Endoscopy Out of Body Detection #

Authors: Holoscan Team (NVIDIA)
Supported platforms: x86_64, aarch64
Language: C++
Last modified: May 13, 2025
Latest version: 1.0
Minimum Holoscan SDK version: 1.0.3
Tested Holoscan SDK versions: 2.9.0
Contribution metric: Level 1 - Highly Reliable

Endoscopy Out of Body Detection Workflow

Overview#

This application performs real-time detection of whether an endoscope is inside or outside the body during endoscopic procedures. For each input frame, the application:

Classifies the frame as either "in-body" or "out-of-body"
Provides a confidence score for the classification
Outputs either to the console or to a CSV file (when analytics is enabled)

Note: This application does not include visualization components.

Quick Start#

Run the following command to build and launch the application on a supported Holoscan platform:

./dev_container build_and_run endoscopy_out_of_body_detection --language cpp

Prerequisites#

NVIDIA Holoscan SDK (version 0.5 or higher)
A supported Holoscan platform or workstation with a CUDA-capable NVIDIA GPU
CMake build system
FFmpeg (for data conversion)

Data Requirements#

Model and Sample Data#

The endoscopy detection model and sample datasets are available on NGC. The package includes:

Pre-trained ONNX model for out-of-body detection: out_of_body_detection.onnx
Sample endoscopy video clips (MP4 format): sample_clip_out_of_body_detection.mp4

Data Preparation (optional)#

The application requires the input videos to be converted to GXF tensor format. This conversion happens automatically during building, but manual conversion can be done following these steps:

Download and extract the data:
```
unzip [NGC_DOWNLOAD].zip -d <data_dir>
```

Convert the video to GXF tensor format using the provided script:

ffmpeg -i <INPUT_VIDEO_FILE> -fs 900M -pix_fmt rgb24 -f rawvideo pipe:1 | \
python convert_video_to_gxf_entities.py --width 256 --height 256 --channels 3 --framerate 30

Note: The conversion script (convert_video_to_gxf_entities.py) is available in the Holoscan SDK repository.

Organize the data directory as follows:

data/
└── endoscopy_out_of_body_detection/
  ├── LICENSE.md
  ├── out_of_body_detection.onnx
  ├── sample_clip_out_of_body_detection.gxf_entities
  ├── sample_clip_out_of_body_detection.gxf_index
  └── sample_clip_out_of_body_detection.mp4

Configuration#

The application uses endoscopy_out_of_body_detection.yaml for configuration. Key settings include:

Input video parameters in the replayer section
Model parameters in the inference section
Analytics settings for data export

Building#

./dev_container build
./dev_container launch
./run build endoscopy_out_of_body_detection cpp
./run launch endoscopy_out_of_body_detection cpp

For more information, see the Holohub README.md.

Running the Application#

Basic Usage#

From your build directory:

applications/endoscopy_out_of_body_detection/endoscopy_out_of_body_detection \
  --config endoscopy_out_of_body_detection.yaml \
  --data ../data/endoscopy_out_of_body_detection

Analytics Mode#

To enable analytics and export results to CSV:

Set enable_analytics: true in the configuration file:

# endoscopy_out_of_body_detection.yaml
enable_analytics: true

Configure analytics output (optional):

# Set output directory (default: current directory)
export HOLOSCAN_ANALYTICS_DATA_DIRECTORY="/path/to/output"

# Set output filename (default: data.csv)
export HOLOSCAN_ANALYTICS_DATA_FILE_NAME="results.csv"

The application will create:

A directory named after the application
Subdirectories with timestamps for each run
CSV files containing frame-by-frame classification results

Output Format#

Console Mode: Displays "Likely in-body" or "Likely out-of-body" along with confidence scores for each frame.
Analytics Mode: Outputs a CSV file with frame-by-frame classification results in the following format:

In-body,Out-of-body,Confidence Score
1,0,0.972432
1,0,0.902066

Endoscopy Out of Body Detection#