Higher Fields, Deeper Insights: DL-based $B₁⁺$ prediction at 10.5T

Motivation: Deep learning (DL)-based channel-wise $B₁⁺$-mapping at 10.5T can substantially reduce necessary $$B₁⁺$ calibration times, but the network's inner workings remain unclear. Goal(s): Analyzing the impact of individual 80Rx head-coil elements to predict the 16Tx elements, using interpretability methods for DL-based $B₁⁺$-mapping and gaining insights to the network's decision-making. Approach: Localizers and $$B₁⁺$-maps collected at 10.5T using a 16Tx/80Rx head-coil were supplied to a DL network to rapidly predict $$B₁⁺$-maps while evaluating its reliance on specific Rx-channels. Results: Reducing Rx-channels from 80 to as few as 4 improves accuracy, suggesting redundancy; feature permutation maps further support redundancy in Rx-channels for DL-based $B₁⁺$-mapping. Impact: The study suggests that training a neural network to predict $$B₁⁺$-maps for a 16Tx/80Rx head coil at 10.5T might not require all coil elements, highlighting methods to identify redundant elements to optimize training speed and specific applications.