SelExNet: A Self-Supervised Physics-Informed Framework for Multi-Channel Joint RF and Gradient Waveform Optimization in 2D Spatially Selective Excitation

Xiao, Yuliang; Rock, Jason; Wu, Zhe; Near, Jamie; Chiew, Mark; Graham, Simon J.

SelExNet A Self-Supervised Physics-Informed Framework for Multi-Channel Joint RF and Gradient Waveform Optimization in 2D Spatially Selective Excitation

Yuliang Xiao^1,2,#, Jason Rock^1,2, Zhe Wu³, Jamie Near^1,2, Mark Chiew^1,2, Simon J. Graham^1,2

¹Sunnybrook Research Institute ²University of Toronto ³Siemens Healthcare Limited

Magnetic Resonance in Medicine 2026

^#Indicates Corresponding Author

Abstract Code

With SelExNet, a tailored pTx RF pulse and spiral gradient waveform are generated from input B0/B1+ maps to produce high-fidelity spatially selective excitation that closely matches the target pattern under patient-specific field conditions and achieve reduced field-of-view imaging at 7T. — With SelExNet, a tailored pTx RF pulse and spiral gradient waveform are generated from input B₀ and B₁⁺ maps to produce high-fidelity spatially selective excitation that closely matches the target pattern under patient-specific field conditions and achieve reduced field-of-view (rFOV) imaging at 7T.

Abstract

SelExNet is a self-supervised framework for 2D spatially selective excitation that jointly optimizes radiofrequency (RF) pulses and gradient waveforms, with extension to multi-channel transmission MRI. It couples neural RF and gradient generators with a differentiable Bloch simulator, enabling pulse optimization directly from desired excitation outcomes without requiring pre-designed target pulses.

The framework designs RF pulses and parameterized variable-density spiral gradient waveforms for both single- (sTx) and multi-channel (pTx) transmission, and supports patient-specific adaptation using measured, previously unseen B₀ and B₁⁺ maps. Joint RF-Gradient optimization improves excitation fidelity over RF-only optimization. In phantom experiments, patient-specific fine-tuning restores target geometry and uniformity under field inhomogeneity. In-vivo studies demonstrate anatomically precise excitation, sharper target boundaries, and reduced off-target signal.

Network Overview

SelExNet uses a Feature Compression and Selection Module to model plausible B₀/B₁⁺ field configurations with PCA and Gaussian mixture modeling. The RF module produces complex-valued RF waveforms for each transmit channel, while the gradient module predicts a small set of parameters for a variable-density spiral trajectory, including the number of spiral turns, radial and angular rates, and maximum excitation k-space extent.

Generated RF and gradient waveforms are passed through a differentiable Bloch simulator. The loss combines profile fidelity terms, RF amplitude and smoothness regularization, gradient amplitude and slew-rate constraints, and weight decay. Minute-scale patient-specific transfer learning fine-tunes the pretrained model using measured field maps, providing fast adaptation to individual field inhomogeneity conditions.

SelExNet framework with feature compression, neural RF and gradient generators, and differentiable Bloch simulation — SelExNet: a self-supervised physics-informed framework for joint optimization of multi-channel RF and gradient pulses, consists of the: (a) Feature Compression and Selection Module; and the (b) Multi-Channel Joint RF-Gradient Optimization Network for selective excitation.

Experimental Results

Single-transmit phantom validation across AI, square, and triangle targets

Representative phantom results using SelExNet in the sTx configuration (N_c = 1) under ideal conditions without pronounced B₀/B₁⁺ inhomogeneities. Representative excitation patterns for letter, square, and triangular targets demonstrate accurate geometry reproduction with minimal background excitation.

RF-only optimization compared with joint RF-Gradient optimization

Qualitative comparison of RF-only optimization and joint RF-Gradient optimization for the "AI" target with SelExNet under sTx conditions and ideal fields for MRI experiments in a uniform phantom.

Patient-specific transfer learning under field inhomogeneity

Patient-specific transfer learning restores UT geometry and uniformity under previously unseen B₀/B₁⁺ inhomogeneity.

In-vivo single-transmit brain excitation results

In-vivo sTx brain experiments show close agreement between simulated and measured "UT" excitation under patient-specific field maps.

In-vivo pTx reduced field-of-view MRI with SelExNet-designed excitation

At 7T, SelExNet-designed pTx excitation confines signal to the prescribed rFOV while preserving anatomical detail.

Quantitative Evaluation

Method	SSIM ↑	PSNR (dB) ↑	NRMSE ↓	BSR ↓
Single-channel Transmission
GPS-RF	0.681	24.9	0.065	0.045
SelExNet-RF-Only	0.937	26.2	0.053	0.016
SelExNet-Joint	0.956	27.6	0.042	0.012
Multi-channel Transmission
Vendor-provided demo pTx	0.613	18.6	0.182	0.157
SelExNet-Joint	0.964	28.5	0.037	0.007

BSR is the background-to-signal ratio. Lower BSR indicates stronger suppression of excitation outside the target ROI.

Accepted Abstract

BibTeX

@article{xiao2026selexnet,
  title={SelExNet: A Self-Supervised Physics-Informed Framework for Multi-Channel Joint RF and Gradient Waveform Optimization in 2D Spatially Selective Excitation},
  author={Xiao, Yuliang and Rock, Jason and Wu, Zhe and Near, Jamie and Chiew, Mark and Graham, Simon J.},
  journal={Magnetic Resonance in Medicine},
  year={2026}
}

More Works

Self-Supervised Cyclic Diffeomorphic Mapping

AutoSeg4SinonasalCT

AutoSeg4ETICA