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

Area of Science:

• Medical Imaging
• Machine Learning
• Artificial Intelligence

Background:

• Longitudinal data from repeated Magnetic Resonance Imaging (MRI) scans offer superior diagnostic and prognostic value compared to single scans.
• A significant limitation in applying machine learning to sequential MRI analysis is the scarcity of paired longitudinal datasets.
• Existing methods struggle with the instability and overfitting common in low-data scenarios for generating sequential medical imaging features.

Purpose of the Study:

• To address the data scarcity challenge in longitudinal MRI analysis by generating absent follow-up imaging features.
• To enable accurate predictions of MRI developments over time and enrich limited datasets through imputation.
• To improve the application of machine learning for crucial sequential tasks in medical imaging.

Main Methods:

• Proposed Self-Conditioned Diffusion with Gradient Manipulation (SECONDGRAM), a novel neural diffusion model.
• Incorporated self-conditioned learning to leverage larger, unlinked MRI datasets.
• Utilized gradient manipulation to enhance stability and mitigate overfitting in low-data settings.

Main Results:

• SECONDGRAM demonstrated superior performance in modeling MRI patterns compared to existing baseline methods on the UK Biobank dataset.
• The method effectively generated absent follow-up imaging features, enabling predictions of MRI progression.
• Enriching training datasets with SECONDGRAM-imputed data led to improved downstream machine learning task performance.

Conclusions:

• SECONDGRAM effectively addresses the scarcity of longitudinal MRI data by generating realistic follow-up scans.
• The proposed method enhances the utility of limited datasets for machine learning applications in medical imaging.
• This approach holds significant potential for improving diagnostic and prognostic capabilities through advanced MRI analysis.