Brain magnetic resonance imaging serves as a cornerstone of preoperative neurosurgical assessment. Neural fields represent an emerging machine learning approach capable of super-resolution reconstruction and novel view synthesis without requiring large training datasets. Ten 1.5-Tesla brain MRI sequences (nine anisotropic and one isotropic) were used to train patient-specific neural field models using a closed source machine learning framework (Radscaler). Image quality assessment was performed on reconstructions upscaled by factors of 2, 3, and 4 relative to original resolution. The method achieved favorable quality metrics across all scaling factors: mean (sd) SSIM of 0.85 (0.04), MS-SSIM of 0.95 (0.01), and LPIPS of 0.09 (0.04). Neural field reconstruction enabled enhanced visualization of micro-anatomical structures through improved spatial resolution and interpolation of intermediate views not present in the original acquisition. These findings demonstrate that neural fields provide a clinically viable approach for volumetric MRI super-resolution and novel view synthesis, particularly valuable for addressing anisotropic acquisition limitations in neurosurgical planning.
Nicolás González (2025) studied this question.