A Transitional Study for Clinical Implementation of Deep Learning-Accelerated Brain MRI: DWI, FLAIR, and SWI in a Healthy Cohort

Article information

J Neurosonol Neuroimag. 2026;18(1):16-25

Publication date (electronic) : 2026 June 30

doi : https://doi.org/10.31728/jnn.2026.00179

Hee Sang Oh, MD, MS^*, Yo Han Jung, MD, PhD^†, Bio Joo, MD, PhD^*, Mina Park, MD, PhD^*, Sang Hyun Suh, MD, PhD^*, Kwon-Duk Seo, MD, PhD^*, Kyung-Yul Lee, MD, PhD^†, Sung Jun Ahn, MD, PhD^,*

^*Department of Radiology, Gangnam Severance Hospital, Yonsei University College of Medicine, Seoul, Korea

^†Department of Neurology, Gangnam Severance Hospital, Yonsei University College of Medicine, Seoul, Korea

Correspondence: Sung Jun Ahn, MD, PhD Department of Radiology, Gangnam Severance Hospital, Yonsei University College of Medicine, 211 Eonju-ro, Gangnam-gu, Seoul 06273, Korea Tel: +82-2-2019-4551 Fax: +82-2-3462-5472 E-mail: aahng77@yuhs.ac

Received 2026 January 8; Revised 2026 March 14; Accepted 2026 March 16.

Abstract

INTRODUCTION

Magnetic resonance imaging (MRI) is a pivotal diagnostic tool; however, its clinical utility is often hampered by lengthy acquisition times, which increase the likelihood of motion artifacts and patient discomfort.1,2 Although deep-learning-based reconstruction (DLR) has emerged as a promising solution for accelerating imaging without compromising the signal-to-noise ratio (SNR), rigorous technical validation in a controlled environment is a prerequisite before its widespread clinical implementation.3-6

Recent concerns have been raised regarding the potential of using DLR algorithms to treat subtle pathological or anatomical features, such as noise, leading to over-smoothing or loss of critical details,6,7 and it is therefore essential to evaluate the fidelity of accelerated sequences (DWI, FLAIR, and SWI) in preserving the normal brain parenchyma and vascular structures8-10 before their application in clinical practice.

As such, this study aimed to evaluate the technical feasibility and image quality of a deep learning-accelerated MRI protocol and establish a baseline for the structural integrity and image fidelity of the most commonly utilized brain sequences (DWI, FLAIR, and SWI) before broader clinical application. Healthy volunteers were recruited, and the accelerated sequences were compared with those obtained using conventional methods.

SUBJECTS AND METHODS

Imaging protocol

The MRI examinations were performed using a 3 T Siemens scanner (Vida, Erlangen, Germany). Participants underwent both conventional and Accel-DLR brain MRI protocols, including DWI, SWI, and FLAIR. The detailed acquisition parameters for each sequence and protocol, including the repetition time, echo time, flip angle, matrix size, slice thickness, field of view, and acceleration factors, are provided in Table 1.

Table 1.

Detailed MR parameters of DWI, SWI, and FLAIR

Image analysis

Image quality was evaluated through both quantitative and qualitative analyses, following a previously described methodology.3

For quantitative analysis, the SNR on FLAIR, DWI, and SWI was calculated by one radiologist (O.H.S, with four years of experience), using a manually drawn region of interest (ROI) (Fig. 1). One ROI was placed within the normal-appearing white matter to measure the parenchymal signal (S_parenchyma), and a second ROI was placed in the background air outside the cranium to estimate noise (S_noise). The SNR was calculated using the following formula:

Fig. 1.

Representative images from conventional DWI and Accel-DLR DWI protocols. Representative images from conventional DWI (left) and Accel-DLR DWI (right) protocols. To calculate the SNR, one ROI was placed within the normal-appearing white matter to measure the parenchymal signal intensity (S_parenchyma) and a second ROI was positioned in the artifact-free background air to estimate the background noise (S_noise). Accel-DLR, accelerated deep learning-based reconstruction; DWI, diffusion-weighted imaging; ROI, region of interest.

