A Novel Contrastive Loss and Clustering Approach for Particle Detection in Cryo-EM

Detecting particles in cryo-electron microscopy (cryo-EM) is labor-intensive and time-consuming. Meanwhile, existing datasets often suffer from false positives and negatives in annotations, making the problem more challenging. To tackle these issues, we propose a novel detection solution enhanced by integrating specifically designed contrastive loss and a novel deep clustering mechanism. The contrastive loss improves the model's ability to distinguish between particles and non-particles by maximizing the separation of their learned embeddings. The deep clustering mechanism further organizes these embeddings into distinct groups, effectively grouping similar particles together. This way, our method effectively addresses the challenges of noisy labels and ensures more robust learning and better particle detection. Furthermore, we fully annotated a previously partially labeled test set, ensuring a more comprehensive evaluation of model performance. The experiment results show the effectiveness of tailored loss functions and the importance of addressing the unique challenges in cryo-EM particle detection.