CryoDETR: a Deformable DETR-Based Method for Particle Picking in Cryo-EM Micrographs

In the workflow of single-particle cryo-electron microscopy structure reconstruction, particle picking remains a bottleneck, particularly for micrographs with low signal-to-noise ratio, extremely low contrast, and particle-dense distributions. We propose CryoDETR, an end-to-end particle picking framework based on Deformable DETR. While preserving the simplicity of the original architecture, it introduces two enhancements tailored for cryo-EM: (1) we design a Hybrid Attention Encoder in the encoder, which combines dense self-attention and deformable attention; only the last-scale feature is processed by dense attention to integrate global context and enlarge the effective receptive field for sparse sampling, reducing unstable feature representation of deformable encoders under complex backgrounds; (2) we implement mixed O2O–O2M supervision in the decoder by attaching a lightweight O2M auxiliary head to the intermediate query features produced by cross-attention, explicitly supervising object queries to improve candidate prediction quality. Experiments on five representative CryoPPP subsets show that CryoDETR achieves the highest average F1 score among the compared methods, improving average recall by 7.6% points and average F1 score by 4.1% points, with particularly clear gains on low-contrast and particle-dense subsets. Our code is available at https://github.com/Lily00725/CryoDETR.