DM-Net: a physics-model-independent direct mapping approach for calibration-free multi-coil MRI

Deep learning-based multi-coil magnetic resonance image reconstruction has been actively investigated and increasingly applied in clinical settings. However, most models are physics-model-based approaches, which either require precalculation of coil sensitivity or fail to achieve optimal performance without taking coil sensitivity into consideration. Inspired by ESPIRiT, we propose a physics-model-independent direct mapping approach, namely DM-Net, which aims for optimal reconstruction without explicitly using coil sensitivity. A densely connected convolutional network is employed, where coil sensitivity and channel correlation are intrinsically exploited. For comparison, other reconstruction models that mimic SENSE and GRAPPA are implemented. DMwS-Net is a direct mapping approach that incorporates precalculated coil sensitivity maps as additional input; and CCR-Nets provide coil-by-coil reconstruction, where individual coil images are jointly predicted and combined with or without coil sensitivity. The proposed models are trained and tested on 5440 images from 17 subjects acquired with 3DFT (Fourier transform). DM-Net achieves superior performance, suggesting the feasibility of excluding precalculated coil sensitivity from model input. Moreover, DM-Net can be applied on k-space data without a fully sampled calibration region, which may improve image quality by sampling high frequency regions more densely. This work expands the scope of DL-based image reconstruction and provides evidence that a direct mapping framework without using coil sensitivity may lead to reliable, simplified and faster reconstruction.