Missing Wedge Completion via Unsupervised Learning with Coordinate Networks.

Area of Science:

• Structural biology
• Biophysics
• Computational imaging

Background:

• Cryogenic electron tomography (cryo-ET) provides nanoscale 3D imaging of biological samples.
• The missing wedge artifact, caused by incomplete angular data, degrades reconstruction quality.
• Supervised deep learning methods (CNNs) reduce artifacts but require extensive, representative pretraining data, risking inaccuracies.

Purpose of the Study:

• To introduce and evaluate an unsupervised learning approach for cryo-ET reconstruction.
• To overcome limitations of supervised methods, specifically pretraining requirements and data scarcity issues.
• To improve 3D reconstruction quality by mitigating the missing wedge artifact.

Main Methods:

• Developed a proof-of-concept unsupervised learning method using coordinate networks (CNs).
• Optimized network weights directly against input projections, eliminating the need for pretraining.
• Assessed performance using in silico data with voxel-based image quality metrics and a directional Fourier Shell Correlation (FSC) metric.

Main Results:

• The unsupervised CN approach reduced reconstruction runtime by 3-20× compared to supervised methods.
• Demonstrated improved shape completion and significant reduction of missing wedge artifacts.
• Achieved better image quality metrics in real space and via the novel directional FSC metric.

Conclusions:

• Unsupervised learning with CNs offers a viable alternative to supervised methods for cryo-ET reconstruction, especially when training data is limited.
• This approach accelerates the reconstruction process and enhances 3D structural detail.
• The study provides insights into both supervised and unsupervised deep learning strategies for advancing cryo-ET data processing.