DEEP GENERATIVE STORM MODEL FOR DYNAMIC IMAGING.

Area of Science:

• Medical Imaging
• Machine Learning
• Image Reconstruction

Background:

• Dynamic imaging often requires high data acquisition, leading to undersampling challenges.
• Existing deep learning methods typically need large fully-sampled datasets for training, which are often unavailable.

Purpose of the Study:

• To introduce a novel generative smoothness regularization on manifolds (SToRM) model for dynamic image recovery.
• To address the limitations of existing methods in handling highly undersampled data and the lack of training data.

Main Methods:

• A generative framework representing image time series as a smooth non-linear function of latent vectors using a deep convolutional neural network (CNN).
• Joint estimation of CNN generator parameters and latent vectors from undersampled measurements via stochastic gradient descent.
• Penalization of generator gradient norm and latent vector temporal gradients to enforce smoothness.

Main Results:

• Significant reduction in memory demand compared to previous analysis-based SToRM models.
• Introduction of spatial regularization inherent to the CNN model.
• Development of efficient progressive approaches to minimize computational complexity.

Conclusions:

• The generative SToRM model effectively recovers dynamic image data from highly undersampled measurements.
• The method offers reduced memory requirements and enhanced spatial regularization without extensive training data.
• The approach is computationally efficient, making it suitable for practical applications.