Endoscopic Tool Segmentation
Performs endoscopic tool segmentation using selectable algorithm backends, including feature-based and machine learning models.
Input
One 2D image set. Only 8-bit RGB images with a single slice per image are supported.
Output
Label image with two classes: “Background” and “Endoscopic Tool”. Visualization includes label names and colors.
Description
The algorithm supports multiple segmentation backends, selectable via the parameter endoscopicToolSegmentationAlgorithm. Available backends may include feature-based methods and machine learning models (e.g., “Custom Model”).
For the “Custom Model” backend, a machine learning model, Torch, ONNX or TensorRT is loaded from a user-specified path (modelConfigurationPath). The model processes the input image set and outputs a segmentation label image. An example model is the MONAI model (model.ts) at https://huggingface.co/MONAI/endoscopic_tool_segmentation/tree/0.6.2/models/.
```yaml Version: ‘8.0’ Type: NeuralNetwork Name: CustomEndoscopicToolSegmentation Description: Engine:
Name: torch ModelFile: model.ts ForceCPU: false Verbose: false InputFields: [Image] OutputFields: [ToolSeg] DisableJITRecompile: true
PreprocessingInputFields: [Image] PredictionOutput: [Image] Sampling: - SkipUnpadding: True PreProcessing: - ResampleDims:
target_dims: 736 480 1
MakeFloat:
- NormalizeUniform:
min: 0 max: 1
PostProcessing: - Softmax: - ArgMax: - ResampleToInput: MaxBatchSize: 2 ```
Configuration parameters:
endoscopicToolSegmentationAlgorithm: Selects the segmentation backend.
modelConfigurationPath (for Custom Model): Path to the model configuration file.
Output labels:
Background (label 0)
Endoscopic Tool (label 1, color: green)
The output is a label image with these classes, suitable for visualization and further analysis.