Dynamic Imaging Using a Deep Generative SToRM (Gen-SToRM) Model.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Dynamic image data acquisition often involves undersampled measurements, leading to challenges in reconstruction.
• Traditional deep learning methods, like convolutional neural networks (CNNs), typically require extensive fully sampled training data.
• Existing manifold models for dynamic image reconstruction can be computationally intensive and memory-demanding.

Purpose of the Study:

• To introduce a novel generative model for dynamic image reconstruction from highly undersampled data.
• To develop a method that reduces memory requirements and computational complexity compared to traditional manifold models.
• To improve the quality of reconstructed dynamic images while minimizing the need for fully sampled training datasets.

Main Methods:

• The proposed model, Smoothness Regularization on Manifolds (SToRM), uses a deep convolutional neural network (CNN) to represent non-linear transformations from low-dimensional latent vectors.
• It jointly estimates generator parameters and latent vectors directly from undersampled measurements.
• Spatial regularization is achieved by penalizing the norm of the non-linear mapping's gradients, and temporal smoothness is enforced by penalizing latent vector gradients.

Main Results:

• The SToRM model enables dynamic image recovery from highly undersampled measurements.
• It achieves significant reductions in memory demand compared to conventional manifold models.
• The method demonstrates improved image quality and reconstruction performance through efficient progressive training-in-time and approximate cost functions.

Conclusions:

• The SToRM model offers an effective solution for dynamic image reconstruction from undersampled data.
• This generative approach provides superior image quality and memory efficiency.
• The computational optimizations facilitate faster and more robust image reconstructions.