Progressively volumetrized deep generative models for data-efficient contextual learning of MR image recovery

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Image Reconstruction

Background:

• Magnetic Resonance Imaging (MRI) provides versatile imaging contrasts but is limited by scan time.
• Undersampled data acquisition is a common strategy to reduce MRI scan time.
• Current recovery methods either process data volumetrically (capturing global context but complex) or cross-sectionally (simpler but ignoring longitudinal context).

Purpose of the Study:

• To introduce a novel progressive volumetrization strategy for generative models (ProvoGAN) to address limitations in MRI data recovery.
• To develop a method that balances global context capture with model complexity for improved MRI reconstruction.
• To enhance the quality and diversity of MRI data acquired under time constraints.

Main Methods:

• ProvoGAN employs a progressive volumetrization strategy, decomposing complex volumetric recovery into ordered cross-sectional mappings.
• The model serially processes data across individual rectilinear dimensions, optimizing task order.
• Generative models are utilized within this progressive framework for image recovery.

Main Results:

• ProvoGAN effectively captures global context and recovers fine-structural details across all dimensions.
• The method maintains low model complexity and demonstrates improved learning behavior compared to existing approaches.
• ProvoGAN achieved superior performance on mainstream MRI reconstruction and synthesis tasks.

Conclusions:

• ProvoGAN offers a superior approach to MRI image recovery from undersampled data.
• The progressive volumetrization strategy effectively balances model complexity and performance.
• This method has the potential to significantly improve MRI acquisition efficiency and data quality.