New framework enables scalable training for advanced MRI reconstruction networks

Researchers have developed a novel memory-efficient training framework for unrolled variational networks used in accelerated MRI reconstruction. This new method allows for training with significantly longer and wider cascades of networks, overcoming the memory limitations of standard end-to-end training on GPUs. The framework employs an activation-light backward pass and a shared time-conditioned network to reduce memory usage and parameter count, respectively, while maintaining reconstruction quality. Tested on the fastMRI dataset for knee and brain scans, the approach consistently outperforms other local training methods and matches or surpasses the performance of E2E-VarNet with fewer parameters and lower peak memory usage. AI

IMPACT This research could enable more complex and accurate AI models for medical imaging, potentially leading to faster and more precise diagnoses.

RANK_REASON The cluster contains a research paper detailing a new technical framework for a specific scientific application (MRI reconstruction). [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• E2E-VarNet
• Emiljo Mëhillaj
• Fastmri
• graphics processing unit
• magnetic resonance imaging
• Unrolled variational networks

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →