Image quality evaluation of neonatal brain magnetic resonance imaging using a deep learning reconstruction algorithm: A quantitative and multireader study using variable denoising levels at 3 tesla Authors: Zain Alvi, Eduardo P. Reis, Mateus A. Esmeraldo, David Shin, Suchandrima Banerjee, Hisham M. Dahmoush, Andrew Campion, Stefanie Chambers, Yanniklas Kravutske, Sergios Gatidis, Bruno P. Soares Published: 2026 ABSTRACT Objectives:Neonatal imaging is particularly challenging because newborns are prone to head motion, which can degrade image quality and complicate interpretation. Improving brain magnetic resonance imaging (MRI) image quality may help reduce diagnostic uncertainty and facilitate the nuanced assessment of early myelinating structures in the neonatal brain. Although deep learning reconstruction algorithms designed to improve MRI image quality have been evaluated in pediatric imaging, they have not been specifically studied in exclusively neonatal populations. This pilot study sought to evaluate image quality improvement through the employment of a deep learning reconstruction algorithm in neonatal brain imaging.Materials and Methods:3D T1-weighted brain MRIs were obtained in a small cohort of 15 neonates. A deep learning reconstruction algorithm was applied to the image sets using low, medium, and high levels of denoising. Three radiologists qualitatively rated image quality (signal-to-noise ratio [SNR], presence of artifacts, and overall clarity) on a 4-point scale of eight early myelinating structures. Objective apparent SNR (aSNR) and apparent contrast-to-noise ratio (aCNR), based on signal intensities of white and gray matter, were measured across all three denoising levels.Results:Evaluation by radiologists indicated an overall increase in all image quality categories and increased conspicuity of the early myelinating structures as the level of denoising increased. Objective aSNR and aCNR values also increased progressively with denoising, with significant differences observed for nearly all pairwise comparisons.Conclusion:Our findings suggest improvement in image quality with the use of the deep learning reconstruction algorithm in 3D T1-weighted neonatal brain MRI, though the small sample size and pilot design limit generalizability. Diagnostic accuracy and clinical outcomes were not evaluated and warrant future investigation. INTRODUCTION Neonatal brain magnetic resonance imaging (MRI) is challenging due to the high probability of head motion since it is usually performed with little or no sedation.[1] Motion during image acquisition causes artifacts that may affect the diagnostic capability and confidence of the interpreting radiologist.[1,2] Therefore, fast and motion-robust imaging techniques are desirable, as they may help reduce uncertainty when image quality is compromised.[3] Multiple strategies have been developed to shorten MRI acquisition times, including parallel imaging, compressed sensing, and radial k-space sampling. While these techniques can substantially accelerate imaging, they often come at the cost of reduced signal-to-noise ratio (SNR), residual aliasing or reconstruction artifacts, and occasionally blurring or altered image texture that can obscure subtle pathology.[4] In neonatal brain MR imaging, employing a 3D Gradient Echo T1-weighted sequence is preferable to depict early myelinating brain structures, as myelin appears hyperintense on T1-weighted images.[5] The 3D acquisition also provides higher anatomical details and the flexibility to view the structures in any spatial orientation through reformats. Early myelinating structures are key brain areas to assess as the presence of myelin correlates with functionality. Recently, deep learning reconstruction techniques have shown promise in image quality enhancement by reducing noise, improving SNR, and allowing for shorter image acquisition times.[6] Traditional motion-compensating techniques can be used in tandem with deep learning reconstruction. Together, these approaches may provide higher-quality images and improve the reliability of neonatal neurological diagnoses.[7,8] A specific denoising tool, adaptive image reconstruction deep learning (AIR Recon DL) (GE HealthCare), has shown promise in improving image quality.