CRISP: A modular platform for cryo-EM image segmentation and processing with Conditional Random Field

Distinguishing signal from background in cryogenic electron microscopy (cryo-EM) micrographs is a critical processing step but remains challenging owing to the inherently low signal-to-noise ratio (SNR), contaminants, variable ice thickness, and densely packed particles of heterogeneous sizes. Recent image-segmentation methods provide pixel-level precision and thus offer several advantages over traditional object-detection approaches: segmented-blob mass can be computed to suppress false-positive particles, particle centering can be improved by leveraging the full brightness profile, and irregularly shaped particles can be identified more reliably. However, low SNR makes it difficult to obtain accurate pixel-level annotations for training segmentation models, and, in the absence of systematic evaluation platforms, most segmentation pipelines still rely on ad-hoc design choices.

Here, we introduce a modular platform that automatically generates high-quality segmentation maps to serve as reference labels. The platform supports flexible combinations of segmentation architectures, feature extractors, and loss functions, and it integrates novel Conditional Random Fields (CRFs) with class-discriminative features to refine coarse predictions into fine-grained segmentations. On synthetic data, models trained with our reference labels achieve pixel-level accuracy, recall, precision, Intersection-over-Union (IoU), and ${\text{F}}_1$ scores all exceeding 90%. We further show that the resulting segmentations can be used directly for particle picking, yielding higher-resolution 3D density maps from real experimental datasets; these reconstructions match those curated by human experts and surpass the results of existing particle-picking tools. To facilitate further research, we release our methods as the open-source package CRISP, available at https://github.com/phonchi/CryoParticleSegment.