ProxNF: Neural Field Proximal Training for High-Resolution 4D Dynamic Image Reconstruction

Area of Science:

• Biomedical Imaging
• Computational Imaging
• Machine Learning in Medical Imaging

Background:

• Accurate spatiotemporal image reconstruction is crucial for biomedical research but hindered by data undersampling and high computational demands.
• Existing methods struggle with memory constraints for high-resolution 3D+time datasets.
• Neural fields (NFs) offer a promising solution by representing dynamic imaging data as continuous functions.

Purpose of the Study:

• To introduce ProxNF, a novel neural field training approach for efficient spatiotemporal image reconstruction.
• To leverage proximal splitting methods to optimize neural field parameter estimation.
• To overcome challenges of data incompleteness and computational burden in dynamic imaging.

Main Methods:

• ProxNF utilizes proximal splitting to decouple imaging operator computations from neural network parameter updates.
• The method evaluates gradients in the image domain using subsampled data.
• A fully supervised learning strategy is employed for updating neural field parameters.

Main Results:

• Demonstrated effectiveness in two numerical phantom studies.
• Successfully applied to in-vivo tumor perfusion imaging in small animal models.
• Validated using dynamic contrast-enhanced photoacoustic computed tomography (DCE PACT).

Conclusions:

• ProxNF offers an effective and computationally efficient solution for spatiotemporal image reconstruction.
• The proposed method shows significant potential for advancing dynamic biomedical imaging applications.
• ProxNF successfully addresses key limitations in current dynamic imaging reconstruction techniques.