3333

View Abstract / View Presentation

On the predictability of B0 maps at ultra-high fields using deep learning

Vidya Prasad¹, Sharun S Thazhackal¹, Ashvin Srinivasan¹, Suja Saraswathy¹, Jaladhar Neelavalli², and Umesh Rudrapatna²
¹Philips Research, Philips, Bangalore, India, ²Philips Healthcare, Philips, Bangalore, India

Deep learning can predict complete B₀ maps and spherical harmonics for shimming with the input being deformation fields obtained from structural image co-registration to a template. Our results indicate that the predicted UHF B₀ brain maps work as effectively as acquired B₀ maps for shimming.

Figure 1. Learning Pipeline Overview : CT images were used to generate synthetic B₀ maps via analytical computations1 (Y_{ground_truth}). Each CT image was aligned to a template to generate a 3D composite deformation field in x,y,z directions (X). Two experiments were then conducted to predict (Y_predicted) the 1) n^th (4^th) order orthogonal spherical harmonics using a modified 3D EfficientNet-B0¹¹ and 2) whole B₀ maps using a residual variant of a 3D UNet¹³.

Figure 5. SD of residuals after shimming from image-to-coefficient model @ 7T. Ground truth and predicted B₀ maps were reconstructed using upto the 1^st, 2^nd, 3^rd and 4^th order SH coefficients (excluding 0^th term) respectively. The residuals of the reconstructed maps were computer per order. Reported is the average standard deviation (y-axis) per order (x-axis) in Hz across test set subjects of the predicted and GT residual maps. Metrics were computed within the brain region only.