Target Expert Review Group: Senior Pediatric Radiologists, Perinatologists, and Medical Image Computing Scientists.
Abstract
This retrospective, multi-centric study evaluates the consistency and potential systematic bias introduced by three state-of-the-art super-resolution reconstruction (SRR) pipelines—Neural Slice-to-Volume Reconstruction (NeSVoR), NiftyMIC, and Slice-to-Volume Reconstruction ToolKit (SVRTK)—on quantitative fetal brain MRI measurements. Eighty-four T2-weighted fetal brain scans, collected across three European hospitals using 1.5T and 3T scanners, were reconstructed and assessed.
Quantitative analysis showed that statistically significant variations in 2-D biometric measures, performed by four expert raters, consistently remained negligible, falling below the 0.8 mm isotropic voxel width. Multivariate analysis confirmed that inter-rater variability contributed larger, albeit still small, effects (up to 1.55 mm for skull biparietal diameter) than the choice of SRR algorithm. Automated 3-D volumetry revealed small systematic effects, generally around 1%, with the largest deviation (2.7%) noted for extra-cerebral cerebrospinal fluid (CSF).
Qualitative assessment by four neuroradiologists indicated systematic differences in the visual quality of the reconstructed volumes, particularly concerning white matter intensity and sharpness, with SVRTK and NiftyMIC generally preferred over NeSVoR, which often produced white matter alterations. Experts were hesitant to fully substitute SRR volumes for low-resolution stacks in primary radiological assessment. The findings support the pooling of quantitative measurements from studies utilizing different high-quality SRR methods across varied acquisition settings (scanners, centers, raters), thereby facilitating the construction of large-scale normative neurodevelopmental models.
Biometry and Volumetry in Multi-Centric Fetal Brain Magnetic Resonance Imaging
Study Design and Scope (0:00): This retrospective, multi-centric study investigated T2-weighted fetal brain MRI scans (2009–2023) from 84 healthy subjects across three hospitals (H1, H2, H3) to assess measurement consistency across three super-resolution reconstruction (SRR) pipelines: NeSVoR, NiftyMIC, and SVRTK.
Data Characteristics (0:00): Subjects were distributed across three gestational age (GA) bins ([21, 28) weeks, [28, 32) weeks, [32, 36) weeks). Data were acquired using Siemens 1.5T and 3T scanners with variable low-resolution settings (e.g., in-plane resolution 0.55–1.12 mm; slice thickness 2.8–3.5 mm). All reconstructions were performed at 0.8 mm isotropic resolution.
Biometric Measurements (13:00): Five standard 2-D biometric measurements were performed by four clinical experts: Length of the Corpus Callosum (LCC), Height of the Vermis (HV), brain and skull Biparietal Diameters (bBIP, sBIP), and Transverse Cerebellar Diameter (TCD).
Biometric Results (Table 2): Biometric analysis demonstrated statistically significant differences induced by the SRR methods (e.g., $P<0.001$ for sBIP and TCD). Crucially, these differences consistently remained small, below the 0.8 mm voxel width (e.g., maximum median difference of 0.4 mm for sBIP).
Rater Variability: Multivariate analysis using a GAMLSS model confirmed that rater-related effects were consistently larger than SRR effects, with the maximum observed variability attributable to the rater being 1.55 mm (2.5% variability) for sBIP.
Automated Volumetry (10:00): Automated 3-D volumetry was performed using the deep learning-based Brain vOlumetry and aUtomated parcellatioN (BOUNTI) method, measuring five structures: extra-cerebral CSF, Cortical Gray Matter (GM), Cerebellum, Supratentorial Brain Tissue (ST), and total Lateral Ventricles.
Volumetric Results (Table 3): Volumetric measurements showed small but systematic variability between SRR methods, generally in the order of 1% to 2%. The largest observed deviation was 2.7% for extra-cerebral CSF (NeSVoR vs NiftyMIC). Growth curves generally aligned with prior literature, though Cortical GM was consistently overestimated compared to Kyriakopoulou et al. [16] and underestimated compared to Machado-Rivas et al. [28].
Qualitative Assessment (Table 4, 5): Neuroradiologists qualitatively assessed six subjects. NeSVoR performed poorly in white matter assessment (layering and intensity) and blurriness compared to SVRTK and NiftyMIC. Experts often rated SVRTK images as being of sufficiently good quality, but overall consensus showed hesitation to fully use SRR volumes in place of low-resolution stacks for clinical evaluation.
Conclusion on Bias (Discussion): The study concludes that the choice of SRR method does not introduce large systematic biases in 2-D or 3-D measurements when sufficient quality is achieved. This consistency across centers, scanners, and SRR pipelines supports integrating derived quantitative data into unified normative frameworks for prenatal neurodevelopmental studies, despite observed textural differences that influence expert perception.