Variational Manifold Learning From Incomplete Data: Application to Multislice Dynamic MRI.

Area of Science:

• Medical Imaging
• Machine Learning
• Data Science

Background:

• Deep learning manifold learning, including variational autoencoders (VAEs), typically requires fully sampled data.
• Acquiring fully sampled data is challenging in dynamic and high-resolution Magnetic Resonance Imaging (MRI).
• This limitation hinders the application of advanced deep learning techniques in critical medical imaging scenarios.

Purpose of the Study:

• To develop a novel variational approach for manifold learning from undersampled MRI data.
• To enable accurate reconstruction and alignment of dynamic imaging datasets without requiring complete sampling.
• To address the data acquisition challenges in high-resolution and free-breathing, ungated cardiac MRI.

Main Methods:

• A variational autoencoder (VAE) framework is proposed, utilizing a decoder fed by latent vectors.
• A conditional density of latent vectors is estimated from undersampled measurements using back-propagation, approximating the unavailable fully sampled data.
• The method is applied to joint alignment and recovery of multi-slice, free-breathing, and ungated cardiac MRI data from highly undersampled measurements.

Main Results:

• The proposed scheme successfully learns manifolds from undersampled data, overcoming limitations of traditional VAEs.
• Experimental results demonstrate effective joint alignment and reconstruction of dynamic cardiac MRI.
• The approach shows utility in improving the quality of reconstructions from highly undersampled measurements.

Conclusions:

• The novel variational approach enables effective manifold learning from undersampled MRI data.
• This method significantly advances dynamic imaging alignment and reconstruction in scenarios with incomplete data.
• The technique holds promise for improving MRI applications where full data acquisition is impractical.