker, its consistent elevation in PCOS cohorts positions it as a risk-associated biomarker that may contribute to susceptibility profiling, particularly in women presenting with unexplained hyperandrogenism in the context of known environmental exposures. At the molecular level, regulatory biomarkers such as microRNAs (miRNAs) have emerged as promising tools for PCOS detection. Researchers have shown that several circulating miRNAs are consistently altered in PCOS, suggesting their potential as non-invasive diagnostic markers.6 At the tissue level, miR-483-5p and miR-486-5p are reduced in cumulus cells of mature (metaphase II) oocytes, indicating possible disruption in communication between the oocyte and surrounding granulosa cells.43,44 Exosomal miRNAs further expand the diagnostic landscape. miR-424-5p is upregulated in follicular-fluid-derived exosomes in PCOS and induces granulosa-cell senescence by targeting CDCA4, linking the molecular alterations with follicular dysfunction.44 Network-level studies identify abnormal interactions between miRNAs and mRNAs in PCOS. These changed miRNAs are strongly linked to markers of insulin sensitivity and inflammation, such as HOMA-IR and pro-inflammatory cytokines. This suggests that miRNAs may reflect both reproductive and metabolic changes in PCOS.45 Metabolomic profiling offers a further biochemical approach for PCOS detection. Using SWATH mass spectrometry, studies analysing follicular fluid have identified distinct metabolic signatures, encompassing amino acids, lipid species, and energy-related metabolites, that differentiate PCOS patients from controls.46 Because follicular fluid closely reflects the ovarian microenvironment, these metabolomic alterations may reveal subtle or early-stage PCOS phenotypes not captured by routine serum hormone testing. These disruptions in energy metabolism which are aligned with impaired insulin signaling and GLUT4 trafficking defects prevalent in PCOS.47 Collectively, these emerging molecular and spectroscopic approaches address dimensions of PCOS heterogeneity that conventional hormonal and imaging criteria do not fully capture. Their diagnostic value remains contingent on assay standardisation, cycle-phase controls, and validation in ethnically diverse cohorts. Representative quantitative findings for emerging biomarkers are summarised in Table 4. Table 4Summary of Emerging Molecular and Spectroscopic Biomarkers for PCOS Diagnosis, Including miRNA, Metabolomic, and SERS-Based Approaches with Reported Quantitative PerformanceApproach CategoryStudy (Ref)Biomarker/ TargetBiological SampleKey FindingDiagnostic RelevanceSpectroscopic(Nanotechnology-enabled)Momen pour.41Chemerin-associated spectral signatures (SERS)Follicular fluid, plasmaDistinct spectral profiles differentiate PCOS from controls using multivariate analysisDemonstrates potential of label-free, high-sensitivity diagnostic platformsEnvironmental BiomarkerHu.42Bisphenol A (BPA)Serum / plasmaSignificantly elevated BPA levels in women with PCOS; associated with hyperandrogenism and insulin resistanceRisk-associated biomarker reflecting environmental contribution to PCOSmiRNA (Cellular level)Shi.43miR-483-5p, miR-486-5pCumulus cells (MII oocytes)Significant downregulation in PCOS, indicating altered oocyte–granulosa cell interactionReflects impaired follicular microenvironment and reproductive dysfunctionExosomal miRNAYuan.44miR-424-5pFollicular fluid exosomesUpregulated in PCOS; induces granulosa cell senescence via CDCA4 targetingLinks molecular signaling to follicular dysfunction and ovarian ageing processesmiRNA–mRNA NetworkQin.45Regulatory interaction networksTissue / computational datasetsDysregulated networks correlated with insulin resistance (HOMA-IR) and inflammation markersIntegrates reproductive and metabolic dysfunction at systems levelMetabolomicsSun.46Amino acids, lipids, energy metabolitesFollicular fluidDistinct metabolic signatures identified using SWATH-MS differentiating PCOS from controlsCaptures ovarian microenvironment alterations not detected in serum markersInsulin Signaling Pathway(Mechanistic Support)Van Gerwen.47GLUT4 trafficking, insulin signaling componentsCellular / molecular studiesImpaired insulin signaling and glucose transport mechanisms underlying metabolic dysfunctionProvides mechanistic basis linking metabolomic alterations to insulin resistance in PCOS Table 4Summary of Emerging Molecular and Spectroscopic Biomarkers for PCOS Diagnosis, Including