QwenCryoMarker: a universal post-processing framework for contamination-aware particle cleaning

Cryo-electron microscopy (cryo-EM) micrographs are frequently contaminated by carbon edges, ice crystals, ethane bubbles and other high-contrast artifacts. These contaminants trigger abundant false positives in automated particle pickers, severely hampering downstream 3D reconstruction. Existing methods either avoid contamination implicitly (requiring dataset-specific tuning) or rely on rule-based filters that fail on complex contamination patterns. Here, we present QwenCryoMarker, a universal post-processing framework that converts the outputs of arbitrary particle pickers into clean, high-precision particle sets. Our pipeline consists of two core stages: (i) a visual large model (Qwen-Image-Edit-2511) fine-tuned via supervised learning to generate pixel-accurate binary contamination masks from raw micrographs and (ii) a lightweight contamination-aware filtering module that discards particles when the contamination proportion within their surrounding circular region exceeds a predefined threshold ratio. Contamination is defined as all micrograph regions unsuitable for reliable particle picking and subsequent 3D reconstruction, including carbon edges, ice crystals, ethane bubbles, miscellaneous debris and dense protein aggregates. The framework features plug-and-play deployment: it requires no per-dataset parameter tuning or extra retraining, and maintains compatibility with classical pickers (blob detection, template matching) as well as deep learning-based pickers (Topaz, crYOLO etc.). We validate QwenCryoMarker on five diverse CryoPPP benchmark datasets across four representative particle pickers. Quantitatively, our method consistently boosts precision with an average absolute gain of 0.009 and lifts the F₁-score, while recall only drops slightly by an average of 0.008. Segmentation benchmarking shows our model reaches a mean intersection over union (IoU) of 0.629, surpassing that of the state-of-the-art MicrographCleaner (0.551) by 14.2%. We further compare against multiple segmentation baselines: U-Net (0.448), DeepLabV3+ (0.488), SAM (0.475), ASOCEM (0.195) and IceBreaker (0.433). A downstream reconstruction case study on EMPIAR-10017 verifies that particles filtered by QwenCryoMarker yield cleaner 2D class averages and higher resolution 3D density maps (3.88 versus 3.97 Å). Qualitative visualization also confirms that QwenCryoMarker stably eliminates false particles located on carbon films and ice crystals, independent of the upstream particle-picking algorithm. By encapsulating contamination suppression as a universal, model-agnostic post-processing module, QwenCryoMarker offers a practical, robust, easy-to-deploy toolkit that greatly improves particle-set quality without modifying existing cryo-EM workflows. The framework is fully open-source and can be seamlessly integrated into mainstream cryo-EM processing pipelines.