Delta-Diffusion: Modeling Longitudinal Brain Amyloid-PET Trajectories via Conditional Poisson Diffusion Bridge

Researchers introduce Delta-Diffusion, a novel AI framework using conditional Poisson Diffusion Bridges to synthesize longitudinal brain PET imaging for tracking amyloid accumulation in neurodegenerative diseases. The method addresses limitations of existing generative models by anchoring predictions to baseline patient scans and incorporating clinical progression patterns, potentially reducing the need for costly repeated imaging procedures.

Delta-Diffusion represents a meaningful advancement in medical AI by tackling a specific clinical bottleneck: longitudinal brain imaging for Alzheimer's disease monitoring. Traditional PET imaging, while diagnostically valuable, imposes significant burdens through high costs and cumulative radiation exposure, limiting its practical deployment in clinical settings. The research team identified that existing generative models struggle with two critical failures—identity drift (losing subject-specific characteristics) and baseline replication bias (failing to model true disease progression).

The innovation lies in reformulating the synthesis problem as a conditional diffusion process anchored to baseline scans, transforming it from pure generation into trajectory modeling. By incorporating Poisson perturbations grounded in PET imaging physics and using adaptive modulation based on clinical intervals, the framework captures the heteroscedastic noise inherent to medical imaging. The architecture's emphasis on volume-of-interest balanced objectives ensures the model prioritizes high-risk amyloid accumulation regions rather than averaging across the entire brain.

For healthcare systems and pharmaceutical development, this approach offers practical value by potentially reducing imaging frequency while maintaining disease tracking capability. The validation across 542 subjects from two cohorts suggests reasonable generalizability. However, the work remains within academic research; clinical translation would require regulatory approval and validation against prospective patient outcomes.

The broader significance centers on AI's role in computational medicine—moving beyond pattern recognition toward physics-informed generative modeling that respects domain constraints. Future developments should examine transferability to other neurodegenerative conditions and whether computational predictions can substitute for actual clinical imaging in longitudinal studies.

• →Delta-Diffusion uses physics-grounded Poisson diffusion anchored to baseline scans to model disease progression rather than generate images from noise
• →The framework addresses identity drift and baseline bias, longstanding failures in generative models applied to medical imaging
• →Validation on 542 subjects demonstrates superior performance in capturing longitudinal amyloid deposition patterns
• →Potential clinical applications include reducing radiation exposure and operational costs in Alzheimer's disease monitoring
• →The approach exemplifies physics-informed AI that respects domain constraints rather than purely data-driven generation