[9] The AI model was trained to minimize random image noise and Gibbs ringing artifact (also known as truncation artifact) in complex MRI data based on ideal images. By substantially improving SNR and edge sharpness while reducing Gibbs ringing artifacts, the algorithm can enable shorter scan times by allowing operators to optimize the acquisition protocol. AIR Recon DL is a raw-data-based MR reconstruction integrated into the scanner pipeline; it uses a convolutional neural network (4.4 million parameters across around 10,000 kernels) with rectified linear unit activation and no bias terms to perform blind denoising.[9] It was trained on more than 10,000 near-perfect and conventional datasets spanning varied contrasts and shapes and using diverse augmentations (e.g., rotation, Gaussian noise, phase manipulation).[10] The algorithm was designed to function with any patient and pathology.[9] While this tool has been used to optimize adult neurological MRI examinations in cases of tumor identification, and although it has been tested in the pediatric population, it has not yet been specifically investigated in neonatal brain MRI.[6,11] This gap is notable given the unique technical and diagnostic demands of neonatal neuroimaging. Accordingly, the goal of this pilot investigation is to assess the ability of AIR Recon DL to improve image quality in 3D T1-weighted neonatal brain imaging at 3T. MATERIAL AND METHODS Subjects The study was designed as a retrospective pilot case series. A consecutive sample of 15 neonatal scans acquired in May 2024 was selected, irrespective of image quality. These patients underwent clinically indicated T1-weighted MRI at a single medical center. These neonates, who required MRI as part of their clinical care, had no specific exclusion criteria. The institutional review board approved the protocol. The imaging findings of each patient, along with basic demographic data, are summarized in Table 1. Table 1:Patient demographics and MRI findings.Age at MRISexImaging findings4 daysFemaleWhite matter injury in watershed territories and signs of medullary venous congestion or thrombosis16 daysFemaleTrace posterior fossa subdural hemorrhage2 daysMaleModerate supratentorial ventriculomegaly6 daysMaleTrace subdural blood products along the bilateral cerebral convexities7 daysFemaleNormal1 dayMaleLeft retrocerebellar arachnoid cyst7 daysFemalePosterior periventricular white matter lesions7 daysMaleMicrohemorrhages in the cerebellum and parenchymal injury with medullary congestion21 daysMaleExtensive acute parenchymal injury and ectopic posterior pituitary with hypoplastic stalk1 dayMaleAgenesis of the corpus callosum with absent septum pellucidum and ventriculomegaly6 daysMaleLeft parieto-occipital scalp subgaleal and superficial fluid collection4 daysMaleNormal10 daysMaleBilateral deep periatrial white matter injury3 daysMaleLeft retinal detachment with intraventricular blood products and mild ventriculomegaly15 daysMaleIntraparenchymal and intraventricular hemorrhage with ventriculomegalyMRI: Magnetic resonance imaging Patient demographics and MRI findings. MRI: Magnetic resonance imaging Procedure All MRI examinations were performed on 3T GE HealthCare scanners (SIGNA Premier, Discovery MR750, SIGNA positron emission tomography/magnetic resonance, and SIGNA Architect; GE HealthCare, Waukesha, WI) using standard head coils. The acquisition parameters for the T1-weighted axial sequence were repetition time = 7.3 ms, echo time = 3 ms, inversion time = 600 ms, flip angle = 8°, number of excitations = 1, spatial resolution = 0.9 × 0.9 × 1 mm3, bandwidth = 244 Hz/pixel, acceleration = 1.5 × 1, and scan time = 4:30 min. Subsequently, three additional image sets were generated in Digital Imaging and Communications in Medicine format for each examination using the AIR Recon DL deep learning reconstruction algorithm from the raw k-space data for each patient at low, medium, and high denoising levels.