DINOSim: Training-Free Semantic Segmentation on Microscopy Images

We present DINOSim, a training-free method for segmenting objects in microscopy images without the need for large annotated datasets or additional model training. DINOSim builds on the pretrained DINOv2 image encoder, which captures semantic information from images. By comparing the encoder's features of image patches to those of a user-selected reference, DINOSim generates pseudo-labels that guide semantic segmentation. Subsequently, a k-nearest neighbors framework is used to refine predictions across new images. Our experiments show that DINOSim can effectively segment previously unseen objects in diverse microscopy datasets, offering performance comparable to supervised approaches while avoiding the need for costly manual labeling. We also investigate how different choices of user prompt selection and model size affect accuracy and generalization. To make the method widely accessible, we provide an open-source napari plugin (github.com/AAitorG/napari-DINOSim), enabling researchers to apply DINOSim to their own data easily. Overall, DINOSim offers a fast, flexible and practical solution for bioimage analysis, particularly valuable in resource-constrained settings.