Visualization of the Trochlear Nerve Using Deep Learning–enhanced 3D T2-weighted MR Imaging at 3T

Abstract

Purpose

Cranial nerve imaging with 3T MRI commonly uses 3D fast imaging employing steady-state acquisition (3D-FIESTA); however, this sequence has limitations in achieving higher spatial resolution and provides poor tissue contrast between cranial nerves and adjacent vascular structures. We evaluated 3D T2-weighted imaging (T2-CUBE) with deep learning–based reconstruction (DLR) for cranial nerve visualization, focusing on the trochlear nerve, the smallest cranial nerve with the longest intracranial course, and compared it with T2-CUBE without DLR and 3D-FIESTA.

Methods

Ten healthy male volunteers (age, 23–40 years; mean age, 32 years) underwent T2-CUBE with and without DLR, and 3D-FIESTA at 3T. Two neuroradiologists independently evaluated trochlear nerve visualization in 4 anatomical segments (origin from the midbrain, cisternal, tentorial, and anterior portion of its cavernous segments) using a 3-point scale, and SNR of the pons (SNR_PONS) and cerebrospinal fluid (SNR_CSF) were calculated.

Results

T2-CUBE with DLR achieved a 100% visualization across all trochlear nerve segments and demonstrated significantly better visualization than both T2-CUBE without DLR and 3D-FIESTA (P < 0.025). T2-CUBE without DLR showed 67.5%–100% visualization across the 4 segments. 3D-FIESTA showed 32.5%–80% visualization of the origin from the midbrain, cisternal, and tentorial segments, with no visualization of the cavernous segments. DLR increased SNR_PONS and SNR_CSF by factors of 1.8–2.5 (SNR_PONS: 14.1 vs 5.7; SNR_CSF: 31.8 vs 17.5, respectively; P < 0.001). T2-CUBE with DLR demonstrated significantly higher SNR_PONS than 3D-FIESTA (14.1 vs 6.4, P < 0.001), while SNR_CSF was comparable (31.8 vs 36.3, P = 0.20).

Conclusion

T2-CUBE with DLR at 3T provided a significantly better trochlear nerve visualization than T2-CUBE without DLR and 3D-FIESTA. This technique may extend beyond the trochlear nerve to other cranial nerves and to the evaluation of neurovascular compression in the cistern, with the potential to become the new standard for cisternal imaging.

Introduction

Cranial nerves in the cistern are involved in various pathological conditions, including neurovascular compression, acoustic neuromas, and inflammatory diseases.¹ MRI is the standard modality for visualizing the anatomical course and pathological changes of cranial nerves. On 3T MRI, 3D balanced steady-state free precession (bSSFP) sequences, such as 3D fast imaging employing steady-state acquisition (3D-FIESTA) and constructive interference in the steady state, have demonstrated high sensitivity for cranial nerve visualization.^2–6 These sequences provide excellent contrast between the cerebrospinal fluid and neural structures, enabling visualization of the cranial nerves. However, bSSFP sequences have inherent limitations, including sensitivity to flow and magnetic susceptibility artifacts. In addition, their tissue contrast is relatively low, sometimes making it difficult to distinguish small cranial nerves from adjacent vascular structures. Moreover, achieving higher spatial resolution requires prolonged TRs, which extend TEs, resulting in image quality degradation and banding artifacts.⁷ These limitations restrict higher spatial resolution with bSSFP sequences, necessitating alternative approaches for cranial nerve imaging. Compared with 3T imaging, 7T MRI using 2D T2-weighted imaging (T2WI) sequences offers significant improvements in cisternal cranial nerve depiction.^2,8 It provides superior visualization of fine structures, such as the trochlear nerve. However, 7T MRI remains limited to specialized centers because of high capital investment and maintenance requirements, which restrict its widespread clinical application. Therefore, to further enhance cranial nerve visualization in routine clinical practice, it is essential to develop novel applications of high-resolution T2 contrast that utilize the widely available 3T MRI systems. However, it has been pointed out that 3D T2-weighted fast spin-echo imaging (commonly implemented as CUBE on GE HealthCare systems) may have a reduced spatial resolution resulting from imaging time constraints.