miRNA, Metabolomic, and SERS-Based Approaches with Reported Quantitative PerformanceApproach CategoryStudy (Ref)Biomarker/ TargetBiological SampleKey FindingDiagnostic RelevanceSpectroscopic(Nanotechnology-enabled)Momen pour.41Chemerin-associated spectral signatures (SERS)Follicular fluid, plasmaDistinct spectral profiles differentiate PCOS from controls using multivariate analysisDemonstrates potential of label-free, high-sensitivity diagnostic platformsEnvironmental BiomarkerHu.42Bisphenol A (BPA)Serum / plasmaSignificantly elevated BPA levels in women with PCOS; associated with hyperandrogenism and insulin resistanceRisk-associated biomarker reflecting environmental contribution to PCOSmiRNA (Cellular level)Shi.43miR-483-5p, miR-486-5pCumulus cells (MII oocytes)Significant downregulation in PCOS, indicating altered oocyte–granulosa cell interactionReflects impaired follicular microenvironment and reproductive dysfunctionExosomal miRNAYuan.44miR-424-5pFollicular fluid exosomesUpregulated in PCOS; induces granulosa cell senescence via CDCA4 targetingLinks molecular signaling to follicular dysfunction and ovarian ageing processesmiRNA–mRNA NetworkQin.45Regulatory interaction networksTissue / computational datasetsDysregulated networks correlated with insulin resistance (HOMA-IR) and inflammation markersIntegrates reproductive and metabolic dysfunction at systems levelMetabolomicsSun.46Amino acids, lipids, energy metabolitesFollicular fluidDistinct metabolic signatures identified using SWATH-MS differentiating PCOS from controlsCaptures ovarian microenvironment alterations not detected in serum markersInsulin Signaling Pathway(Mechanistic Support)Van Gerwen.47GLUT4 trafficking, insulin signaling componentsCellular / molecular studiesImpaired insulin signaling and glucose transport mechanisms underlying metabolic dysfunctionProvides mechanistic basis linking metabolomic alterations to insulin resistance in PCOS Summary of Emerging Molecular and Spectroscopic Biomarkers for PCOS Diagnosis, Including miRNA, Metabolomic, and SERS-Based Approaches with Reported Quantitative Performance Insulin Resistance and Obesity in PCOS Insulin resistance (IR) represents one of the most important metabolic hallmarks of PCOS, contributing to hyperandrogenism, chronic anovulation, and increased long-term cardiometabolic risk. Obesity, particularly central adiposity, exacerbates IR through inflammatory adipokine dysregulation and impaired lipid handling. Together, IR and obesity define a clinically distinct metabolic PCOS subtype that requires timely biochemical identification.48 Women with PCOS have well-documented abnormalities in the IRS-1/PI3K/AKT/GLUT4 signalling cascade in skeletal muscle and adipose tissue, resulting in reduced glucose uptake and compensatory hyperinsulinaemia.48 Hyperinsulinism suppresses SHBG, stimulates ovarian theca-cell androgen synthesis, and raises circulating free testosterone. These mechanisms are supported by experimental models in which oxidative stress impairs insulin signalling in skeletal muscle under PCOS-like conditions,49 and in which pharmacological therapies such as berberine restore GLUT4 expression and PI3K/AKT/MAPK pathway activity.50 The sustained growth of IR-focused PCOS research, documented across nearly four decades of published literature, reflects the increasing recognition of IR as a central, therapeutically relevant disease mechanism rather than a secondary complication.51 Translating these mechanistic insights into clinical practice, a range of biochemical markers have been developed to identify and quantify insulin resistance in PCOS. In clinical practice, traditional indicators such as fasting glucose, fasting insulin, HOMA-IR, QUICKI, and oral glucose tolerance tests are still essential for identifying IR.52 Several newer biomarkers offer improved diagnostic precision. Copeptin, a surrogate measure of vasopressin, shows a high association with IR in PCOS.53 InsuTAG combines fasting insulin and triglyceride levels and provides a composite indicator of adipose–insulin metabolic dysfunction.54 Fetuin-B level is high in newly diagnosed PCOS and is correlated with IR. These levels decrease after treatment with GLP-1 receptor agonists.55 In infertile women with PCOS, lipoprotein ratios, especially TG/HDL-C, have been linked to IR.56 Notable differences in trace elements such as zinc and selenium distinguish PCOS women with IR from those without.57 The adiponectin-to-leptin ratio, often lower in PCOS, correlates