[9] In the qualitative analysis, three board-certified radiologists (one pediatric neuroradiologist with 20 years of experience, one pediatric radiologist with 10 years of experience, and one neuroradiology fellow) were recruited to assess each image independently, blinded to one another’s scores. The image sets were anonymized and presented in a random arrangement, blinded to the readers, who were unaware of the reconstruction type and denoising level of each image. The sets included the original acquisition and images processed with low, medium, and high levels of denoising. Radiologists evaluated the images on a 1–4 scale (1 = worst, 4 = best) for perceived SNR, overall image quality (OIQ), and the presence of additional artifacts (ADA) such as artifacts related to motion, field inhomogeneity, and ghosting. ADA assessed artifacts beyond the algorithm’s primary targets (Gibbs ringing and random noise) whose reduction is already captured in the SNR and OIQ scores.[9] ADA therefore specifically evaluates whether the algorithm affects artifact types it was not designed to address. Data collection also involved radiologists rating the conspicuity and confidence in evaluating early myelinating structures, including the medulla, dorsal pons, decussation of superior cerebellar peduncles, inferior colliculi, subthalamic nuclei, globi pallidi, posterior limb of the internal capsules, and perirolandic cortices, using the same 1–4 ranking scale. Interobserver agreement among the three readers was assessed using the intraclass correlation coefficient (ICC, two-way random effects, absolute agreement, single measures) for each evaluated category. For quantitative analysis, two circular 5 mm regions of interest (ROIs) were placed in the right frontal white matter (WM) and right putamen, representative white- and gray-matter structures whose size and signal homogeneity reliably accommodate a 5 mm ROI, using 3D Slicer, following methodology previously validated, without modification.[12-14] ROIs were placed on the axial slice at the level of the basal ganglia where both structures were most clearly delineated, avoiding areas of partial volume averaging, visible artifacts, or anatomical boundaries. The putamen ROI was centered within the largest axial cross-section of the right putamen. The frontal WM ROI was placed on the same axial slice, in a region of homogeneous WM signal free from sulci, vessels, or adjacent gray matter (GM). The same ROI size, slice, and anatomical location were applied consistently across all denoising levels for each subject. The placement was performed by consensus between two board-certified radiologists: A pediatric neuroradiologist with 20 years of experience and a postdoctoral research fellow in pediatric neuroradiology. Quantitative metrics were extracted within both ROIs on a per-case and denoising level basis, with which the following measures were calculated: Apparent SNR (aSNR) and apparent contrast-to-noise ratio (aCNR), based on signal intensities of white and GM.[13] Terms and measures Perceived SNR: The radiologist’s perception of the clarity of the image in differentiating signal from noise. Presence of ADA: Any ADAs present in the image other than Gibbs ringing that could limit diagnostic ability or confidence, such as motion artifacts and field inhomogeneity. OIQ: Radiologist’s overall evaluation of the image based on a combination of perceived SNR, sharpness, and overall artifact burden in a single score. OIQ differs from ADA in scope: while OIQ captures the overall effect of denoising on image quality as a whole, ADA exclusively evaluates artifacts unrelated to the algorithm’s primary function. The two metrics are therefore complementary. aSNR: Calculated based on the local variability within the tissue ROI. where μ is the mean signal intensity and σ is the standard deviation within the ROI.[13] aCNR: Calculated as the ratio of the difference in mean signal intensities between WM and GM to the combined standard deviations in their respective ROIs. where μ is the mean signal intensity and σ is the standard deviation for the respective tissue type.[13] Statistical tests Qualitative scores were analyzed separately for image quality categories and early myelinating structures. A Friedman test assessed the overall effect of denoising level on reader scores, with post hoc pairwise comparisons performed using the Nemenyi test (Original vs. low, low vs. medium, and medium vs. high). P < 0.05 was considered statistically significant. For quantitative metrics, aSNR and aCNR were evaluated across subjects at each denoising level. Per-condition means and standard errors (SEs) were computed and visualized as mean ± standard error (SE). A one-way repeated-measures analysis of variance was conducted for each metric (aSNR in GM, aSNR in WM, and aCNR), with denoising level as the within-subject factor (n = 15). Where significant, post hoc pairwise comparisons were performed using paired t-tests with Holm correction. All analyses and figures were produced in Python (version 3.13.5) using pandas, numpy, matplotlib, SciPy/Stats, and scikit-posthocs packages. RESULTS The scores for the 15 neonatal brain MRIs for each patient with various denoising levels are presented. Table 2 shows the scores for SNR, ADA, and OIQ. Table 3 shows the scores of the individual rankings for the visibility of the different early myelinating structures. Interobserver agreement ranged from fair to substantial agreement across all evaluated categories (ICC = 0.38–0.77); [Table 4]. Table 2:Mean image quality score±standard deviation for SNR, presence of ADA, and OIQ across denoising levels.CategoryOriginalLowMediumHighSNR2.20±0.682.67±0.623.13±0.523.60±0.51ADA2.80±0.862.93±0.883.20±0.773.53±0.74OIQ2.20±0.682.53±0.642.93±0.593.53±0.52The Friedman test revealed a significant effect of denoising level on reader scores (p<0.001). Post hoc Nemenyi pairwise comparisons, performed on scores across all image quality categories, identified significant differences between all denoising level pairs (Original vs. low, low vs. medium, and medium vs. high; p<0.05). ADA: Additional artifacts, SNR: Signal-to-noise ratio, OIQ: Overall image quality Mean image quality score±standard deviation for SNR, presence of ADA, and OIQ across denoising levels. The Friedman test revealed a significant effect of denoising level on reader scores (p<0.001). Post hoc Nemenyi pairwise comparisons, performed on scores across all image quality categories, identified significant differences between all denoising level pairs (Original vs. low, low vs. medium, and medium vs. high; p<0.05). ADA: Additional artifacts, SNR: Signal-to-noise ratio, OIQ: Overall image quality Table 3:Mean image quality score±standard deviation of early myelinating structures (GP, IC, MED, PC, PLIC, PONS, decussation of SCP, SN), at low, medium, and high levels of denoising compared to the original acquisition.CategoryOriginalLowMediumHighGP2.80±0.562.80±0.563.20±0.563.87±0.35IC2.73±0.592.93±0.803.40±0.913.67±0.82MED2.73±0.592.73±0.703.33±0.903.67±0.62PC2.47±0.522.67±0.722.93±0.593.47±0.52PLIC2.53±0.742.67±0.723.07±0.803.73±0.80PONS2.73±0.593.00±0.653.13±0.523.73±0.59SCP2.67±0.722.93±0.703.33±0.823.67±0.82SN2.80±0.772.87±0.833.27±0.703.80±0.56The Friedman test revealed a significant effect of denoising level on visibility scores (p<0.001). Post hoc Nemenyi pairwise comparisons, performed on scores across all structures, identified significant differences between all denoising level pairs (original vs. low, low vs. medium, and medium vs. high; p<0.001). GP: Globi pallidi, IC: Inferior colliculi, MED: Medulla, PC: Perirolandic cortices, PLIC: Posterior limb of the internal capsules, PONS: Dorsal pons, SCP: Decussation of superior cerebellar peduncles, SN: Subthalamic nuclei Mean image quality score±standard deviation of early myelinating structures (GP, IC, MED, PC, PLIC, PONS, decussation of SCP, SN), at low, medium, and high levels of denoising compared to the original acquisition. The Friedman test revealed a significant effect of denoising level on visibility scores (p<0.001). Post hoc Nemenyi pairwise comparisons, performed on scores across all structures, identified significant differences between all denoising level pairs (original vs. low, low vs. medium, and medium vs. high; p<0.001). GP: Globi pallidi, IC: Inferior colliculi, MED: Medulla, PC: Perirolandic cortices, PLIC: Posterior limb of the internal capsules, PONS: Dorsal pons, SCP: Decussation of superior cerebellar peduncles, SN: Subthalamic nuclei Table 4:Interobserver agreement for all evaluated image quality categories and early myelinating structures.CategoryICC95% Confidence intervalp-valueSNR0.3840.086–0.6880.004ADA0.7660.542–0.905<0.001OIQ0.6710.401–0.861<0.001MED0.5000.185–0.768<0.001PONS0.5130.214–0.774<0.001SCP0.6440.371–0.846<0.001IC0.6870.428–0.868<0.001SN0.7070.454–0.878<0.001GP0.6150.331–0.831<0.001PLIC0.6450.352–0.849<0.001PC0.6120.324–0.830<0.001ICC: Intraclass correlation coefficient, SNR: Signaltonoise ratio, ADA: Presence of additional artifacts, OIQ: Overall image quality, MED: Medulla, PONS: Dorsal pons, SCP: Decussation of superior cerebellar peduncles, IC: Inferior colliculi, SN: Subthalamic nuclei, GP: Globi pallidi, PLIC: Posterior limb of the internal capsules, PC: Perirolandic cortices. ICC interpretation: Fair=0.21–0.40, Moderate 0.41–0.60, Substantial 0.61–0.80. p<0.05 was considered statistically significant. Interobserver agreement for all evaluated image quality categories and early myelinating structures. ICC: Intraclass correlation coefficient, SNR: Signaltonoise ratio, ADA: Presence of additional artifacts, OIQ: Overall image quality, MED: Medulla, PONS: Dorsal pons, SCP: Decussation of superior cerebellar peduncles, IC: Inferior colliculi, SN: Subthalamic nuclei, GP: Globi pallidi, PLIC: Posterior limb of the internal capsules, PC: Perirolandic cortices. ICC interpretation: Fair=0.21–0.40, Moderate 0.41–0.60, Substantial 0.61–0.80. p<0.05 was considered statistically significant. The Friedman test revealed a significant overall effect of denoising level on reader scores (p < 0.001). Post hoc Nemenyi pairwise comparisons identified significant differences between all denoising level pairs based on overall rank sums (Original vs. Low, Low vs. Medium, and Medium vs. High; p < 0.05), with scores increasing from original to high denoising [Table 2 and Figure 1]. The same pattern was observed for visibility scores of early myelinating structures (original vs. low, low vs. medium, and medium vs. high; p < 0.001), with scores increasing from original to high denoising [Table 3 and Figure 2]. Image quality assessment of denoised images in the study cohort. Mean overall image quality, signal-to-noise ratio, and presence of additional artifacts were scored by readers for images processed with low, medium, and high-level denoising compared with the original acquisition. ADA: Additional artifacts, OIQ: Overall image quality, SNR: Signal-to-noise ratio. Image quality evaluation of early myelinating brain structures on denoised magnetic resonance imaging acquisitions. Mean overall quality scores of the globi pallidi, inferior colliculi, medulla, perirolandic cortices, posterior limb of the internal capsules, dorsal pons, decussation of the superior cerebellar peduncles, and subthalamic nuclei are shown for images processed with low, medium, and high-level denoising compared with the original acquisition. GP: Globi pallidi, IC: Inferior colliculi, MED: Medulla, PC: Perirolandic cortices, PLIC: Posterior limb of the internal capsules, PONS: Dorsal pons, SCP: Decussation of superior cerebellar peduncles, SN: Subthalamic nuclei. Quantitative analysis further supported the qualitative reader evaluations, demonstrating significant improvements in aSNR and aCNR with increasing denoising levels [Table 5 and Figure 3]. All values are reported as mean ± SE. Mean GM aSNR increased stepwise from 11.77 ± 1.07 in the original images to 21.73 ± 1.82 at the high denoising setting. Similarly, WM aSNR rose from 6.48 ± 0.71 (original) to 12.91 ± 1.07 (high). Apparent CNR between WM and GM also improved consistently, increasing from 3.96 ± 0.31 at baseline to 7.38 ± 0.55 with high denoising. All pairwise comparisons were statistically significant (p ≤ 0.04), with the exception of WM aSNR between the original and the low level of denoising. Table 5:Quantitative evaluation of aSNR and aCNR.MetricOriginal (Mean±SE)Low (Mean±SE)Medium (Mean±SE)High (Mean±SE)Post hoc comparisons (Holm-corrected)aSNR (GM)11.77±1.0713.77±1.2516.93±1.4821.73±1.82All pairwise differences significant (p<0.005)aSNR (WM)6.48±0.717.20±0.519.23±0.6612.91±1.07All significant except Original versus lowaCNR (WM-GM)3.96±0.314.57±0.335.51±0.487.38±0.55All pairwise differences significant (p≤0.04)Values are reported as mean±SE. Mean aSNR in GM, WM, and WM-GM aCNR increased across denoising levels, with all pairwise comparisons statistically significant (p≤0.04) except for WM aSNR between the original and low denoising settings. High-level denoising consistently produced the greatest improvements. SE: Standard error, GM: Gray matter, WM: White matter, aCNR: Apparent contrasttonoise ratio, aSNR: Apparent signaltonoise ratio Quantitative evaluation of aSNR and aCNR. Values are reported as mean±SE. Mean aSNR in GM, WM, and WM-GM aCNR increased across denoising levels, with all pairwise comparisons statistically significant (p≤0.04) except for WM aSNR between the original and low denoising settings. High-level denoising consistently produced the greatest improvements. SE: Standard error, GM: Gray matter, WM: White matter, aCNR: Apparent contrasttonoise ratio, aSNR: Apparent signaltonoise ratio Quantitative evaluation of denoising performance on neonatal brain magnetic resonance imaging. Bar graph shows the mean ± standard error of apparent signal-to-noise ratio (aSNR) in gray matter (GM), aSNR in white matter (WM), and apparent contrast-to-noise ratio (aCNR) across original acquisition and three denoising levels: Low, medium, and high. SE: Standard error. DISCUSSION The results showed that images that undergo higher denoising levels by the deep learning reconstruction algorithm typically displayed better scores in terms of SNR, ADA, and OIQ [Figures 4-6]. These qualitative findings were further confirmed by quantitative analysis, which demonstrated significant improvements in both aSNR and aCNR across all denoising levels [Figure 3]. The denoising level high, specifically, was the most consistent in receiving the highest scores across all categories. Similarly, the highest level of denoising displayed more clarity in identifying the early myelinating structures, suggesting that the AIR Recon DL deep learning reconstruction algorithm may improve image assessment of T1-weighted neonatal brain MRIs. Seven-day-old female neonate with normal brain magnetic resonance imaging findings. Axial T1-weighted images at the level of the pons demonstrate different levels of deep learning denoising: Original acquisition and low, medium, and high levels. Seven-day-old female neonate with normal brain magnetic resonance imaging findings. Axial T1-weighted images at the level of the subthalamic nuclei and globi pallidi demonstrate different levels of deep learning denoising: Original acquisition and low, medium, and high levels. Seven-day-old female neonate with normal brain magnetic resonance imaging findings. Axial T1-weighted images at the level of posterior limbs of the internal capsules demonstrate different levels of deep learning denoising: Original acquisition and low, medium, and high levels. Various studies have explored strategies to mitigate MRI motion artifacts. Retrospective motion correction methods have been used in pediatric and adult populations,[15,16] but they remain limited in addressing severe or unpredictable neonatal motion.[17,18] Prospective motion correction, while promising, faces technical challenges and high costs.[19,20] In this context, AIR Recon DL enables protocol optimizations that can result in shorter examinations, thereby indirectly reducing the likelihood of motion occurring and offering a valuable complementary solution to these existing challenges. Integrating deep learning reconstruction into motion correction pipelines offers new opportunities to enhance image quality. Deep learning reconstruction algorithms leverage large datasets and complex models to potentially improve diagnostic utility.[21,22] Ongoing development of these tools has the potential to reduce sedation and repeat scans in neonates, improving patient safety and lowering imaging costs.[23,24] In neonatal brain MRI, techniques must balance diagnostic accuracy with practicality and safety. Prior studies stress the importance of minimizing scan duration and motion artifacts, which can compromise diagnostic interpretation.[25] Our findings align with prior research supporting deep learning reconstruction as a tool to enhance image quality.