In MRI, spatial resolution and SNR are inversely related. Improving spatial resolution increases noise levels, traditionally requiring either acceptance of image quality degradation or extension of acquisition times to clinically impractical durations to achieve an adequate SNR.^1,9 To address this challenge, deep learning–based reconstruction (DLR) technology, specifically GE HealthCare’s AIR Recon DL (ARDL), has been developed. This technique achieves effective noise reduction while preserving sharp edges, thereby improving image quality in thin-slice MRI acquisitions.^10–12 Iwamura et al. reported that applying DLR to 2D T2WI enabled high-resolution, high-quality evaluation in patients with multiple sclerosis,¹³ whereas Ishimoto et al. demonstrated improved visualization of pituitary adenoma evaluation using high-resolution 3D T1-weighted imaging combined with DLR.¹⁴ Based on these findings, we hypothesized that combining 3D T2WI with DLR on 3T MRI would enable high-resolution, high-quality cisternal imaging within clinically acceptable acquisition times.

In this study, we chose the trochlear nerve as our evaluation target. It is the smallest cranial nerve (diameter 0.3–1 mm) and has the longest intracranial course (60 mm).^8,15,16 In addition, it follows a complex anatomical pathway, coursing through the trochlear groove within the tentorium cerebelli and subsequently through the cavernous sinus,^8,17 making it an optimal target for the evaluation of high-resolution imaging techniques. Previous studies using 3T MRI have reported visualization of the trochlear nerve;^1,16,18,19 however, these studies were primarily limited to the brainstem and cisternal segments. A comprehensive depiction of the entire trochlear nerve pathway, from the brainstem to the cavernous sinus, has been demonstrated only with 7T MRI.⁸ Therefore, in this study, we aimed to compare the trochlear nerve visualization capabilities of T2-CUBE with DLR, T2-CUBE without DLR, and conventional 3D-FIESTA sequences at 3T MRI and establish an optimal imaging protocol for clinical applications.

Materials and Methods

The Institutional Review Board of Hirosaki University approved this study involving healthy volunteers (No. 2025-019). This prospective study was conducted in accordance with the principles of the Declaration of Helsinki. Written informed consent was obtained from all participants prior to the scans. In this study, the corporate co-authors (A.N. and T.W.) provided support only for MRI acquisition technique optimization and ARDL technology and were not involved in image dataset preparation, image evaluation, or statistical analysis.

Participants

This prospective study included 10 healthy male volunteers (age, 23–40 years; mean age, 32 years) without any ocular movement disorders. Self-reported medical histories included cerebral aneurysm (n = 1), sinusitis (n = 1), and Kawasaki disease (n = 1); the remaining 7 participants reported no relevant medical history.

MRI sequences

MRI was performed using a 3T system (SIGNA Premier 3.0T; GE HealthCare, Milwaukee, WI, USA). All volunteers underwent T2-CUBE sequences with and without DLR, a vendor-supplied deep-learning algorithm (ARDL; GE HealthCare), as well as conventional 3D-FIESTA. Both DLR and non-DLR images were reconstructed from the same T2-CUBE dataset acquired in a single scan. The acquisition parameters for each sequence are presented in Table 1. All sequences were obtained in the coronal plane, including the midbrain, pons, and upper medulla. To determine the imaging parameters for 3D-FIESTA, a pilot study was conducted in which 1 volunteer underwent imaging at 3 different spatial resolutions: 0.4 mm isovoxel (acquisition time: 11 mins 19s), 0.6 mm isovoxel (acquisition time: 7 mins 35s), and 0.8 mm isovoxel (acquisition time: 4 mins 1s). During the 0.4 mm isovoxel acquisition, the volunteer reported significant chest discomfort, presumably due to peripheral nerve stimulation. Based on safety concerns and the prolonged acquisition time, we determined that clinical implementation at 0.4 mm resolution was not feasible. Thus, in this study, we used the coronal 3D-FIESTA imaging with parameters as follows: voxel size, 0.6 × 0.6 × 0.6 mm; acquisition time, 7 mins 35s. The coronal T2-CUBE with and without DLR parameters were as follows: voxel size, 0.3 × 0.3 × 0.8 mm; acquisition time, 8 mins 15s. Theoretically, an acquisition time of 16 mins 30s would be required for T2-CUBE without DLR to achieve SNR equivalent to that of T2-CUBE with DLR. However, such a prolonged acquisition time raised concerns regarding volunteer burden and the potential impact of motion artifacts; therefore, this acquisition was not performed in the present study. Consequently, the 3 imaging methods (T2-CUBE with DLR, T2-CUBE without DLR, and 3D-FIESTA) were compared with approximately equivalent acquisition times.