SNR=SparenchymaSnoise

Qualitative image assessments were performed independently by two radiologists who were blinded to the reconstruction technique. Images were presented in a randomized order with a minimum washout period of two weeks between evaluations to minimize recall bias. The readers assessed the images using a Likert scale. Structural delineation was rated on a 3-point scale, while overall image quality and artifact severity were scored on a 5-point scale. In all categories, higher scores indicated superior image quality, clearer anatomical definitions, and fewer artifacts.

For the FLAIR sequence, structural delineation was assessed based on the sharpness of the subarachnoid space and gray-white matter differentiation, as well as the clarity of deep anatomical structures, including the caudate nucleus head, lentiform nucleus, posterior limb of the internal capsule, midbrain, pons, and cerebellum. The FLAIR images were also evaluated for pulsation, CSF flow, and motion artifacts. For the DWI sequence, structural delineation was assessed based on deep gray matter and gray-white matter differentiation, as well as the clarity of the brainstem and cerebellum. The DWI images were also evaluated for air-interface susceptibility and motion artifacts. For the SWI sequence, structural delineation was assessed based on the conspicuity of the cortical vessels and intramedullary veins and the sharpness of the deep nuclei, including the globus pallidus, red nucleus, and substantia nigra. The SWI images were also evaluated for susceptibility to motion artifacts. The overall image quality was comprehensively assessed by evaluating both the general diagnostic quality and the level of overall image noise across all sequences.

RESULTS

Qualitative analysis

The interobserver reliability for qualitative image quality showed substantial to good agreement across all sequences (Table 2). For absolute agreement (ICC[2,1]), the values ranged from 0.60 to 0.84, with Accel-DLR sequences generally demonstrating higher reliability compared to conventional protocols.

Table 2.

Inter-observer reliability for qualitative image quality of conventional and deep learning-accelerated protocols.

For DWI, Accel-DLR images achieved significantly higher scores for overall image quality (4.47±0.12 vs. 3.29±0.25; p<0.001) and structural delineation (2.82±0.11 vs. 2.52±0.12; p<0.001). Artifact scores did not differ between the two methods (4.28±0.12 vs. 4.28±0.08; p=1.000).

For SWI, Accel-DLR resulted in statistically improved overall image quality (4.12±0.28 vs. 3.88±0.28; p=0.021) and structural delineation (2.69±0.22 vs. 2.55±0.21; p=0.015). No statistically significant difference was observed in artifact scores between Accel-DLR and conventional images (4.21±0.18 vs. 4.13±0.22; p=0.059).

For FLAIR, the overall image quality was comparable between Accel-DLR and conventional imaging, with no statistically significant difference (4.00±0.43 vs. 3.91±0.20; p=0.332). However, Accel-DLR demonstrated significantly lower scores for structural delineation (2.68±0.09 vs. 2.85±0.16; p=0.002) and artifact (3.62±0.15 vs. 3.96±0.15; p<0.001) (Figs. 2–4).

Fig. 2.

Qualitative assessment of structural delineation in conventional and Accel-DLR images. Accel-DLR, accelerated deep-learning reconstruction; FLAIR, fluid attenuated inversion recovery; DWI, diffusion weighted image; SWI, susceptibility weighted image; GM, gray matter; WM, white matter. *p<0.05, **p<0.01, ***p<0.001. Asterisks indicate statistical significance.

Fig. 3.

Qualitative assessment of artifacts in conventional and Accel-DLR images. Accel-DLR, accelerated deep-learning reconstruction; FLAIR, fluid attenuated inversion recovery; DWI, diffusion weighted image; SWI, susceptibility weighted image. ***p<0.001. Asterisks indicate statistical significance.

Fig. 4.

Representative cases between conventional and Accel-DLR protocols. (A) On DWI, Accel-DLR demonstrates improved structural delineation of gray-white matter differentiation (arrowheads). (B) Similarly, Accel-DLR enhances the delineation of cerebellar structures on DWI (arrowheads). (C) On SWI, the conspicuity of cortical vessels is increased (arrowheads). (D) On T2-FLAIR, while the Accel-DLR protocol exhibits overall reduced noise, magnified views of the midbrain reveal less distinct boundaries of midbrain structures, such as the red nuclei (arrowheads). Accel-DLR, accelerated deep learning-based reconstruction; DWI, diffusion-weighted imaging; SWI, susceptibility-weighted imaging; FLAIR, fluid-attenuated inversion recovery.

