Temporally-aware diffusion model for brain progression modelling with bidirectional temporal regularisation.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Accurate prediction of future brain structure changes from MRIs is crucial for clinical outcome assessment and disease progression analysis.
• Existing methods struggle with time-interval relationships, lack clinical utility through interpolation, and disregard essential 3D anatomical context.

Purpose of the Study:

• To propose a novel 3D Temporally-Aware Diffusion Model (TADM-3D) for accurate prediction of brain progression in MRI volumes.
• To enhance temporal modeling by integrating a pre-trained Brain-Age Estimator (BAE) and introducing Back-In-Time Regularisation (BITR).

Main Methods:

• Developed TADM-3D, a 3D diffusion model incorporating a pre-trained Brain-Age Estimator to guide MRI generation based on time intervals.
• Implemented Back-In-Time Regularisation (BITR) to train the model for bidirectional prediction (baseline to follow-up and vice-versa).
• Trained and evaluated the model on the OASIS-3 dataset and validated on the external NACC dataset.

Main Results:

• TADM-3D accurately predicts brain progression on MRI volumes, outperforming existing methods by better capturing temporal dynamics.
• The integration of BAE and BITR significantly improves the temporal accuracy and clinical utility of generated future MRIs.
• The model demonstrated strong generalisation performance on an external dataset.

Conclusions:

• TADM-3D offers a robust solution for longitudinal brain MRI prediction, addressing limitations of current approaches.
• The proposed model enhances the assessment of disease progression and clinical outcomes at the patient level.
• This work provides a valuable tool for neuroscience research and clinical practice, with code available for reproducibility.