Table 1.

Magnetic resonance imaging acquisition parameters

	3D-T2WI	3D-FIESTA
Pulse sequence	CUBE	FIESTA-C
Acquisition plane	Coronal	Coronal
TR (ms)	2000	4.8
TE (ms)	100	1.9
RF (degrees)	VRFA	60
FOV (mm)	120	180
Matrix	400 × 400	300 × 300
In-plane resolution (mm)	0.3 × 0.3	0.6 × 0.6
Slice thickness (mm)	0.8	0.6
Slab thickness (mm)	170	300
Echo train length	80	—
BW (Hz/pixel)	312.5	606.1
Parallel imaging factor	2	—
Number of signals averaged	1	1
DLR	+/–	–
Scan time	8:15	7:35

3D-FIESTA, 3D fast imaging employing steady-state acquisition; BW, bandwidth; DLR, deep learning–based reconstruction; RF, refocus flip; VRFA, variable refocus flip angle.

Principle of the AIR recon DL algorithm

The details of the ARDL algorithm have been described by Lebel.²⁰ In brief, the ARDL used in this study is a deep learning–based MRI reconstruction technique that processes raw k-space data to produce high-quality images with reduced noise, suppressed Gibbs ringing artifacts, and improved edge sharpness. This algorithm is integrated directly into the scanner’s reconstruction pipeline, enabling the generation of both conventional and ARDL-reconstructed images from a single set of raw k-space data. The system also provides user-controlled denoising levels and is designed to operate in a scale-invariant and robust manner across various anatomical regions. Trained on millions of augmented image pairs, ARDL has acquired generalization capability across diverse imaging conditions and contributes to improved diagnostic confidence compared to conventional reconstruction methods.¹⁴

Image analyses

Quantitative evaluation of image quality

For the 3 images (T2-CUBE with and without DLR, as well as 3D-FIESTA), the SNR of the pons and cerebrospinal fluid (CSF) were calculated by the consensus of 2 neuroradiologists (S.I., with 17 years of experience in neuroradiology, and T.K., with 4 years of experience in neuroradiology) according to a previous study.¹ Circular or ovoid ROIs with diameters of 3–5 mm were placed on the pons and CSF in the cerebellopontine cistern on coronal images at the level where the facial nerve detachment point was visible (Fig.1). For CSF measurements, the side with more uniform signal was selected. This coronal plane at the level of the facial nerve detachment point was chosen because the cistern was relatively wide, allowing ROI placement without including vessels or nerves, and because flow artifacts were minimal, resulting in more homogeneous signal intensity compared to other cisternal regions. Consistency of ROI location was ensured across the 3 images. The mean and standard deviation (SD) of signal intensity for each ROI were recorded.

Fig. 1.

The example of placements of ROIs (circles) for quantitative analyses on T2-CUBE with DLR. Circular or ovoid ROIs with diameters of 3–5 mm were placed on the pons (white circle) and cerebrospinal fluid in the cerebellopontine cistern (black circle) on coronal images at the level where the facial nerve detachment point was visible.CUBE, 3D fast spin-echo sequence (GE HealthCare proprietary name); DLR, deep learning–based reconstruction.

The SNR of the pons (SNR_PONS) and CSF (SNR_CSF) were calculated as the ratio of mean signal intensity to the SD of the pons and CSF, respectively.

Qualitative evaluation of image quality

Trochlear nerve visualization was evaluated by 2 neuroradiologists, A (K.W.) and B (S.I.), each with 17 years of experience in neuroradiology. All imaging sequences were evaluated using multiplanar reconstruction and reviewed sequentially in the following order: 3D-FIESTA, T2-CUBE without DLR, and T2-CUBE with DLR. Within each sequence, all cases were randomized by neuroradiologist C (T.K., with 4 years of experience in neuroradiology) before being presented to the neuroradiologist A and B for evaluation. A 10-day washout period was implemented between reading sessions to minimize recall of previous evaluations. Additionally, the neuroradiologists were blinded to the sequence names presented at each session to minimize learning effects and expectation bias. Each trochlear nerve was independently evaluated by 2 neuroradiologists following a standardized anatomical and rating protocol.