with the severity of IR and reflects adipose inflammation.58 MicroRNAs offer a multi-omic perspective on metabolic dysfunction. Aberrant miRNA–mRNA regulatory networks show strong correlations with HOMA-IR, fasting insulin, TNF-α, and IL-6, indicating that miRNAs may simultaneously capture chronic inflammation and IR.46 In addition to insulin resistance, obesity further amplifies metabolic dysfunction in PCOS through multiple systemic mechanisms like increased leptin, decreased adiponectin, and altered levels of resistin, chemerin, and inflammatory cytokines.59 Non-alcoholic fatty liver disease (NAFLD) has a particularly strong association with PCOS; meta-analyses show that NAFLD prevalence remains markedly higher in PCOS even after controlling for BMI,60,61 and hepatic steatosis in adolescents with obesity and PCOS appears to be primarily driven by enhanced de novo lipogenesis.62 Elevated uric acid/creatinine ratios and decreased SHBG are additional obesity-related indicators linked to metabolic PCOS phenotypes.63,64 Altered circulating intelectin-1 further reflects metabolic dysregulation associated with IR. Preliminary findings also suggest that extra-ovarian organs, such as the cornea, may show IGF-1-related sensitivity to systemic IR in PCOS.65 To integrate these complex metabolic and biochemical signals, machine-learning approaches have been increasingly explored to integrate anthropometric parameters, metabolic biomarkers, adipokines, and reproductive hormones which have shown promise in identifying women at high metabolic risk. AI-driven hormonal profiling tools, PCOS-focused mobile applications, and statistical modelling approaches demonstrate early success in classifying metabolic PCOS phenotypes and predicting IR severity.66,67 IR and obesity are not peripheral features but central pathophysiological drivers of PCOS that carry independent diagnostic value. Classical biochemical indices remain the clinical standard, while a growing range of novel biomarkers, spanning hormonal surrogates, adipokine ratios, organ-level manifestations, and miRNA signatures, are steadily improving the metabolic diagnostic toolkit. Machine-learning integration of these signals represents a promising route toward earlier identification of high-risk phenotypes, particularly in women whose conventional metabolic markers appear within normal range. Key insulin resistance biomarkers and their correlation with HOMA-IR are summarised in Table 5. Table 5Correlation of Insulin Resistance–Related Biomarkers with HOMA-IR in PCOS, Including Metabolic, Adipokine, and Lipid-Based IndicatorsBiomarker CategoryStudy (Ref)BiomarkerKey ObservationClinical RelevanceHormonal surrogateAli et al,52CopeptinAssociated with insulin resistance and vasopressin activityEarly marker of metabolic stressComposite indexOmma et al,53InsuTAGCombines insulin and triglycerides to reflect metabolic dysfunctionSensitive indicator of adipose–insulin interactionHepatokineMokou et al,54Fetuin-BElevated in PCOS; reduced after metabolic treatmentReflects hepatic contribution to insulin resistanceLipid ratioGhaffarzad et al,55TG/HDL-CElevated in PCOS, especially in infertile womenIndicator of dyslipidemia-linked IRAdipokine balanceOlszanecka et al,57Adiponectin / Leptin ratioReduced in PCOS; correlates with inflammation and IR severityReflects adipose tissue dysfunctionTrace elementsKanafchian et al,56Zinc, SeleniumAltered levels in PCOS with IRSuggests micronutrient involvement in metabolic regulation Table 5Correlation of Insulin Resistance–Related Biomarkers with HOMA-IR in PCOS, Including Metabolic, Adipokine, and Lipid-Based IndicatorsBiomarker CategoryStudy (Ref)BiomarkerKey ObservationClinical RelevanceHormonal surrogateAli et al,52CopeptinAssociated with insulin resistance and vasopressin activityEarly marker of metabolic stressComposite indexOmma et al,53InsuTAGCombines insulin and triglycerides to reflect metabolic dysfunctionSensitive indicator of adipose–insulin interactionHepatokineMokou et al,54Fetuin-BElevated in PCOS; reduced after metabolic treatmentReflects hepatic contribution to insulin resistanceLipid ratioGhaffarzad et al,55TG/HDL-CElevated in PCOS, especially in infertile womenIndicator of dyslipidemia-linked IRAdipokine balanceOlszanecka et al,57Adiponectin / Leptin ratioReduced in PCOS; correlates with inflammation and IR severityReflects adipose tissue dysfunctionTrace elementsKanafchian et al,56Zinc, SeleniumAltered levels in PCOS with IRSuggests micronutrient