Accelerated and quantitative three-dimensional molecular MRI using a generative adversarial network

Abstract

Purpose: To substantially shorten the acquisition time required for quantitative three-dimensional (3D) chemical exchange saturation transfer (CEST) and semisolid magnetization transfer (MT) imaging and allow for rapid chemical exchange parameter map reconstruction.

Methods: Three-dimensional CEST and MT magnetic resonance fingerprinting (MRF) datasets of L-arginine phantoms, whole-brains, and calf muscles from healthy volunteers, cancer patients, and cardiac patients were acquired using 3T clinical scanners at three different sites, using three different scanner models and coils. A saturation transfer-oriented generative adversarial network (GAN-ST) supervised framework was then designed and trained to learn the mapping from a reduced input data space to the quantitative exchange parameter space, while preserving perceptual and quantitative content.

Results: The GAN-ST 3D acquisition time was 42-52 s, 70% shorter than CEST-MRF. The quantitative reconstruction of the entire brain took 0.8 s. An excellent agreement was observed between the ground truth and GAN-based L-arginine concentration and pH values (Pearson's r > 0.95, ICC > 0.88, NRMSE < 3%). GAN-ST images from a brain-tumor subject yielded a semi-solid volume fraction and exchange rate NRMSE of $3.8\pm 1.3\%$ and $4.6\pm 1.3\%$ , respectively, and SSIM of $96.3\pm 1.6\%$ and $95.0\pm 2.4\%$ , respectively. The mapping of the calf-muscle exchange parameters in a cardiac patient, yielded NRMSE < 7% and SSIM > 94% for the semi-solid exchange parameters. In regions with large susceptibility artifacts, GAN-ST has demonstrated improved performance and reduced noise compared to MRF.

Conclusion: GAN-ST can substantially reduce the acquisition time for quantitative semi-solid MT/CEST mapping, while retaining performance even when facing pathologies and scanner models that were not available during training.

Keywords: chemical exchange saturation transfer; generative adversarial network; magnetic resonance fingerprinting; magnetization transfer; pH; quantitative imaging.

FIGURE 1

Generative adversarial network (GAN)-saturation transfer (ST) architecture. (A) A conditional GAN framework, receives N raw, molecular information encoding semi-solid magnetization transfer/chemical exchange saturation transfer images, and is trained to simultaneously output the quantitative proton volume fraction and the exchange rate maps. (B) A fully connected neural network, receiving the full-length raw magnetic resonance fingerprinting image series (M > N) pixelwise, as well as T₁, T₂, and B₀ maps, and yielding the reference proton volume fraction and exchange rate maps. The output of this network was used for training GAN-ST.

FIGURE 2

Generative adversarial network (GAN)-saturation transfer (ST) in vitro image results. (A,B) L-arginine concentration (A) and exchange rate (B) maps from GAN-ST-based reconstruction, obtained with N = 9. Vials are numbered 1–6. (C,D) Full-length chemical exchange saturation transfer-magnetic resonance fingerprinting-based L-arginine concentration (C) and exchange rate (D) maps, obtained with M = 30. (E,F) GAN-ST-based (N = 9) concentration (E) and pH (F) maps. (G,H) Concentration (G) and pH (H) maps obtained using gold-standard non-MRI measures.

FIGURE 3

Statistical analysis and quantitative assessment of Generative adversarial network (GAN)-saturation transfer (ST) performance in vitro. (A,B) Correlation between GAN-ST based and chemical exchange saturation transfer-magnetic resonance fingerprinting-based concentration (A) and exchange rate (B) maps across the entire three-dimensional volume of an L-arginine phantom. (E,F) Box plots showing the distribution of per-vial GAN-ST-based L-arg concentration (E) and pH (F) maps with measured values indicated. Vial numbers are based on Figure 2A. (C,D,G,H) Structural similarity index and normalized root mean squared error for concentration/exchange rate (C,D) and concentration/pH (G,H) maps.

FIGURE 4

Quantitative semisolid magnetization transfer (MT) parameter maps from a healthy volunteer, scanned at a site and scanner model that were not used during training. (A–D) Generative adversarial network (GAN)-saturation transfer (ST)-based semi-solid MT proton volume fraction maps, obtained with N = 9. (E-H) chemical exchange saturation transfer (CEST)-magnetic resonance fingerprinting (MRF)-based semisolid MT proton volume fraction maps, obtained with M = 30. (I-L) GAN-ST-based semi-solid MT proton exchange rate maps, obtained with N = 9. (M-P) CEST-MRF-based semi-solid MT proton exchange rate maps, obtained with M = 30. The red arrows indicate regions with susceptibility artifacts.

FIGURE 5

Quantitative semisolid magnetization transfer (MT) parameter maps from a glioblastoma patient. (A–D) Generative adversarial network (GAN)-saturation transfer (ST)-based semi-solid MT proton volume fraction maps, obtained with N = 9. (E–H) chemical exchange saturation transfer (CEST)-magnetic resonance fingerprinting (MRF)-based semisolid MT proton volume fraction maps, obtained with M = 30. (I-L) GAN-ST-based semi-solid MT proton exchange rate maps, obtained with N = 9. (M-P) CEST-MRF-based semi-solid MT proton exchange rate maps, obtained with M = 30. The red arrows indicate regions with susceptibility artifacts.

FIGURE 6

Statistical analysis and quantitative assessment of the generative adversarial network (GAN)-saturation transfer (ST) performance in the in vivo brains of a tumor patient (A–D) and a healthy volunteer (E–H). (A,B) Correlation between all GAN-ST-based proton semi-solid magnetization transfer (MT) proton volume fractions (A) and exchange rates (B) for the entire brain in the WM/GM, and the corresponding pixel values obtained using chemical exchange saturation transfer (CEST)-magnetic resonance fingerprinting (MRF). Notably, the GAN-based f_ss values in the WM are in better agreement with MRF refernce than the GM (Pearson’s r = 90 compared to 0.74, respectively, p < 0.001), due to the myelin-rich content of the WM. (E,F) A similar analysis for the healthy human volunteer scanned at a site and scanner that were not available during training. (C,D,G,H) Structural similarity index metric and normalized root mean squared error for the tumor patient (C,D) and healthy volunteer (G,H).

FIGURE 7

Quantitative semi-solid magnetization transfer (MT) parameter maps from the calf muscle of a cardiac patient. (A–D) Generative adversarial network (GAN)-saturation transfer (ST)-based semi-solid MT proton volume fraction maps, obtained with N = 9. (E–H) chemical exchange saturation transfer (CEST)-magnetic resonance fingerprinting (MRF)-based semisolid MT proton volume fraction maps, obtained with M = 30. (I–L) GAN-ST-based semi-solid MT proton exchange rate maps, obtained with N = 9. (M–P) CEST-MRF-based semi-solid MT proton exchange rate maps, obtained with M = 30.

FIGURE 8

Statistical analysis and quantitative assessment of the generative adversarial network (GAN)-saturation transfer (ST) performance in the calf-muscle of a cardiac patient. (A) Structural similarity index metric. (B) Normalized root mean squared error