Segmentation Models

The segmentation model registry (~/.optari/config/segmentation_models.json) declares every model OPTARI can run, independent of the segmentation preset. See Segmentation for the workflow and Data & Integrations for the adapter pattern behind it.

Registry format

segmentation_models.json

{
  "models": [
    {
      "adapter_class": "UKErUSSegAdapter",
      "id": "c_unet_gastrocnemius-transverse",
      "filename": "c_unet_ta_20260928_gastrocnemius.onnx",
      "description": "trained on all MSOT2 samples, structural loss weight: 0.0",
      "input_height": 224,
      "input_width": 224,
      "class_names": {
        "0": "background",
        "1": "Haut",
        "2": "Faszie1",
        "3": "Muskel1",
        "4": "Muskel2",
        "5": "Membran",
        "7": "Vorlaufstrecke",
        "8": "SAT",
        "9": "Gel"
      },
      "default_class": "Muskel1",
      "postprocessing_config": {
        "keep_largest_per_class": [1, 3, 7, 8],
        "combine_class_groups": [[2, 6]],
        "remove_small_objects_config": [[1, 50], [3, 50], [4, 50], [7, 50], [8, 50], [9, 50], [2, 25]]
      },
      "url": "https://link.to/modelfile.onnx"
    }
  ]
}

Field	Meaning
id	Unique model identifier, referenced by segmentation presets (model_id) and by segmentation.default_model in config.json.
adapter_class	Name of the ModelAdapterBase subclass implementing this model's pre/postprocessing.
filename	ONNX weights filename, expected under ~/.optari/models/.
input_height / input_width	Expected input frame size for the model.
class_names	Class-id → name map for the model's segmentation output (max class id is 255)
default_class	Class selected by default in the Segmentation dock.
postprocessing_config	Postprocessing steps for the mask. Integers refer to class IDs: keep only the largest connected component per class, merge specific class groups, and remove small objects below a pixel threshold.
url	Model download URL. If the weights file is missing locally, OPTARI tries to download it from here on first use.

Pretrained models

OPTARI comes with pretrained models for common clinical examination sites. The weights are not bundled with the installer but are download automatically into ~/.optari/models/ the first time a given model is used, keeping the base installation small. Settings → Models shows whether a given model's weights are already installed.

Adding a custom model

See Extending OPTARI in the developer guide.