Quantitative analysis

For DWI, Accel-DLR demonstrated a statistically significant increase in SNR compared with the conventional protocol (104.90±41.30 vs. 88.75±26.70; p=0.0028).

For SWI, the mean SNR was higher in the Accel-DLR images than in the conventional images (569.38±541.39 vs. 392.13±229.36). However, this difference was not statistically significant (p=0.108).

In contrast, for FLAIR, the SNR was significantly lower with Accel-DLR than with conventional images (38.37±9.93 vs. 93.27±36.43; p<0.001). A detailed quantitative analysis is presented in Table 3.

Table 3.

Quantitative comparison of FLAIR, DWI, and SWI sequences in conventional and Accel-DLR images

DISCUSSION

Our study demonstrated that deep learning-based reconstruction (Accel-DLR) significantly reduced the acquisition time of core brain sequences (DWI, FLAIR, and SWI) by over 50% while maintaining overall image fidelity with sequence-dependent differences. Unlike previous studies that focused solely on clinical throughput,1,12 we evaluated whether the accelerated algorithm could preserve normal anatomical structures and contrast before it is deployed in pathologic conditions; as such, this work provides foundational technical validation of the algorithm. Although the present study did not directly quantify motion artifacts, the reduction in acquisition time (447.6 to 218.4 seconds) may enhance robustness in clinical settings where patient motion is prevalent. This enhanced efficiency may improve the practical feasibility of multisequence MRI protocols in clinical settings.

Technical evaluation of the accelerated DWI sequence demonstrated that Accel-DLR improved quantitative and qualitative image metrics compared with the conventional accelerated protocol. DWI is widely recognized as a fundamental sequence for detecting diffusion abnormalities in a broad range of neurological conditions.8,13,14 However, accelerated acquisition is often associated with reduced spatial resolution and increased image noise.15,16 In the present study, the Accel-DLR protocol increased the DWI SNR by approximately 18% (p=0.0028) and significantly improved structural delineation (p<0.001).

For DWI, these findings suggest that deep learning reconstruction preferentially reduces background noise while preserving the diffusion-related signal contrast, resulting in sharper anatomical boundaries in qualitative assessments. Given that supervised architectures may suppress subtle low-contrast features, such as early gray-white matter differentiation, our results demonstrated improved structural delineation without evidence of boundary blurring in this healthy cohort. Rather than merely producing a visually smoother appearance, Accel-DLR was associated with an enhanced clarity of the anatomical interfaces. However, although these findings support the preservation of diffusion-related contrast under accelerated conditions, we did not directly evaluate lesion detectability. Therefore, validation in pathological cases is necessary to determine whether these improvements translate into maintained or enhanced detection of subtle lesions.

SWI is widely used in neuroimaging because of its sensitivity to paramagnetic and diamagnetic substances, such as deoxygenated blood and calcium, and it plays an important role in characterizing susceptibility-related contrasts, including those relating to hemorrhage and vascular abnormalities.9,17 In this study, Accel-DLR was associated with a statistically significant improvement in SWI structural delineation (p=0.015). Although the quantitative SNR showed a non-significant increase with relatively large variability, the qualitative assessment demonstrated clearer visualization of the cortical vessels and intramedullary veins. These findings suggest that the susceptibility-related contrast was maintained under accelerated reconstruction in this healthy cohort. Nevertheless, the potential risk of attenuating small susceptibility-related pathological features such as distal thrombi or microbleeds cannot be excluded. Therefore, further evaluation of patients with confirmed pathologies is necessary to determine whether lesion conspicuity is preserved under clinical conditions.

