CoilDrop-MRI: Self-supervised physics-guided MRI reconstruction with coil dropout

Researchers introduce CoilDrop-MRI, a self-supervised deep learning method that improves accelerated MRI reconstruction by strategically dropping data across receiver coils rather than only in k-space. Validated across multiple hospital sites and field strengths, the approach matches supervised methods' quality without requiring fully sampled training data, offering practical efficiency gains for medical imaging.

CoilDrop-MRI represents a meaningful advancement in medical imaging reconstruction by addressing a technical limitation in existing self-supervised learning approaches. Traditional methods partition acquired MRI data exclusively within k-space (spatial frequency domain), but this innovation applies coil-wise dropout—dropping entire receiver coil signals—to create input-target training pairs. This exploits signal correlation patterns that previous methods overlooked, enabling the neural network to learn more robust reconstruction strategies.

The technique emerges from the broader push to reduce MRI scan times and training data requirements in clinical settings. Accelerated MRI has long required either fully sampled reference data for training or compromised image quality. Self-supervised learning bridges this gap by learning from undersampled data itself, reducing scanning burden on patients. CoilDrop-MRI improves upon this foundation through physics-informed design integrated into unrolled network architectures that mirror actual MRI physics.

For healthcare institutions and imaging device manufacturers, this development carries substantial practical value. The method demonstrates consistent performance across 0.3T to 3T field strengths and multiple imaging modalities (T1, T2, T2-FLAIR, diffusion MRI), indicating strong real-world applicability. By achieving supervised-quality results without fully sampled training data, hospitals can implement faster MRI protocols while maintaining diagnostic accuracy, directly reducing scan times and operational costs. The strong generalization across imaging conditions suggests the approach will transfer well to new clinical environments.

Future developments likely include broader adoption across MRI vendors' reconstruction pipelines and extension to even more specialized sequences. The demonstrated data efficiency indicates potential applications in emerging imaging modalities where reference data is particularly scarce.

• →CoilDrop-MRI applies coil-dimension dropout rather than only k-space partitioning, enabling more efficient self-supervised MRI reconstruction training.
• →Method achieves supervised-level image quality without requiring fully sampled reference data, reducing data requirements for clinical MRI applications.
• →Validation across multiple sites, field strengths (0.3T-3T), and imaging modalities demonstrates strong generalization and practical clinical utility.
• →Physics-guided unrolled architectures in both image and k-space domains provide theoretical grounding and architectural flexibility.
• →Strong data efficiency and robust cross-condition generalization position this as a production-ready framework for hospital imaging systems.