OncoFlow: Context-Conditioned Flow Matching for Brain MRI Forecasting

Forecasting how brain tumors will evolve on MRI could enable earlier assessment of disease progression and treatment response. This remains challenging because tumor evolution is rapid and heterogeneous, shaped not only by prior imaging but also by treatment interventions and underlying tumor biology. Existing methods often fail to capture these clinically meaningful changes, while conventional evaluation metrics are dominated by stable anatomy and can therefore reward near-identity predictions. We introduce OncoFlow, a clinically conditioned 3D flow-matching framework for patient-specific MRI forecasting. Its source distribution spatially interpolates between the latest MRI and Gaussian noise, progressively reducing dependence on the observed anatomy as the forecast horizon increases. A multipath velocity network integrates visit-aligned longitudinal imaging memory with global conditioning on treatment, genomics, forecast horizon, and generative time. To evaluate whether predicted scans capture disease evolution rather than merely reconstruct static anatomy, we introduce Rpattern, a Residual Pattern Alignment metric that quantifies the spatial fidelity of predicted changes. Across three longitudinal brain-tumor datasets, OncoFlow achieves state-of-the-art overall performance, consistently improving change fidelity while remaining competitive in whole-image reconstruction. Qualitative and ablation analyses show that soft source conditioning and horizon-dependent decay suppress spurious changes in healthy tissue without collapsing to near-identity forecasts. Crucially, conditioning on treatment and genomic information further improves forecasting performance over imaging history alone, demonstrating that non-imaging clinical variables provide complementary signals about future tumor evolution. Together, these findings establish clinically conditioned generative modeling as a promising route towards forecasts that preserve patient anatomy while remaining sensitive to meaningful, treatment- and biology-driven change.