SECONDGRAM: Self-conditioned diffusion with gradient manipulation for longitudinal MRI imputation

While individual MRI snapshots provide valuable insights, the longitudinal progression in repeated MRIs often holds more significant diagnostic and prognostic value. However, a scarcity of longitudinal datasets, comprising paired initial and follow-up scans, hinders the application of machine learning for crucial sequential tasks. We address this gap by proposing self-conditioned diffusion with gradient manipulation (SECONDGRAM) to generate absent follow-up imaging features, enabling predictions of MRI developments over time and enriching limited datasets through imputation. SECONDGRAM builds on neural diffusion models and introduces two key contributions: self-conditioned learning to leverage much larger, unlinked datasets and gradient manipulation to combat instability and overfitting in a low-data setting. We evaluate SECONDGRAM on the UK Biobank dataset and show that it not only models MRI patterns better than existing baselines but also enhances training datasets to achieve better downstream results over naive approaches.