The trochlear nerve course was anatomically classified into 4 segments: origin from the midbrain, cisternal segment, tentorial segment, and anterior portion of the cavernous segment.⁸ Trochlear nerve visualization was graded on an established 3-point scale (0–2):^2,18 0 = not identified (trochlear nerve could not be identified); 1 = most probably identified (partially visible but complete course could not be depicted); and 2 = clearly identified (complete depiction with high confidence). A total of 80 segments were evaluated (4 segments × 2 sides × 10 volunteers). For cases rated as “0: not identified,” the reasons for non-identification were documented, including flow artifacts, poor image quality, insufficient spatial resolution, and other factors.

Statistical analyses

Statistical analyses were performed using R software (version 4.1.2; https://www.r-project.org/). Image quality scores from both neuroradiologists were analyzed, and interobserver agreement was evaluated using Cohen’s weighted κ values. The strength of agreement was classified according to κ values: fair (0.21–0.40), moderate (0.41–0.60), good (0.61–0.80), and excellent (> 0.80).²¹ When the κ value was > 0.6, indicating sufficiently high interobserver agreement, the evaluation data from both neuroradiologists were combined for the analysis. The scores from the qualitative image analyses comparing T2-CUBE with DLR versus T2-CUBE without DLR and T2-CUBE with DLR versus 3D-FIESTA were assessed using the Wilcoxon signed-rank test. The Bonferroni correction was applied to adjust for 2 pairwise comparisons, resulting in a corrected significance threshold of P < 0.025. The SNR was compared using the paired t test. The ratio was calculated by dividing the ratio of the SNR of the T2-CUBE with DRL over the SNR of the T2-CUBE without DRL.

Results

Quantitative evaluation of image quality

The results of the SNR measurements for the 3 images (T2-CUBE with and without DLR, as well as 3D-FIESTA) are shown in Table 2. The SNR_PONS and SNR_CSF of T2-CUBE were increased by a factor of 1.8–2.5 by applying DLR (mean SNR_PONS: T2-CUBE with DLR 14.1 vs T2-CUBE without DLR 5.7, P < 0.001; mean SNR_CSF: T2-CUBE with DLR 31.8 vs T2-CUBE without DLR 17.5, P < 0.001).

Table 2.

SNR measurements in pons and CSF

	T2-CUBE with DLR	T2-CUBE without DLR	3D-FIESTA	T2-CUBE with DLR is compared to (P value)
				T2-CUBE without DLR	3D-FIESTA
SNRPONS	14.1 (12.0–16.2)	5.7 (5.1–6.4)	6.4 (5.5–7.3)	< 0.001	< 0.001
SNRCSF	31.8 (28.0–35.6)	17.5 (16.0–19.0)	36.3 (30.8–41.7)	< 0.001	0.20

Data are means with 95% confidence intervals in parentheses. The SNR of the pons (SNR_PONS) and CSF (SNR_CSF) were calculated as the ratio of mean signal intensity to the standard deviation of the pons and CSF, respectively. The SNR were compared using the paired t-test. 3D-FIESTA, 3D fast imaging employing steady-state acquisition; CSF, cerebrospinal fluid; CUBE, 3D fast spin-echo sequence (GE HealthCare proprietary name); DLR, deep learning–based reconstruction.

T2-CUBE without DLR showed lower values for both SNR_PONS and SNR_CSF compared to 3D-FIESTA; however, SNR increased with the application of DLR. T2-CUBE with DLR demonstrated significantly higher SNR_PONS compared to 3D-FIESTA (mean SNR_PONS: T2-CUBE with DLR 14.1 vs 3D-FIESTA 6.4, P < 0.001). In contrast, SNR_CSF was almost equal to or slightly lower than that of 3D-FIESTA, with no statistically significant difference (mean SNR_CSF: T2-CUBE with DLR 31.8 vs 3D-FIESTA 36.3, P = 0.20).