involvement in metabolic regulation Correlation of Insulin Resistance–Related Biomarkers with HOMA-IR in PCOS, Including Metabolic, Adipokine, and Lipid-Based Indicators Cross-Domain Comparison of Diagnostic Approaches A cross-domain comparison of the major diagnostic approaches discussed above is presented in Table 6, to highlight differences in performance, clinical maturity, validation status, and practical applicability. The comparison highlights clear differences in the strengths and limitations of each diagnostic domain. Hormonal profiling and ultrasound imaging remain the most clinically established approaches. AI-based methods, on the other hand, show high reported accuracy and good accessibility potential; but they are limited by insufficient external validation and lower interpretability. Molecular approaches are an emerging area. Its potential is the insights of underlying disease mechanisms though currently lacking clinical maturity and standardisation. The findings suggest that no single modality is sufficient in isolation, and that effective PCOS diagnosis will depend on integrating complementary strengths across domains. Table 6Cross Domain Comparison of Major Diagnostic Approaches Discussed AboveCriterionAI-Based MethodsHormonal ProfilingUltrasound ImagingMolecular ApproachesReported accuracyHighModerate to highModerate to highEmergingClinical maturityLimitedWell establishedWell establishedEarly stageExternal validationLimitedStrongStrongLimitedAccessibilityModerate to highModerateModerate to lowLimitedInterpretabilityLimitedHighModerateLimited Table 6Cross Domain Comparison of Major Diagnostic Approaches Discussed AboveCriterionAI-Based MethodsHormonal ProfilingUltrasound ImagingMolecular ApproachesReported accuracyHighModerate to highModerate to highEmergingClinical maturityLimitedWell establishedWell establishedEarly stageExternal validationLimitedStrongStrongLimitedAccessibilityModerate to highModerateModerate to lowLimitedInterpretabilityLimitedHighModerateLimited Cross Domain Comparison of Major Diagnostic Approaches Discussed Above While each domain offers distinct advantages, the progress across diagnostic domains suffers from several common limitations. Many studies are based on relatively small or single-centre cohorts, which limits generalisability across populations and healthcare settings. Variability in diagnostic thresholds, assay protocols, imaging acquisition methods, and dataset composition also complicates direct comparisons between studies. These challenges highlight the need for standardised validation frameworks and multi-centre evaluation before the widespread clinical implementation of emerging diagnostic approaches. Discussion Based on the cross-domain comparison presented in the previous section, PCOS diagnosis is gradually shifting from single-method, clinician-dependent approaches to multimodal, phenotype-focused assessment. Advances in artificial intelligence, imaging, hormonal profiling, and metabolic evaluation are complementary rather than replacing one another. Each approach contributes a distinct type of diagnostic insight. These methods are not interchangeable, as they address different aspects of the condition, vary across healthcare settings, and differ in their level of clinical readiness. Recognizing where they align, where they differ, and where evidence remains limited is essential for interpreting current findings and guiding future research. The Accuracy–Accessibility Trade-off in AI-Based Screening A key challenge in clinical AI is maintaining the balance between model complexity and practical usability. Image-based deep learning models, including CNNs, U-Net variants, and Vision Transformers, offer high morphological detail when supported by high-resolution ultrasound images, well-annotated datasets, and sufficient computational resources. These requirements pose significant challenges in low- and middle-income settings.7–9 In contrast, data-driven models operating on symptom scores and anthropometric measurements are far less resource-intensive and have demonstrated competitive diagnostic performance, with accuracies exceeding 93% in some studies.16 This creates a realistic two-tier screening paradigm. The symptom-and-anthropometric models for initial population-level triage, and imaging-augmented AI for confirmatory assessment in better-resourced settings.7,16 The potential of these AI models remains largely unproven, as the reviewed literature consistently fails to provide