Instance segmentation models

[normanrz]

Detailed Description

Add the capability to train instance segmentation models and run predict+segment. The AI analysis modal needs to be extended to add a new instance segmentation category in the train model tab. The backend needs to coordinate the new category with the worker.

Context

See https://github.com/scalableminds/voxelytics/issues/3858

[AndreaNathansen]

Support for training and segmenting with instance models is already implemented in the VX worker (except for flexible distance threshold for instance inference). The respective interface functions that have to be called by WK are:

• train_instance_model for training
• infer_instances which has to be called with the respective model ID for instance segmentation with custom model. In this regard, the neuron segmentation with custom model can also use infer_neurons with the model ID instead of infer_with_model (which calls infer_neurons under the hood anyway and is only kept for compatibility reasons).

Currently, the default config parameters for instance segmentation are configured for nuclei. To be more flexible and allow segmentation of smaller objects, the distances would have to be configurable in WK:

• max_distance_nm for the instance training
• seed_generator_distance_threshold for the instance segmentation with custom model (needs to be implemented in the worker as well, see scalableminds/voxelytics#4146)

[normanrz]

• max_distance_nm for the instance training

• seed_generator_distance_threshold for the instance segmentation with custom model (needs to be implemented in the worker as well, see scalableminds/voxelytics#4146)

Could we auto-detect these params or have presets?

[AndreaNathansen]

Could we auto-detect these params or have presets?

In the last VX sprint meeting, we decided to make it configurable at first (because that is faster to implement) and invest the time to add heuristics for auto-detection or presets later.

[AndreaNathansen]

The configurable seed_generator_distance_threshold for infer_instances in the VX worker is now implemented (https://github.com/scalableminds/voxelytics/pull/4161).