Swin-cryoEM: Multi-class cryo-electron micrographs single particle mixed detection method

Area of Science:

• Cryo-electron microscopy
• Image analysis
• Structural biology

Background:

• Cryo-electron micrograph images present significant challenges including noise, impurities, and irregular particle features.
• Accurate single particle picking from diverse micrograph types remains a key challenge in the field.
• Existing methods struggle with adaptability across varied image characteristics.

Purpose of the Study:

• To develop an adaptive single particle detection model for cryo-electron microscopy.
• To improve the accuracy and generalization of particle picking from complex micrographs.
• To provide high-quality data for cryo-electron tomography 3D reconstruction.

Main Methods:

• Utilized the MixUp enhancement algorithm for pre-processing.
• Developed a feature perception network using Swin Transformer with channel self-attention.
• Incorporated PReLU activation and Cross-Entropy with softmax for classification.

Main Results:

• The Swin-cryoEM model achieved optimal Average Precision of 95.5% in training.
• Demonstrated superior single particle picking performance in prediction compared to mainstream models.
• Showcased enhanced network tolerance to noise and improved inter-particle information exchange.

Conclusions:

• Swin-cryoEM effectively addresses the challenge of adaptive single particle picking in cryo-electron microscopy.
• The model significantly improves detection accuracy and generalization ability.
• Provides robust data support for high-resolution 3D reconstruction in structural biology.