external validation. High accuracy figures, many exceeding 95%, are predominantly derived from small, demographically homogeneous datasets evaluated using internal cross-validation, a procedure known to yield optimistic performance estimates.14,15 None of the included studies validated their model prospectively or in a real-world clinical workflow. Until multi-centre, externally validated studies are available, these results should be interpreted as signals of feasibility rather than estimates of deployable clinical performance.17 The field would benefit substantially from shared benchmark datasets, pre-registered validation protocols, and standardised reporting of sensitivity, specificity, and likelihood ratios alongside accuracy.26 Imaging and Hormonal Profiling: Complementary, Not Competing A repeated implicit assumption in both the AI and hormonal literature is that imaging and biochemical markers are competing diagnostic strategies. The evidence reviewed here does not support this understanding. Ultrasound-derived AFC and ovarian volume operationalise the morphological criterion of PCOM but carry no information about androgenic status, insulin sensitivity, or ovulatory function.20 Conversely, hormonal panels, LH/FSH ratio, AMH, free testosterone, SHBG, capture the endocrine axis but provide no spatial information about follicular architecture.34,35 The Rotterdam criteria, which require at least two of three features (oligo-anovulation, hyperandrogenism, PCOM), implicitly recognise this complementarity: imaging and hormonal data each independently contribute to meeting diagnostic thresholds.20 What the reviewed evidence adds to this picture is that AMH now occupies an unusual bridging position: it is a hormonal marker whose concentration is directly determined by antral follicle number, making it a biochemical surrogate for morphological PCOM.35,38 This has practical significance: in settings where transvaginal ultrasound is unavailable or poorly tolerated, a sufficiently elevated AMH may partially substitute for the imaging criterion, particularly when combined with clinical hyperandrogenism.35,36 The reliability of this substitution across different ethnic groups, where population-specific AMH thresholds are consistently required, highlights a limitation that deserves explicit attention in any future diagnostic guideline revision.36 Similarly, the integration of AI models trained jointly on imaging and hormonal features, rather than on either modality alone, represents a logical and underexplored next step that could outperform either approach in isolation.32 Metabolic Phenotyping: The Underutilised Diagnostic Dimension Though current diagnostic frameworks for PCOS mainly focus on reproductive features such as menstrual irregularity, hyperandrogenism, and ovarian morphology, the evidence reviewed here shows that insulin resistance and metabolic dysfunction are not just associated conditions but key drivers of the disease.48 Abnormalities in the IRS-1/PI3K/AKT/GLUT4 pathway in skeletal muscle and adipose tissue lead to hyperinsulinaemia. This increases androgen production in theca cells and reduces SHBG levels, contributing to the hormonal profile seen in PCOS.47,49 This two-way relationship between insulin resistance and hyperandrogenism suggests that metabolic markers have direct diagnostic value, not just a supportive role.61,65 A wide range of emerging insulin resistance markers has been studied, including copeptin, fetuin-B, InsuTAG, adipokine ratios, lipoprotein ratios, and miRNA-based signatures.53–57 These efforts reflect the need for more sensitive, affordable, and earlier markers than HOMA-IR. Although none has yet proven superior in large prospective studies, current trends suggest that future PCOS diagnosis may include a metabolic risk score along with reproductive criteria. This would have particular value for identifying the substantial proportion of lean PCOS women in whom classical metabolic markers such as BMI and fasting insulin may appear normal, yet who carry significant insulin signalling dysfunction at the tissue level.48,53 Machine learning approaches integrating multi-omic metabolic data hold particular promise for defining these metabolically distinct subphenotypes.65,66 Translational Readiness and the Path to Clinical Deployment Across the diagnostic domains reviewed, translational readiness varies markedly. Hormonal profiling and transvaginal ultrasound are established clinical tools, albeit with ongoing