Conditional Diffusion Framework for Probabilistic Modeling of Follicular Growth During Controlled Ovarian Stimulation

Controlled ovarian stimulation (COS) in in vitro fertilization (IVF) relies on serial ultrasound monitoring to guide medication adjustments and determine trigger timing. Existing predictive approaches are predominantly deterministic and focus on final oocyte yield or static trigger-day estimation, without explicitly modeling the evolving dynamics and uncertainty of the follicular cohort. This technical note introduces a conditional generative framework for probabilistic modeling of follicular growth trajectories using a diffusion-based time-series formulation. Ultrasound-derived follicle diameters are transformed into a structured ovarian state representation incorporating clinically relevant size bins and leading follicle measurements. A recurrent temporal encoder summarizes historical follicular dynamics and medication exposure into a latent response embedding. Conditioned on this embedding, a diffusion model defines a distribution over plausible future follicular trajectories. This work presents the conceptual and mathematical formulation of the framework. Empirical validation and large-scale evaluation are ongoing.