CUDA-Accelerated-Chest-X-Ray-Classification-with-ONNX!

About the Project

In this project, I built a high-performance pipeline for chest X-ray analysis. The goal was to speed up image preprocessing and model inference using GPU acceleration and ONNX, making it efficient enough to handle large datasets quickly.

Basically, I wanted to see how we can take medical images, prepare them efficiently, and run predictions fast — all while keeping things reproducible and clear.

---

What I Did

1. GPU-Powered Preprocessing

   • Converted chest X-ray images to grayscale using CUDA/Numba, running directly on the GPU.
   • Normalized pixel values so the images are ready for the model.
   • Saved these preprocessed images as .npy files for easy reuse.

2. ONNX Model Inference

   • Took the preprocessed images, replicated channels to match DenseNet-121’s expected input, and resized them to 224×224.
   • Ran predictions using ONNXRuntime, which is faster than standard PyTorch inference.
   • Recorded the predicted class, confidence, and inference time for each image.

3. Performance Comparison

   • Measured inference times for both PyTorch and ONNX.
   • Created bar charts to clearly show the speed improvement with ONNX.

---

Project Structure

project-root/
│
├─ dataset/
│   └─ chest_xray/      # Original X-ray images
│
├─ preprocessed/
│   └─ *.npy            # GPU-preprocessed images
│
├─ notebooks/
│   └─ preprocessing_and_onnx_inference.ipynb
│
├─ models/
│   └─ densenet121.onnx
│
└─ README.md

---

How to Use

1. Preprocessing Images

gray_img = preprocess_image_gpu(img)
np.save("preprocessed/NORMAL/IM-0001.npy", batch_img)

2. Running ONNX Inference

import onnxruntime as ort
ort_session = ort.InferenceSession("models/densenet121.onnx")
outputs = ort_session.run(None, {ort_session.get_inputs()[0].name: batch_resized})

3. Measuring Inference Time

start = time.time()
outputs = ort_session.run(None, {ort_session.get_inputs()[0].name: batch_resized})
end = time.time()
inference_time_ms = (end - start) * 1000

---

Tools and Libraries I Used

• NumPy – For array handling and batching.
• OpenCV – To resize and manipulate images.
• Numba – For CUDA-powered preprocessing.
• ONNXRuntime – For fast GPU inference.
• PyTorch – DenseNet-121 baseline comparison.
• Matplotlib & Pandas – For visualizations and tables.

---

Results

• Preprocessing on GPU was fast and efficient.
• ONNX inference ran significantly faster than PyTorch, helping speed up predictions.
• The pipeline is reproducible and works well with larger datasets.

---

Next Steps

• Implement TensorRT acceleration for even faster inference.
• Add batch processing to handle multiple images at once on HPC clusters.
• Explore fine-tuning DenseNet on pediatric X-ray datasets for better medical accuracy.

---