NeRP: Implicit Neural Representation Learning With Prior Embedding for Sparsely Sampled Image Reconstruction.

Area of Science:

• Medical Imaging
• Computational Imaging
• Machine Learning

Background:

• Image reconstruction is an inverse problem crucial for creating computational images from sensor data.
• Reconstructing images from sparsely sampled data presents significant challenges due to limited measurements.

Purpose of the Study:

• To propose a novel methodology, Neural Representation learning with Prior embedding (NeRP), for reconstructing images from sparsely sampled measurements.
• To demonstrate the generalizability and robustness of NeRP across different imaging modalities and applications.

Main Methods:

• NeRP leverages implicit neural representation learning combined with prior image information.
• The method integrates the physics of sparsely sampled measurements to generate an accurate image representation.
• Unlike traditional deep learning methods, NeRP requires only a prior image and sparse measurements, not large-scale training datasets.

Main Results:

• NeRP successfully reconstructs computational images from limited, sparsely sampled measurements.
• The methodology is shown to be generalizable, applicable to various imaging modalities including computed tomography (CT) and magnetic resonance imaging (MRI).
• NeRP demonstrates robustness in capturing subtle image changes critical for applications like tumor progression assessment.

Conclusions:

• NeRP offers a powerful and data-efficient approach to image reconstruction from sparse data.
• The method's versatility across imaging modalities and its ability to detect subtle changes highlight its potential impact in medical diagnostics.
• NeRP represents a significant advancement in overcoming the challenges of sparse data in image reconstruction.