TransFACT predicts embryo viability using day-4 time-lapse video analysis

Current bovine embryo transfer practice relies on a single expert assessment on day seven post-insemination, leading to high rates of pregnancy loss. Time-lapse videomicroscopy captures detailed early development but is difficult to exploit due to complex motion patterns and time-consuming analysis. To address this, researchers introduce TransFACT, a transformer-based framework that models early developmental stages from 2D time-lapse videos over the first four days. It combines frame-level temporal features with stage-level representations, using developmental stages as auxiliary supervision to predict transferability on day four. The model leverages an existing action recognition method, achieving superior performance over competitors in predicting embryo transferability.

TransFACT's ability to assess embryo viability three days earlier than standard practice could significantly reduce pregnancy losses and improve breeding efficiency in cattle. The technique also shows potential for adaptation to other species, including human IVF, where early, non-invasive prediction of embryo quality remains a major challenge. By automating the analysis of time-lapse microscopy, TransFACT offers a scalable, objective tool for embryologists, moving from subjective day-seven snapshots to data-driven, temporal decisions. The paper is accepted at ICIP 2026 (Tampere, Finland) and available on arXiv.