Few-Shot Classification of Cryo-EM Micrographs Using Triplet Loss Embeddings

Micrograph quality assessment in cryo-electron microscopy (cryo-EM) presents a significant challenge: users must either manually screen thousands of micrographs or expend substantial computational resources processing potentially low-quality data. While few-shot learning has been applied to particle picking and subtomogram classification in cryo-EM, its application to micrograph-level quality assessment remains largely unexplored. We present a framework combining few-shot learning with cryo-EM micrograph classification using triplet loss embeddings. By combining real-space and Fourier-space information in an embedding network trained with triplet loss, we achieve competitive performance across multiple EMPIAR datasets using as few as 1-5 labeled examples per class. Our approach demonstrates improvements over traditional cross-entropy training, which often collapses to predicting only the majority class in the few-shot regime. These results suggest a practical framework for rapid adaptation of automated micrograph screening to new experimental conditions with minimal manual labeling.