PyTorch Installation Guide for Medical Imaging Research

By Dominic Velasco, MD (@domvmd)

This guide helps nuclear medicine researchers set up PyTorch with GPU support for deep learning in medical imaging (e.g., PET/CT, SPECT reconstruction).

1. Why PyTorch for Medical Imaging?

Key Benefits:

GPU acceleration for large 3D medical datasets (e.g., 4D dynamic PET scans).

Flexible data pipelines (custom DICOM/NIfTI loaders).

Pretrained models for tasks like tumor segmentation (e.g., MONAI framework).

2. Prerequisites

Hardware: NVIDIA GPU (e.g., RTX 3070/3090, A100) with ≥12GB VRAM (for 3D volumes). Driver ≥535.86.10 (Download).

Software: Windows 10/11 or Linux (recommended for GPU support). Miniconda (Python 3.9).

3. Step-by-step Installation

Here’s a streamlined, medical-researcher-friendly integrating all requested components (Miniconda, environment setup, Jupyter integration, and PyTorch installation):

---

Step-by-Step Installation Guide

Target Audience: Nuclear Medicine Residents with Minimal Coding Experience

---

1. Install Miniconda

Purpose: Isolate dependencies and simplify package management.
Steps:

1. Download Miniconda for Windows (or Linux).
2. Run the installer. At the "Advanced Options" screen:
   • Check ✅ "Add Miniconda3 to my PATH environment variable" (avoids manual setup).
   • Check ✅ "Register Miniconda3 as my default Python".

---

2. Create a Conda Environment

Purpose: Avoid conflicts with existing Python projects.

# Open Anaconda Prompt (Windows) or Terminal (Linux)
conda create -n medical_pytorch python=3.9
conda activate medical_pytorch

---

3. Install Jupyter Notebook & Kernel

Purpose: Run code interactively for medical imaging experiments.

# Install Jupyter Notebook and ipykernel inside the environment
conda install jupyter ipykernel

# Link your environment to Jupyter
python -m ipykernel install --user --name=medical_pytorch --display-name="PyTorch (Medical)"

---

4. Install PyTorch with CUDA 12.4

Purpose: Leverage GPU acceleration for large medical datasets (e.g., 3D PET/CT scans).

conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia

---

5. Verify Installation in Jupyter

1. Launch Jupyter:

   jupyter notebook

2. Create a new notebook and select the PyTorch (Medical) kernel.
3. Run this code:

   import torch
   
   # GPU check (critical for medical imaging workflows)
   print(f"GPU Available: {torch.cuda.is_available()}")  # Must return "True"
   print(f"GPU Memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB")
   
   # Test tensor operations on GPU
   pet_scan_simulated = torch.randn(128, 128, 128).cuda()  # Simulate 3D PET scan
   print(f"Tensor on {pet_scan_simulated.device}")

Expected Output:

GPU Available: True  
GPU Memory: 8.0 GB  # RTX 3070 has 8GB VRAM  
Tensor on cuda:0

---

Troubleshooting for Medical Researchers

• Jupyter Kernel Not Found:
  • Ensure you ran python -m ipykernel install after activating medical_pytorch.
• CUDA Not Detected:
  • Update NVIDIA drivers here.
  • Reinstall PyTorch:

    conda install pytorch pytorch-cuda=12.4 -c pytorch -c nvidia

---

Next Steps for Medical Imaging

Include in your repo’s examples/ folder:

• Loading DICOM/NIfTI files with SimpleITK.
• Training a tumor segmentation model with MONAI.

---

Sample GitHub README Snippet

## Quick Start  
1. Install Miniconda.  
2. Copy-paste these commands into Anaconda Prompt:  
   ```bash
   conda create -n medical_pytorch python=3.9
   conda activate medical_pytorch
   conda install jupyter ipykernel
   python -m ipykernel install --user --name=medical_pytorch --display-name="PyTorch (Medical)"
   conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia
   jupyter notebook

3. Verify GPU support using the provided test code.

This structure ensures **clarity** and **reproducibility** for clinicians new to deep learning. Let me know if you’d like a full repo template!