Video frame interpolation neural network for 3D tomography across different length scales

Area of Science:

• Multidisciplinary scientific imaging
• Advanced materials characterization
• Medical imaging analysis

Background:

• Three-dimensional (3D) tomography is crucial across science and medicine.
• Image resolution limitations, often anisotropic, hinder data precision.
• Enhancing tomographic data quality is essential for accurate analysis.

Purpose of the Study:

• To apply a neural network for enhancing 3D tomographic image resolution.
• To achieve isotropic cubic-voxel resolution in tomographic reconstructions.
• To demonstrate the versatility of the enhancement method across diverse scientific fields.

Main Methods:

• Utilizing a neural network originally designed for video-frame interpolation.
• Applying the method to focused ion beam-scanning electron microscopy (FIB-SEM) data.
• Testing the approach on magnetic resonance imaging (MRI) and X-ray computed tomography (CT) datasets.

Main Results:

• Achieved cubic-voxel resolution, significantly improving 3D image quality.
• Demonstrated successful application in materials science (graphene), neuroscience (brain MRI), and medical imaging (abdomen CT).
• Validated accuracy through computer-vision metrics and improved precision of physical quantities (e.g., porosity, diffusivity).

Conclusions:

• The neural network-based approach offers a versatile image-augmentation strategy.
• Optimizes 3D tomography acquisition by enhancing resolution and preserving information content.
• Provides a powerful tool for improving data quality in various scientific imaging domains.