Transmit Radiofrequency Field (B1+) Map Prediction Using Machine Learning

Quantitative magnetic resonance imaging (qMRI) enables measurement of tissue parameters such as longitudinal T1 and transverse T2 relaxation times, which can reveal microstructural changes relevant to neurological disease. Accurate T1 and T2 mapping requires modelling of the signal and knowledge of the actual flip angle distribution, typically obtained from transmit radiofrequency field (B1⁺) mapping. However, B1⁺ acquisitions are often excluded from clinical and large-scale research protocols, limiting the reliability of downstream quantitative analyses. This thesis investigates the use of deep learning to predict B1⁺ maps in the brain from routinely acquired anatomical MR images. The Alberta 300 dataset was used including 267 healthy adult subjects (ages 19–90, 151 females) acquired at 3T at the Edmonton site. Available anatomical images included: volumetric T1-weighted magnetization-prepared rapid gradient echo (MPRAGE), and dual-echo proton density (PD) and T2-weighted turbo spin echo. In addition, a B1+ mapping sequence was included in each study, enabling a gold standard for model development. A 3D generative adversarial network (GAN) was trained to synthesize subject-specific B1⁺ distributions from the corresponding anatomical MR images (240 subjects for training, 27 for inference). Multiple input combinations (up to two channels) were tested to identify the best-performing configuration. Both whole-cohort (n = 267) and age-classified models (three groups of 89 subjects each) were evaluated using structural similarity (SSIM), mean absolute percentage difference (APD), and regional analyses across whole brain and subcortical regions. To further increase the number of inference subjects and assess model robustness, a four-fold randomized cross-validation was conducted, expanding the test set from 27 to 108 subjects. The GAN-predicted B1+ maps showed strong agreement with measured B1⁺ values. Among the input combinations, the single-channel T1-weighted input yielded the best whole-brain accuracy (APD = 3.17%, SSIM = 96.0%, averaged for all of 27 inference subjects in 3D space). After age separation, performance improved further (APD = 2.60%, SSIM = 97.0%, averaged for the same 27 inference subjects in 3D space across the three age groups). The four-fold randomized cross-validation confirmed stable performance (APD = 2.49%, SSIM = 96.2%, averaged for all of 108 inference subjects in 3D space across the three age groups). Regional analysis of B1⁺ maps across seven regions of interest showed errors typically below 3%, with the lowest error in gray matter (APD = 2.28%, averaged for all of 108 inference subjects in 3D space across the three age groups). When integrated into a T2 mapping pipeline that required dual echo PD and T2-weighted images and a B1+ map for accurate modelling, the predicted B1⁺ maps produced quantitative T2 values within 1.16% of those obtained using measured B1⁺ across the whole brain. Regional T2 errors were generally below 1%, with the lowest discrepancy in the putamen (APD = 0.28%, averaged for all of 108 inference subjects in 3D space across the three age groups). These findings demonstrate that accurate B1⁺ estimation can be achieved directly from standard MR contrasts, enabling retrospective correction of existing datasets and reducing reliance on direct transmit field mapping. This approach has the potential to make quantitative MRI more accessible in both research and clinical settings.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.