Qualitative evaluation of image quality

Overall assessment

The qualitative evaluation results of the trochlear nerve segments by neuroradiologists A and B are presented in Table 3. The interobserver agreement showed a weighted κ value of 0.823, indicating excellent agreement. Because κ > 0.6, the results from both the neuroradiologists were combined. The percentages of cases rated 1 or 2 (probable or definite) are shown in Table 4. Using T2-CUBE with DLR, all trochlear nerve segments achieved a 100% rate of cases evaluated as 1 or 2. T2-CUBE with DLR demonstrated significantly better visualization than T2-CUBE without DLR and 3D-FIESTA for all segments (P < 0.025).

Table 3.

Qualitative evaluation of trochlear nerve segment image quality by 2 neuroradiologists

		T2-CUBE with DLR	T2-CUBE without DLR	3D-FIESTA
		(2▪1▪0)	(2▪1▪0)	(2▪1▪0)
Origin from the midbrain	A	11▪9▪0	9▪4▪7	8▪4▪8
	B	11▪9▪0	9▪5▪6	10▪5▪5
Cisternal segment	A	14▪6▪0	11▪8▪1	12▪3▪5
	B	14▪6▪0	11▪4▪5	10▪7▪3
Tentorial segment	A	20▪0▪0	11▪9▪0	0▪6▪14
	B	20▪0▪0	10▪10▪0	1▪6▪13
Anterior portion of its cavernous segment	A	13▪7▪0	5▪10▪5	0▪0▪20
	B	12▪8▪0	7▪10▪3	0▪0▪20

The data are presented as the number of cases with visualization scores of 2/1/0 for each sequence. The visualization scores were defined as follows: 2 = clearly identified, 1 = most probably identified, and 0 = not identified. Each neuroradiologist evaluated 20 trochlear nerve segments (10 volunteers × 2 sides) in each sequence. 3D-FIESTA, 3D fast imaging employing steady-state acquisition; A, neuroradiologist A; B, neuroradiologist B; CUBE, 3D fast spin-echo sequence (GE HealthCare proprietary name); DLR, deep learning–based reconstruction.

Table 4.

Combined visualization rates of the trochlear nerve segments using each imaging sequence

	T2-CUBE with DLR	T2-CUBE without DLR	3D-FIESTA	T2-CUBE with DLR is compared to (P value*)
				T2-CUBE without DLR	3D-FIESTA
Origin from the midbrain	100%	67.5%	67.5%	< 0.025	< 0.025
Cisternal segment	100%	85%	80%	< 0.025	< 0.025
Tentorial segment	100%	100%	32.5 %	< 0.025**	< 0.025
Anterior portion of its cavernous segment	100%	80%	0%	< 0.025	< 0.025

Data represent the percentage of cases with visualization scores ≥1 (most probably or clearly identified) based on combined evaluations from both radiologists after confirming adequate interobserver agreement (κ > 0.6). Statistical comparisons were performed using the Wilcoxon signed-rank test with a Bonferroni correction. *Statistically significant difference (P < 0.025). **For the tentorial segment, both T2-CUBE with and without DLR achieved 100% visualization (score ≥1). The statistically significant difference favoring T2-CUBE with DLR reflects an improvement in the distribution of ordinal scores (0, 1, 2) as assessed by the Wilcoxon signed-rank test, rather than the binary detection rate.3D-FIESTA, 3D fast imaging employing steady-state acquisition; CUBE, 3D fast spin-echo sequence (GE HealthCare proprietary name); DLR, deep learning–based reconstruction.

Comparison between T2-CUBE with and without DLR

Compared with T2-CUBE without DLR, T2-CUBE with DLR demonstrated a significantly better visualization of all trochlear nerve segments (P < 0.025). For T2-CUBE without DLR, the percentages of cases rated 1 or 2 were 67.5% at the origin from the midbrain (Fig. 2), 85% at the cisternal segment (Fig. 3), and 80% at the anterior portion of the cavernous segment, while some trochlear nerves were evaluated as 0 (not identified). The reasons for non-identification were poor image quality and insufficient spatial resolution. Using T2-CUBE with DLR, no trochlear nerve segments were evaluated with 0. All segments that had been evaluated 0 on T2-CUBE without DLR improved to 1. For the tentorial segment, almost half of the cases were rated 1, and all the remaining cases were rated 2 for T2-CUBE without DLR (Fig. 4). However, all T2-CUBE with DLR cases were rated 2, as all segments previously rated 1 were upgraded to 2.

