NKI-AI/direct

What it solves

DIRECT is a PyTorch toolkit designed to accelerate MRI reconstruction. It addresses the challenge of turning undersampled multi-coil k-space data (raw MRI data) into high-quality images, providing a standardized way to handle sampling, reconstruction, and validation.

How it works

The toolkit provides an end-to-end pipeline that manages the entire process from raw data to final image. It includes:

• Data Handling: Support for various MRI volumes (static, dynamic, multislice) and a library of sampling masks (Cartesian, radial, spiral, etc.).
• Model Library: A wide array of reconstruction architectures, including vSHARP, RecurrentVarNet, RIM, and transformer-based reconstructors like ViT and UFormer.
• Training Options: Support for fully supervised, self-supervised (SSDU), and joint supervised/self-supervised (JSSL) learning paradigms.
• Optimization: Tools for distributed multi-GPU training, mixed precision, and conditional pipelines that can adjust networks based on acceleration factors.

Who it’s for

It is primarily for researchers and developers working on medical imaging, specifically those focused on deep learning-based MRI reconstruction and acceleration.

Highlights

• Challenge-Winning Models: Includes implementations of models like vSHARP and RecurrentVarNet that have won major MRI reconstruction competitions.
• Comprehensive Sampling: Offers a diverse set of k-space masks and a learned Adaptive Dynamic Sampler (ADS).
• Extensive Metrics: Built-in support for MRI-specific validation metrics such as SSIM, pSNR, and NMSE.
• Model Zoo: Provides access to pretrained models via Hugging Face for quick deployment and reproduction.