Imaging Models
Microscopy, fluorescence imaging, and cryo-EM analysis
168 models in this category
What biological imaging foundation models do
Biological imaging foundation models are trained on large-scale microscopy datasets — spanning fluorescence confocal imaging, cryo-electron microscopy, cell painting, brightfield, and live-cell timelapse — to learn visual representations that generalize across instruments, imaging protocols, and specimen types. Models like Cellpose learn generalist cell and nucleus segmentation that transfers across cell lines and imaging conditions without retraining, while others focus on image restoration, super-resolution, or phenotypic profiling from high-content screening data. The defining feature of this class is pretraining that reduces the annotation burden historically required for each new imaging experiment.
Applications in microscopy and drug screening
Cell segmentation is the most widely adopted application, with Cellpose in particular becoming near-standard infrastructure for fluorescence microscopy analysis. Phenotypic profiling from cell painting assays — where cells are stained with a panel of dyes and imaged to capture morphological features — has been substantially advanced by foundation model embeddings that capture biological signal more efficiently than classical handcrafted features. Cryo-EM particle picking and 3D reconstruction represent a specialized but high-impact domain where deep learning has largely replaced manual curation, shortening the path from raw micrographs to solved structures.
Notable Models
Top-rated imaging models from our evaluations
Generalist deep learning algorithm for cell and nucleus instance segmentation using simulated diffusion flows, without per-dataset retraining.
Biomedical vision-language model trained contrastively on 15M PubMed Central figure-caption pairs for zero-shot classification, retrieval, and VQA.
Promptable 3D foundation model for volumetric CT segmentation, covering over 200 anatomical categories through point, box, and free-text prompts.
Promptable foundation model for universal medical image segmentation, fine-tuned from SAM on 1.57M image-mask pairs across 10 imaging modalities.
Generalist medical multimodal LLM for image understanding, visual question answering, and report generation across twelve-plus imaging modalities.
Virtual staining models that translate label-free light microscopy into fluorescent-equivalent predictions of nuclei and plasma membranes.
Frequently asked questions
What is a biological imaging foundation model?
A biological imaging foundation model is a neural network pretrained on large collections of biological images — microscopy, cryo-EM, histology, or cell painting — to learn visual representations that transfer across imaging tasks without task-specific training from scratch. These models enable cell segmentation, image restoration, phenotypic profiling, and structural reconstruction with reduced annotation requirements. Cellpose is among the most widely adopted examples for cell and nucleus segmentation.
How does Cellpose differ from a general image segmentation model?
Cellpose uses a gradient flow representation that is particularly well-suited to the circular, overlapping morphology of cells and nuclei, and it was pretrained on diverse microscopy images spanning many cell types and imaging modalities. General image segmentation models like SAM (Segment Anything Model) were trained predominantly on natural images and require adaptation or fine-tuning to perform well on microscopy data, where textures, boundaries, and object statistics differ substantially from everyday photographs.
What is cell painting and why are foundation models useful for it?
Cell painting is a high-content imaging assay where cells are stained with five to eight fluorescent dyes targeting distinct cellular compartments — nucleus, mitochondria, endoplasmic reticulum, actin — and imaged to produce a rich morphological profile. Foundation models trained on cell painting data can extract compact embeddings that capture drug mechanism of action, toxicity, and genetic perturbation effects more efficiently than classical feature extraction pipelines, enabling large-scale phenotypic screening.
Are cryo-EM processing tools considered foundation models?
Some cryo-EM tools — particularly neural network-based particle pickers and ab initio reconstruction approaches trained on large datasets of micrographs — fit the foundation model framing, though the field often uses different terminology. Models pretrained on diverse cryo-EM data that transfer to new specimens with minimal fine-tuning are tracked by bio.rodeo; classic algorithmic tools without a learned representation component are generally excluded.