Self-supervised next view prediction for limited-angle optical projection tomography.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Optical Engineering

Background:

• Optical projection tomography (OPT) reconstructs 3-D structures from 2-D projections, typically requiring a wide angular range (π radians) for reliable results.
• Limited-angle tomography, using fewer projection angles, suffers from missing wedge information, leading to unsatisfactory reconstructions.
• Existing methods struggle to compensate for the information loss in limited-angle scenarios.

Purpose of the Study:

• To develop a novel view prediction technique for limited-angle optical projection tomography.
• To computationally extend the angular range of captured projections without requiring prior label data.
• To improve the quality of 3-D reconstructions from incomplete projection data.

Main Methods:

• A self-supervised learning approach was employed to learn the relationships between captured limited-angle views.
• Unseen projection views were computationally synthesized based on the learned relationships.
• The technique was integrated with an optical tomography system to validate its performance.

Main Results:

• The proposed approach successfully extended the angular range of projections from 60° to nearly 180°.
• High-quality 3-D reconstructions were achieved even with highly incomplete initial measurements.
• The method demonstrated robust generation of new projections for unknown biological samples.

Conclusions:

• The novel view prediction technique effectively overcomes the limitations of missing wedge information in limited-angle tomography.
• This self-supervised method enables high-fidelity 3-D reconstructions from sparse projection data.
• The approach holds significant potential for advancing biological sample analysis using optical tomography.