Title: CoilDrop-MRI: A Self-Supervised, Physics-Guided Approach to MRI Reconstruction via Coil Dropout

Abstract:

Self-supervised deep learning techniques have emerged as a promising solution for accelerating magnetic resonance imaging (MRI) reconstruction, delivering high-fidelity images without the need for fully sampled data during the training phase. Conventional strategies in this domain typically split acquired data into separate input and target subsets to train reconstruction networks; however, these methods generally restrict this partitioning to the spatial frequency (k-space) domain, ignoring the potential of the coil dimension. To address this limitation and fully leverage signal correlations across receiver coils, we introduce CoilDrop-MRI. This novel approach implements coil-wise dropout, utilizing the masked coil data as training targets within a self-supervised learning framework. We demonstrate the integration of CoilDrop-MRI into unrolled architectures, applying it to both image-domain (SENSE) and k-space (SPIRiT) formulations. Furthermore, we validate the method’s adaptability by extending its application to multi-shot, phase-corrected diffusion MRI (dMRI) reconstruction.

Extensive testing across multi-site, multi-field-strength (0.3T, 0.55T, and 3T), and multi-modality datasets—including T1-weighted, T2-weighted, T2-FLAIR, and dMRI scans—confirms that CoilDrop-MRI consistently surpasses current state-of-the-art self-supervised techniques. It achieves image quality on par with supervised methods, all while eliminating the requirement for fully sampled reference data for training. Additionally, CoilDrop-MRI demonstrates superior data efficiency and robust generalization across diverse imaging scenarios, positioning it as a highly practical and versatile framework for self-supervised parallel MRI reconstruction.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC