AI-Based Ovarian Phenotyping Using Follicle Size Distribution Patterns for PCOS Assessment

Background: Polycystic Ovary Syndrome (PCOS) is commonly diagnosed using ultrasound-based follicle counts and size thresholds. However, manual assessment is subjective, time-consuming, and limited to coarse criteria that overlook the broader morphological patterns of follicular development. In particular, follicle size distribution, rather than absolute count alone, remains underutilized as a diagnostic marker. Methods: We present an artificial intelligence–based ovarian phenotyping approach that automatically quantifies follicle size and shape distributions from transvaginal ultrasound images. An AI-driven object detection model was used solely as a measurement engine to extract follicle geometry, enabling downstream analysis of diameter distributions and aspect ratio patterns. Importantly, the proposed methodology operates in a resolution-independent manner, allowing robust morphological analysis even when spatial calibration metadata are unavailable. Results: Across 302 ultrasound images, the derived follicle diameter distributions revealed distinct multimodal patterns, characterized by an over-representation of small antral follicles and secondary size populations consistent with disrupted folliculogenesis in PCOS. Aspect ratio analysis further identified deviations from circular morphology in a subset of follicles. Clustering based on distributional features demonstrated that 67% of cases exhibited high-density, small-follicle dominance, a hallmark of polycystic ovarian morphology. Conclusion: This study introduces AI-based follicle size distribution phenotyping as a complementary and scalable imaging marker for PCOS assessment. By shifting focus from follicle counting to distribution-level morphological signatures, the proposed approach offers a reproducible, operator-independent pathway toward more nuanced ovarian characterization. The framework is designed to be dataset-agnostic and readily extensible to future clinically curated cohorts.