核心概念
Introducing an end-to-end framework for adaptive dynamic subsampling and reconstruction in cardiac MRI, showcasing superior performance at high accelerations.
要約
The study presents a novel approach integrating a DL-based adaptive sampler with a state-of-the-art reconstruction network for dynamic MRI. By leveraging temporal correlations, the method outperforms traditional and DL-optimized techniques. The experiments were conducted on a multi-coil cardiac MRI dataset, demonstrating significant improvements in reconstruction quality, especially at high acceleration factors. The proposed method shows adaptiveness in both phase-specific and unified scenarios, highlighting its potential for real-time dynamic subsampled acquisitions.
統計
Acceleration factor R = 4: SSIM results - Gauss: 0.9930; Rand: 0.9921; Fixed Equi: 0.9933; Opt: 0.9936; Adapt: 0.9936; Ad-I: 0.9934; Ad-I2: 0.9936; Ad-NU: 0.9938; Ad-k: 0.9934; Learned Ad-kI2: 0.9936
Acceleration factor R = 6 (Phase-specific SSIM results): Opt - 0.9963, Adapt - 0.9965
Unified SSIM results at R = 8 (1D): Fixed Gauss - 0.9851, Rand - 0.9798, Fixed Equi - 0.9882, Opt - 0.9880, Adapt - 0.9877, Ad-I - 0.9880, Learned Ad-I2 - 0.9888
Unified SSIM results at R = 8 (2D): Fix Gauss - 0.9954, Opt - 0.9958, Learn Adapt -
Table S1 provides additional pSNR & NMSE results for phase-specific sampling methods at different acceleration factors.
Table S2 showcases pSNR & NMSE results for phase-specific sampling methods in the context of unified schemes.
Table S3 displays pSNR & NMSE outcomes for unified sampling strategies with varying acceleration factors.
Table S4 presents pSNR & NMSE metrics for unified sampling methods considering different acceleration rates.
引用
"The proposed method demonstrates adaptiveness in both phase-specific and unified scenarios."
"Our approach notably outperforms traditional and DL-optimized techniques."