A notable finding of our study was the discrepancy between quantitative and qualitative assessments in FLAIR imaging. Although the overall qualitative image scores were comparable between Accel-DLR and conventional imaging (p=0.332), the SNR was significantly lower for Accel-DLR (38.37 vs. 93.27%, p<0.001). This discrepancy may reflect a known limitation of conventional background-based SNR measurements () in accelerated MRI rather than a direct degradation of image quality. As described by Dietrich et al.10, the spatial and statistical distribution of noise in parallel imaging is governed by the geometry factor, and it is highly dependent on coil configuration and acceleration parameters, such that background air measurements may not accurately represent the true image noise10. Accordingly, background-based SNR measurements can either overestimate or underestimate the effective image quality depending on the image reconstruction method. In this context, the inconsistency between our findings and those of Choi et al., who reported an increased SNR with Accel-DLR,3 may partly reflect reconstruction-dependent noise behavior. Importantly, this methodological limitation applies to both directions of SNR change observed in our study. Thus, the apparent SNR decrease in FLAIR and SNR increase in DWI should be interpreted with caution, as background-based SNR estimation may not fully capture the effective noise characteristics of accelerated or deep learning–reconstructed images. As such, quantitative SNR findings should be interpreted alongside qualitative assessments of structural delineation and overall image quality.

The reduced structural delineation and artifact scores observed for FLAIR with Accel-DLR were most pronounced in deep brain structures, such as the midbrain. Although high-channel receiver arrays generally offer improved g-factor performance at high acceleration factors (R≥3), intrinsic SNR remains lower in central brain regions due to the greater distance from surface coil elements. 18 In this setting, a combination of the reduced intrinsic SNR and acceleration-related noise may reduce the contrast-to-noise ratio of small deep nuclei. Under such conditions, image-domain denoising algorithms may have a limited ability to preserve fine anatomical boundaries, potentially resulting in reduced structural clarity.6,7 These findings suggest that the performance of image-domain deep learning reconstruction is both sequence- and anatomy-dependent, and that caution is warranted when applying high acceleration factors to structures with intrinsically low contrast.

The evaluation of inter-observer reliability demonstrated moderate to high agreement across sequences. Although the absolute agreement (ICC[2,1]) for some sequences was influenced by minor systematic scoring differences between the readers, the consistency measure (ICC[3,1]) remained high across all protocols, particularly for the Accel-DLR sequences (range 0.78–0.85). These findings suggest that despite small differences in absolute scoring thresholds, both radiologists applied similar relative criteria when assessing image quality parameters, such as noise and structural delineation. In this study, the Accel-DLR protocol exhibited reproducible qualitative performance among readers. However, further evaluation in larger and more heterogeneous clinical settings is required to determine the generalizability of our findings.

It is also important to note that conventional and Accel-DLR protocols differ not only in their reconstruction methods but also in several acquisition parameters, including acceleration factors and other sequence parameters. Therefore, the Accel-DLR protocol evaluated in this study represents the combined implementation of accelerated acquisition and deep-learning-based reconstruction, rather than an isolated assessment of the reconstruction algorithm. Changes in the acquisition parameters, particularly higher acceleration factors, may independently influence noise behavior, artifact patterns, and structural delineation. Accordingly, some of the observed differences in image quality between the protocols may reflect the effects of accelerated acquisition itself, in addition to reconstruction-related effects, particularly in sequences such as FLAIR, in which reduced structural delineation and increased artifacts were observed.

This study had several limitations. First, the sample size was small (n=17), limiting the prospective feasibility of this study. The use of a healthy cohort was intended to provide controlled conditions for evaluating reconstruction- related image quality and anatomical fidelity without the confounding effects of pathological lesions or heterogeneous clinical presentations. Although this approach allowed for focused technical assessment, it is not a substitute for validation in patients with diseases. Second, it is important to note that the performance of DLR algorithms may be influenced by specific hardware configurations, such as receiver coil density and gradient strength. Although our results establish a technical foundation for 3T neuroimaging using a specific vendor system, the findings may not be directly generalizable across different MR platforms. Finally, because no pathological cases were included, lesion conspicuity, such as detection of acute ischemia or hemorrhage, was not evaluated. Therefore, subsequent large-scale multicenter trials involving diverse pathological cases, different MR vendors, and various field strengths are essential to confirm the broader clinical reproducibility and diagnostic robustness of this reconstruction approach in real-world clinical practice.

Article information Continued