Blood Cell Detection Model

Ready to deploy blood cell detection model for NVIDIA Jetson platform on DeepStream SDK. Trained with using TAO Toolkit based on DetectNet V2 pre-trained model. The model can detect RBCs and WBCs from peripheral blood smear.

Prerequisities

• NVIDIA Jetson Developer Kits or Production Module Systems
• CSI Camera, USB Camera or h264 Encoded Video
• Ubuntu 18.04 L4T
• NVIDIA JetPack SDK 4.6.1
• NVIDIA DeepStream SDK 6.0.1

Dataset

I used blood cell detection dataset which I collected and annotated when I was in medical school. Visit the dataset repository for further information.

Example Usage

1. Be sure that you properly installed JetPack SDK 4.6.1.
2. Install NVIDIA DeepStream SDK 6.0.1. You can follow the official documentation.
3. Clone DeepStream TAO Apps repository. Notice that we cloned the branch for DeepStream 6.0.1 version. an

git clone -b release/tao3.0_ds6.0.1 https://github.com/NVIDIA-AI-IOT/deepstream_tao_apps/

4. Open deepstream_tao_apps directory, set CUDA version variable and compile.

cd deepstream_tao_apps
export CUDA_VER=10.2
make

5. Clone blood-cell-detection-model repository.

git clone https://github.com/draaslan/blood-cell-detection-model

6. Run ds-tao-detection app with config and test video. This process can be take a while.

./apps/tao_detection/ds-tao-detection -d \
    -c blood-cell-detection-model/inference_config.txt \
    -i blood-cell-detection-model/test.h264 

7. Model output will be saved as h624 format and inference result will be open on a window.

Future Work

• Add video and image demonstration.
• Train model using more data with augmentation.
• Improve inference config.
• Containerize full application.

Licence

See LICENSE for details.