Fig. 2.

Visualization of the trochlear nerve at the origin from the midbrain. Coronal T2-CUBE with DLR (a), T2-CUBE without DLR (b), and 3D-FIESTA (c) images showing the trochlear nerve at the origin from the midbrain (arrows) (30s, male). The nerve is the most clearly visualized with T2-CUBE with DLR (a) compared to T2-CUBE without DLR (b) and 3D-FIESTA (c).3D-FIESTA, 3D fast imaging employing steady-state acquisition; CUBE, 3D fast spin-echo sequence (GE HealthCare proprietary name); DLR, deep learning–based reconstruction.

Fig. 3.

Visualization of the trochlear nerve in the cisternal segment. Coronal T2-CUBE with DLR (a), T2-CUBE without DLR (b), and 3D-FIESTA (c) images showing the trochlear nerve in the cisternal segment (arrows) (30s, male). T2-CUBE with DLR (a) demonstrates clear visualization of the nerve, while T2-CUBE without DLR (b) shows partial visualization with surrounding noise. 3D-FIESTA (c) depicts the nerve with poor delineation owing to low spatial resolution.3D-FIESTA, 3D fast imaging employing steady-state acquisition; CUBE, 3D fast spin-echo sequence (GE Healthcare proprietary name); DLR, deep learning–based reconstruction.

Fig. 4.

Visualization of the trochlear nerve in the tentorial segment. Coronal T2-CUBE with DLR (a), T2-CUBE without DLR (b), and 3D-FIESTA (c) images of the trochlear nerve in the tentorial segment (30s, male). The nerve within the trochlear groove of the tentorium cerebelli is clearly identified with T2-CUBE with DLR (a, arrow) and the most probably identified with T2-CUBE without DLR (b, arrow) but is not identified with 3D-FIESTA (c) owing to merging with the surrounding structures.3D-FIESTA, 3D fast imaging employing steady-state acquisition; CUBE, 3D fast spin-echo sequence (GE HealthCare proprietary name); DLR, deep learning–based reconstruction.

Comparison between T2-CUBE with DLR and 3D-FIESTA

T2-CUBE with DLR showed significantly better visualization than 3D-FIESTA across all segments (P < 0.025). While T2-CUBE with DLR achieved 100% visualization for all segments, 3D-FIESTA achieved only 67.5%–80% visualization for the origin from the midbrain and cisternal segments, with the remaining cases rated 0 because of low spatial resolution and the influence of flow artifacts. For the tentorial segment, visualization with 3D-FIESTA was 32.5%, and 67.5% of the cases were rated as 0. No cavernous segment cases were visualized using 3D-FIESTA (Fig. 5). The reasons for non-identification included the inability to separate the nerve from the tentorial structures because of their proximity and poor contrast between the nerve and cavernous sinus veins.

Fig. 5.

Visualization of the trochlear nerve in the anterior portion of the cavernous segment. Coronal T2-CUBE with DLR (a), T2-CUBE without DLR (b), and 3D-FIESTA (c) images of the trochlear nerve in anterior portion of the cavernous segment (30s, male). The nerve (arrow) is clearly visualized only with T2-CUBE with DLR (a). T2-CUBE without DLR (b) shows partial visualization with difficulty in distinguishing the nerve from the surrounding noise. 3D-FIESTA (c) does not depict the nerve owing to poor contrast with surrounding structures.3D-FIESTA, 3D fast imaging employing steady-state acquisition; CUBE, 3D fast spin-echo sequence (GE HealthCare proprietary name); DLR, deep learning–based reconstruction.

Discussion

In this study, we evaluated the performance of 3D T2WI with DLR on 3T MRI for cranial nerve visualization in healthy volunteers. To the best of our knowledge, this is the first report to assess T2-CUBE with DLR for cranial nerve imaging. T2-CUBE with DLR demonstrated significantly better visualization of the trochlear nerve than 3D-FIESTA and T2-CUBE without DLR.

