Quantized neural representation for lossy cryo-EM compression

Area of Science:

• Structural biology
• Computational imaging
• Data compression

Background:

• Cryo-electron microscopy (cryo-EM) generates large 3D density volumes crucial for structural biology.
• Increasing data scale challenges storage, transmission, and interactive exploration of cryo-EM datasets.
• Efficient compression methods are needed to manage large cryo-EM data.

Purpose of the Study:

• To propose a novel hybrid quantized implicit neural representation (INR) method for cryo-EM volume compression.
• To enable efficient on-demand access to compressed cryo-EM data.
• To evaluate the effectiveness of the proposed method against existing techniques.

Main Methods:

• Developed a hybrid quantized implicit neural representation (INR) technique.
• Compressed cryo-EM volumes using the proposed method.
• Benchmarked against traditional compression and a classic INR compressor.
• Assessed compression efficiency, visual quality, and performance on specific cryo-EM tasks.

Main Results:

• The quantized INR method demonstrated superior storage efficiency compared to traditional methods.
• Achieved high fidelity relevant to cryo-EM analysis tasks, including structure identification and inspection.
• Outperformed a classic INR compressor in compression and quality metrics.

Conclusions:

• The hybrid quantized INR method offers an effective solution for compressing large cryo-EM data.
• This approach balances storage efficiency with the fidelity required for structural biology applications.
• An interactive tool and guidelines are provided to aid users in selecting optimal compression strategies.