Spatial adaptive dual-domain fusion MRI reconstruction algorithm based on lightweight dynamic convolution kernel
TL;DR
A spatially adaptive dual-domain fusion MRI reconstruction algorithm based on lightweight dynamic convolution kernels that effectively balances reconstruction accuracy and efficiency under high-magnification acceleration and holds potential for clinical translation.
Abstract
To address the imbalance in capturing global and local features during high-speed accelerated reconstruction in magnetic resonance imaging, we propose a spatially adaptive dual-domain fusion MRI reconstruction algorithm based on lightweight dynamic convolution kernels. A multi-scale feature dynamic initialization mechanism is designed to simultaneously characterize global structure and local details; a low-rank decomposition-driven lightweight dynamic convolution kernel is constructed to achieve position-by-position adaptive dual-domain fusion; an artifact-aware regulator is introduced to dynamically balance the proportions of direct reconstruction and residual learning to adapt to various under sampling modes; and an amplitude compensation branch is established to restore the tissue amplitude fidelity lost during normalization. Experiments show that the average reconstruction time per image ranges from 0.43 to 0.47 s under two acceleration scenarios, meeting real-time requirements; under Cartesian 8× down sampling, SSIM ≥ 0.91 and PSNR ≥ 30 dB; under Poisson 8× down sampling, SSIM ≥ 0.92 and PSNR ≥ 31 dB, with metrics converging smoothly as the sample size increases. This algorithm effectively balances reconstruction accuracy and efficiency under high-magnification acceleration and holds potential for clinical translation.