refinement of thresholds and automation of image analysis.20,21 AI-based classifiers using routinely collected clinical and biochemical data are approaching a state of readiness for prospective evaluation, conditional on external validation.14–16 Emerging molecular and spectroscopic diagnostics, miRNAs, metabolomics, SERS, nano-biosensors, remain substantially earlier in their development trajectory, with significant gaps in assay standardisation, cycle-phase sensitivity controls, and clinical validation cohorts.40,43,47 A further equity concern is present throughout the literature. The majority of AI and biomarker studies are conducted in East Asian, European, or North American cohorts, with limited representation of South Asian, Sub-Saharan African, or Latin American populations, precisely those in whom underdiagnosis is most prevalent and healthcare infrastructure most constrained.1,2 Population-specific differences in AMH thresholds, androgen reference ranges, and metabolic risk profiles are well documented;36,39 models trained on homogeneous cohorts are unlikely to generalise equitably without deliberate diversity in training and validation data. Any future roadmap for AI-assisted PCOS detection must treat population representativeness not as an aspirational addition but as a core methodological requirement. Implications and Future Directions The evidence suggests that future PCOS diagnostics will be multimodal, resource-adapted, and guided by phenotypic variation.32,64 At the population level, non-invasive AI models using symptoms and anthropometric data can support large-scale screening, particularly in settings with limited laboratory access.7,16 At the clinical level, integrated approaches combining hormonal, imaging, and metabolic data, analysed using AI tools, may enable phenotype-specific diagnosis and better capture the range of PCOS presentations.20,35 At the research level, molecular methods such as circulating miRNA panels and metabolomic signatures may allow earlier identification of at-risk individuals.43,47 Achieving this requires four key steps. First, external validation through multi-centre prospective studies with standardised protocols and phenotype definitions.20 Second, outcome-focused evaluation using clinically relevant endpoints, such as agreement in AFC with expert sonographers and prediction of metabolic risk, rather than relying only on pixel-level metrics.28 Third, equity-focused design with inclusion of diverse populations, especially underrepresented low-SDI groups.1,2 Fourth, biomarker standardisation, ensuring that miRNA panels, metabolomic tests, and spectroscopic methods are validated across sites with appropriate cycle-phase control.43,47 Limitations Several limitations should be noted. First, this is a structured narrative review and not a formal systematic review or meta-analysis. Study selection was based on domain-representative keywords rather than a pre-registered PICO protocol, and no formal quality assessment tools (eg, QUADAS-2 or ROBINS-I) were applied. Second, screening was conducted by a single reviewer without inter-rater reliability assessment, which may introduce bias. Third, the search was limited to English-language studies from two databases, and grey literature was not included, which may underrepresent low- and middle-income settings. Fourth, the search covered studies from 2014 to early 2025, so recent work is not included. Finally, variability in diagnostic criteria and outcomes across studies prevented quantitative synthesis. Reported accuracy values should therefore be interpreted in light of these limitations and those discussed in Detection Using Artificial Intelligence: Machine Learning and Deep Learning. Conclusion PCOS diagnosis is progressing from subjective, single-technique methods toward an integrated and data-driven system. Whereas conventional ultrasound methods are still useful for basic analysis, advanced deep learning models have provided more reliable ovarian features even in challenging imaging conditions. Hormonal indicators, together with molecular and metabolic biomarkers, including miRNAs and insulin resistance markers, increase the likelihood of earlier recognition of high-risk cases. Such multimodal approaches have the potential to reduce diagnostic delays. However, successful clinical translation and future advancement will depend on standardised validation, diverse and representative datasets, and clinically interpretable diagnostic frameworks.