Skip to content

Engine.md

Latest commit

 

History

History
106 lines (77 loc) · 3.49 KB

Engine.md

File metadata and controls

106 lines (77 loc) · 3.49 KB
FastnnUNet Logo

⚙️ FastnnUNet C++ Engine

The FastnnUNet C++ Engine is a high-performance implementation of the FastnnUNet architecture designed for production-level deployment. Built on CUDA operators and TensorRT, this engine provides ultra-fast inference for CT and MRI images, completing the entire inference pipeline in seconds.

✨ Key Features

  • 🚀 High Performance: Optimized C++ implementation with CUDA operators and TensorRT acceleration
  • 🏭 Production Ready: Engineered for stable clinical deployment with robust error handling
  • ⚡ Fast Inference: Complete inference pipeline for 3D medical images in seconds
  • 💾 Memory Efficient: Optimized memory management for handling large volumetric data
  • 🖥️ Platform Compatible: Designed to work on various NVIDIA GPU platforms
  • 🏥 Clinical Integration: Easy integration with clinical workflows and DICOM systems

🔧 Technical Specifications

  • 💻 Implementation: C++17 with CUDA 11.x/12.x support
  • 🚄 Acceleration: TensorRT optimization with FP16/INT8 quantization options
  • 📦 Dependencies: CUDA, cuDNN, TensorRT, OpenCV (minimal)
  • 🎮 GPU Support: NVIDIA GPUs with compute capability 6.0+
  • 📄 Input Formats: Supports NIfTI, DICOM, and raw data formats
  • 🔌 Integration: C++ API available

📊 Performance Metrics

Image Type Size Original nnUNet (PyTorch) FastnnUNet (PyTorch) FastnnUNet (C++/TensorRT)
Brain MRI 256×256×160 12-15s 4-8s 0.3-0.5s
Chest CT 512×512×400 30-40s 14-16s 0.1-0.5s
Abdominal CT 512×512×500+ 40-60s 15-18s 4.5-7.5s

*All metrics measured on NVIDIA RTX 5070Ti 4090 4080 3080 3060 3050 2080Ti GPUs

📋 Usage

💼 C++ API Example

#include "fast_nnunet_tools.h"
#include "fast_nnunet_load_image.h"
#include "fast_nnunet_engine.h"
#include "fast_nnunet_eva.h"

// initialization
const auto Engine = std::make_shared<FastnnUNet::Engine>();
Engine->set_config("configs/nnunet_bone_low_config.ini");
Engine->set_workspace("models/batch" ,false, true);

// load data
const std::string nii_file = "test_image/headneck.nii.gz";
const auto [image, inimg_raw, original_orientation] = Data::LoadData(nii_file);

// infer
const auto output_mask = Engine->infer(inimg_raw, image, true, false, true);

// save result
Tools::save_mask(output_mask, image, "output_mask.nii.gz");

🛠️ Building from Source

Requirements:

  • CUDA 11.x or newer
  • TensorRT 8.x or newer
  • CMake 3.18+
  • GCC 7+ or MSVC 2022+
# Clone repository
git clone https://github.com/username/FastnnUNet.git
cd FastnnUNet/engine

# Build
mkdir build && cd build
cmake ..
make -j8

# Install
make install

📤 Converting Models

The C++ engine uses optimized TensorRT models. Convert your ONNX models using:

# From ONNX
trtexec --onnx path/to/model.onnx --saveEngine engine.trt --fp16 --shapes=input:batch_size x 1 x D x H x W(enable batch infer)

📜 License

This component follows the same license as the main FastnnUNet project.

📝 Citation

If you use the FastnnUNet C++ Engine in your research, please cite:

Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021). 
nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. 
Nature methods, 18(2), 203-211.