The problem — Pre-trained deep-learning models for bioimage analysis are usually distributed as Python scripts and weight files. For the majority of microscopy users working inside Fiji or ImageJ, there is no practical path from a model repository to a running inference step — the tool they use daily and the ecosystem where models live are completely disconnected.
What it is / how it works — deepImageJ is an open-source Fiji plugin (Gómez-de-Mariscal et al., Nature Methods 2021) that bridges that gap. It runs models packaged in the BioImage Model Zoo format (bioimage.io) directly inside ImageJ, guiding the user through the full inference workflow: image pre-processing (normalisation, tiling, channel ordering), model prediction via a local TensorFlow or PyTorch runtime, and post-processing (thresholding, connected-components labelling). The plugin reads the model's description file — which specifies expected input/output shapes, pre/post-processing steps, and test tensors — so the user does not need to know the model architecture to run it correctly. Supported task types include semantic segmentation, instance segmentation, denoising, virtual staining, and object detection. Models exported from ZeroCostDL4Mic in bioimage.io format are directly deployable via deepImageJ.
Where it fits — deepImageJ is an inference consumer at the segment and preprocess stages — the ImageJ-native entry point to the model zoo ecosystem. It makes zoo models usable in the interactive platform that most biologists already have open.
Where it breaks — deepImageJ runs inference on full images or tiles; it does not offer training. Local Java/Python runtime setup can fail on machines with conflicting environments, and the plugin is sensitive to model format versions — a model packaged for an older bioimage.io spec may not load in newer releases. Tiling behaviour must be set correctly for very large images; incorrect overlap padding produces seam artifacts in the output.