MONAI Segmentation Inference Operator

Authors: Holoscan SDK Team (NVIDIA)
Supported platforms: x86_64, aarch64
Language: Python
Last modified: August 5, 2025
Latest version: 1.1.0
Minimum Holoscan SDK version: 1.0.3
Tested Holoscan SDK versions: 2.2.0, 3.2.0
Contribution metric: Level 2 - Trusted

This segmentation operator uses MONAI transforms and Sliding Window Inference to segment medical images.

Overview

The MonaiSegInferenceOperator performs pre-transforms on input images, runs segmentation inference using a specified model, and applies post-transforms. The segmentation result is returned as an in-memory image object and can optionally be saved to disk.

Requirements

• Holoscan SDK Python package
• MONAI
• torch

Example Usage

from pathlib import Path
import torch
from monai.transforms import Compose, LoadImage, ScaleIntensity, EnsureChannelFirst
from holoscan.core import Fragment
from operators.medical_imaging.monai_segmentation_inference_operator import MonaiSegInferenceOperator
from operators.medical_imaging.core import AppContext, IOMapping, IOType, Image

# Initialize the fragment
fragment = Fragment()

# Create app context
app_context = AppContext({})

# Define transforms
pre_transforms = Compose([
    LoadImage(image_only=True),
    EnsureChannelFirst(),
    ScaleIntensity(),
])

post_transforms = Compose([
    # Add your post-processing transforms here
])

# Initialize the segmentation operator
seg_op = MonaiSegInferenceOperator(
    fragment,
    roi_size=(96, 96, 96),  # Example ROI size for 3D images
    pre_transforms=pre_transforms,
    post_transforms=post_transforms,
    app_context=app_context,
    model_name="unet",  # Example model name
    overlap=0.25,
    sw_batch_size=4,
    model_path=Path("/path/to/your/model.pt")  # Replace with your model path
)

API Reference

Python

MonaiSegInferenceOperator

Inherits from: InferenceOperator

This segmentation operator uses MONAI transforms and Sliding Window Inference.

Methods

Method	Description
__init__(fragment)	Creates a instance of this class.
setup(spec)
roi_size()	The ROI size of tensors used in prediction.
roi_size(roi_size)
input_dataset_key()	This is the input image key name used in dictionary based MONAI pre-transforms.
input_dataset_key(val)
pred_dataset_key()	This is the prediction key name used in dictionary based MONAI post-transforms.
pred_dataset_key(val)
overlap()	This is the overlap used during sliding window inference
overlap(val)
sw_batch_size()	The batch size to run window slices
sw_batch_size(val)
inferer()	The type of inferer to use
inferer(val)
compute(op_input, op_output, context)	Infers with the input image and save the predicted image to output
compute_impl(input_image, context)
pre_process(data)	Transforms input before being used for predicting on a model.
post_process(data)	Transforms the prediction results from the model(s).
predict(data)	Predicts results using the models(s) with input tensors.