Unsupervised deep image prior for sparse-view and limited-angle electron tomography.

Area of Science:

• Materials Science
• Imaging Science
• Computational Science

Background:

• Electron tomography (ET) is crucial for 3D nanomaterial characterization.
• Conventional algorithms struggle with limited-angle and sparse-view data, causing artifacts like blurring and streaking.
• These artifacts degrade the quality and interpretability of 3D reconstructions.

Purpose of the Study:

• To introduce a novel unsupervised deep learning (DL) approach, deep image prior (DIP), for degraded tomography acquisitions.
• To evaluate DIP's performance against supervised methods using simulated and experimental data.
• To demonstrate DIP's capability in enabling reliable 3D quantification under challenging imaging conditions.

Main Methods:

• Implementation of deep image prior (DIP), an unsupervised deep learning technique.
• Testing DIP on simulated datasets with limited angular range (e.g., 60°) and large tilt increments (e.g., 10°).
• Application of DIP to experimental electron tomography data.

Main Results:

• DIP achieves performance comparable to supervised DL methods, even without training datasets.
• The method effectively mitigates artifacts associated with limited-angle and sparse-view tomography.
• Reliable 3D quantification was achieved on experimental data under degraded conditions.

Conclusions:

• Deep image prior (DIP) offers a powerful unsupervised solution for improving electron tomography reconstructions.
• DIP enables accurate 3D nanomaterial characterization from highly degraded tomographic data.
• This approach has broad applicability across various materials and imaging modalities.