[5] In particular, this study evaluated a single vendor-specific algorithm, AIR Recon DL, which limits the generalizability of the results. Therefore, the observed improvements may not directly translate to other systems. Interobserver agreement ranged from fair to substantial (ICC = 0.38–0.77), reflecting the inherent subjectivity of qualitative image quality scoring. The lower agreement observed for SNR in particular suggests that reader-dependent variability may influence qualitative assessments, which limits the reproducibility of these findings and should be considered when interpreting the qualitative results. Several other important limitations must be acknowledged. First, this study evaluated image quality metrics only; diagnostic accuracy and clinical outcomes were not assessed, and it remains unknown whether the observed improvements translate into improved diagnostic performance or patient management, which is an essential question that future prospective studies must address. Second, although the algorithm may enable protocol optimizations that reduce scan time, these potential reductions were not quantified or assessed in this study. However, such reductions have been reported in prior literature, specifically in pediatric brain MRI using 3D T1 sequences in patients ranging from 5 months to 16 years old, demonstrating acquisition time decreases of 29.3% for pre-contrast and 40.7% for post-contrast scans.[14] Third, we did not evaluate how the algorithm affects the detection of specific diseases, limiting conclusions about improvements in diagnostic performance. Future studies should specifically evaluate whether improvements in objective image quality metrics translate into improved diagnostic accuracy, as these are not equivalent outcomes.[26] Of particular concern, and as a key limitation of denoising approaches in this population, is the risk of over-smoothing at higher denoising levels, which could obscure subtle findings such as small lesions.[26] Prior literature has suggested that deep learning reconstruction may reduce relative tissue contrast even when lesion conspicuity is preserved,[14] a finding that raises concern for neonatal imaging specifically, where intrinsic tissue contrast is inherently limited in the developing brain due to ongoing myelination. Therefore, subtle abnormalities, such as early hypoxic–ischemic injury or immature myelination patterns, may be especially susceptible to unintended signal attenuation.[27] This concern is compounded by the scarcity and high variability of neonatal imaging data, which is frequently underrepresented in the training datasets of deep learning reconstruction algorithms. Furthermore, the complex and non-transparent nature of these algorithms makes their behavior on out-of-distribution data difficult to interpret or predict, warranting careful vigilance.[27] In addition, the quantitative metrics used, aSNR and aCNR, are apparent measures based on local ROI signal statistics rather than true SNR, and their interpretation may be affected by changes in noise distribution introduced by deep learning reconstruction.[2] Furthermore, only two ROIs were sampled, which may not fully capture the signal heterogeneity of the neonatal brain.[13] Finally, the sample size (n = 15) is an important limitation of this pilot study. The limited statistical power reduces confidence in the stability of the results, and the cohort may not adequately represent the diversity of neonatal presentations encountered in clinical practice, restricting generalizability. These factors underscore the need for larger, more diverse studies evaluating algorithm performance in real-world clinical settings. CONCLUSION In summary, our results suggest improvement in image quality, SNR, and visualization of early myelination in neonatal brain MRI at 3T with the use of AIR Recon DL, a deep learning reconstruction technique. Future research should focus on quantifying how these improvements translate into real diagnostic and clinical benefits, including whether enhanced image quality can improve diagnostic accuracy for certain diseases, reduce the need for repeat scans due to low image quality, or lessen the need for sedation in neonates.