Deep-ER: Deep Learning ECCENTRIC Reconstruction for fast high-resolution neurometabolic imaging

Introduction: Altered neurometabolism is an important pathological mechanism in many neurological diseases and brain cancer, which can be mapped non-invasively by Magnetic Resonance Spectroscopic Imaging (MRSI). Advanced MRSI using non-cartesian compressed-sense acquisition enables fast high-resolution metabolic imaging but has lengthy reconstruction times that limits throughput and needs expert user interaction. Here, we present a robust and efficient Deep Learning reconstruction to obtain high-quality metabolic maps. Methods: Fast high-resolution whole-brain metabolic imaging was performed at 3.4 mm$^3$ isotropic resolution with acquisition times between 4:11-9:21 min:s using ECCENTRIC pulse sequence on a 7T MRI scanner. Data were acquired in a high-resolution phantom and 27 human participants, including 22 healthy volunteers and 5 glioma patients. A deep neural network using recurring interlaced convolutional layers with joint dual-space feature representation was developed for deep learning ECCENTRIC reconstruction (Deep-ER). 21 subjects were used for training and 6 subjects for testing. Deep-ER performance was compared to conventional iterative Total Generalized Variation reconstruction using image and spectral quality metrics. Results: Deep-ER demonstrated 600-fold faster reconstruction than conventional methods, providing improved spatial-spectral quality and metabolite quantification with 12%-45% (P<0.05) higher signal-to-noise and 8%-50% (P<0.05) smaller Cramer-Rao lower bounds. Metabolic images clearly visualize glioma tumor heterogeneity and boundary. Conclusion: Deep-ER provides efficient and robust reconstruction for sparse-sampled MRSI. The accelerated acquisition-reconstruction MRSI is compatible with high-throughput imaging workflow. It is expected that such improved performance will facilitate basic and clinical MRSI applications.

💡 Research Summary

The paper introduces Deep‑ER, a deep‑learning reconstruction framework tailored for non‑Cartesian ECCENTRIC‑encoded whole‑brain 1H‑MRSI at 7 T. ECCENTRIC sampling uses randomized circular k‑space trajectories that simultaneously accelerate two spatial dimensions, achieving acquisition times of 4–9 minutes for 3.4 mm isotropic voxels. Traditional reconstruction with Total Generalized Variation (TGV) is computationally intensive, often requiring hours, which hampers clinical translation. To overcome this, the authors design a network based on the Interlacer architecture, extending it to three dimensions and integrating a non‑Cartesian gridding layer (iNUFT followed by FFT). The network processes each spectral time point independently, reducing memory demands from ~14 GB to ~31 MB per time point, enabling training on standard GPUs.

Key technical innovations include: (1) dual‑domain feature extraction via recurrent Interlacer layers that alternately apply 3‑D convolutions in image space and k‑space, with learnable mixing coefficients (α_i, β_i) to fuse information after appropriate Fourier transforms; (2) handling complex‑valued data by separating real and imaginary channels; (3) leveraging unsuppressed water MRSI as a high‑SNR training target, acquired with a shortened repetition time (TR = 100 ms) to provide fully sampled ground truth; (4) end‑to‑end optimization of the network within a full MRSI pipeline that also includes coil combination (ESPIRiT), B0 correction, water/lipid removal, low‑rank decomposition, and spectral fitting. The loss function combines mean‑squared error and structural similarity (SSIM) to enforce both pixel‑wise fidelity and perceptual quality.

Training employed data from 21 subjects (healthy volunteers) and testing on six subjects (including five glioma patients). Compared to the conventional TGV‑ER, Deep‑ER achieved a 600‑fold speedup, reducing reconstruction time to a few seconds per volume. Quantitatively, signal‑to‑noise ratio (SNR) improved by 12–45 % and Cramer‑Rao lower bounds (CRLB) decreased by 8–50 %, indicating more reliable metabolite quantification. Image quality metrics (PSNR > 38 dB, NMSE < 0.02, SSIM > 0.95) consistently favored Deep‑ER across all test cases.

Clinically, the method produced high‑resolution metabolic maps (N‑AA, Cho, Cr, mI) that clearly delineated tumor heterogeneity and boundaries in glioma patients, suggesting utility for tumor grading, treatment planning, and response monitoring. Importantly, because each spectral point is reconstructed independently, the model generalizes across different acquisition parameters (e.g., echo time, repetition time, field strength, nucleus), making it adaptable to a wide range of MRSI protocols.

In summary, Deep‑ER demonstrates that deep‑learning can surmount the computational bottleneck of non‑Cartesian compressed‑sensing MRSI, delivering near‑instantaneous, high‑quality reconstructions that preserve spectral fidelity essential for accurate metabolite quantification. This advancement paves the way for routine incorporation of high‑resolution spectroscopic imaging into both research and clinical neuro‑imaging workflows, potentially accelerating discoveries in neuroscience and precision medicine.