Previous studies using 3T MRI have shown limitations in consistently identifying the trochlear nerve throughout its course.^1,16,18,19 Similarly, in our study, T2-CUBE without DLR at 3T demonstrated insufficient nerve visualization. Kumar et al. reported that the visualization rates of the trochlear nerve using 7T MRI were 65% at the origin from the midbrain, 93% at the cisternal segment, 100% at the tentorial segment, and 74% at the cavernous segment.⁸ In their study, coronal 2D T2WI turbo spin echo was performed with a 1.2 mm slice thickness, 0.34 × 0.34 mm in-plane resolution, and an 8 mins acquisition time. We used coronal 3D T2WI CUBE with a 0.8 mm slice thickness, 0.3 × 0.3 mm in-plane resolution, and an 8 mins 15s acquisition time, comparable to the parameters used in the 7T study. Consequently, our 3T T2-CUBE with DLR enabled visualization of the trochlear nerve, which was considered difficult with 3T T2-CUBE, and achieved an image quality comparable to that of the previous 7T study.

In this study, the application of DLR resulted in significant improvement in SNR. This is considered to be attributable to the noise reduction effect of DLR, consistent with previous reports.^1,13,14 While 7T MRI inherently provides higher SNR than 3T MRI,^22,23 conventional 3T MRI requires significantly prolonged acquisition times to achieve an SNR equivalent to that of 7T MRI. DLR improves SNR and reduces noise, as reported by Ishimoto et al. and Iwamura et al.,^13,14 allowing the acquisition of high-SNR, low-noise images with acquisition times similar to those of 7T MRI. Moreover, improved SNR enabled by DLR was physically attributable to the smaller voxel size. Therefore, the high spatial resolution of 3D imaging further enhances the visualization of the trochlear nerve. As indicated by Choi et al. and Kumar et al., the trochlear nerve has a small and complex anatomy, making imaging with voxel sizes smaller than the nerve diameter crucial.^8,16 While the 7T study used 2D imaging, our study used 3D imaging with smaller voxel volumes, enabling a clearer depiction of the trochlear nerve and surrounding fine structures. In addition, 3D imaging allows the creation of multiplanar reconstructed images, enabling a more detailed evaluation of the anatomical structures.

The technical superiority of the CUBE sequence likely explains why T2-CUBE with DLR demonstrated significantly better visualization than 3D-FIESTA, the current gold standard, across all segments. CUBE is a fast spin-echo sequence that is inherently less susceptible to flow artifacts than 3D-FIESTA, a bSSFP sequence.² Notably, achieving ultra-high-resolution imaging with bSSFP sequences is inherently challenging. In bSSFP sequences, TE is set to TR/2; therefore, prolonging TR to achieve higher image quality inevitably extends TE. This TE prolongation enhances T2* decay effects, causing echo signal attenuation,⁷ which leads to image quality degradation, even with extended acquisition times. In addition, the parameters required to achieve a high spatial resolution with bSSFP imaging may induce neural stimulation, limiting its use from a patient safety perspective. Consequently, substantial improvement in the spatial resolution of 3D-FIESTA remains challenging, and T2-CUBE with DLR has the potential to become the gold standard for cisternal imaging.

The visualization improvement achieved using T2-CUBE with DLR may extend beyond the trochlear nerve to other cranial nerves and vessels coursing through the cistern, such as the facial, trigeminal, and glossopharyngeal nerves. Although future studies are needed to evaluate its utility in clinical cases such as facial nerve vascular compression and cisternal tumors involving the cranial nerves, T2-CUBE with DLR is presumed to be superior to 3D-FIESTA in providing T2 contrast. In this study, 3D-FIESTA showed poor contrast between the nerves, vessels, and brainstem, making it difficult to distinguish these adjacent structures. Conversely, the CUBE sequence provided clearer T2 contrast among individual structures, enabling a detailed depiction of the complex anatomical course of the trochlear nerve.

In this study, favorable results were obtained using DLR developed by GE Healthcare (ARDL). Noise reduction technologies based on DLR have been developed by multiple vendors, and their usefulness in improving image quality and reducing acquisition time has been reported.^12,24–26 Therefore, similar results may be obtained using DLR from other vendors. However, because the DLR algorithms vary across vendors, our results may not directly apply to similar algorithms from other vendors, and additional investigations would be required.

This study had a few limitations. First, it included only a few relatively young and healthy male volunteers, excluding pediatric and older participants. This was not an intentional selection but rather reflected the composition of volunteers who responded to recruitment in accordance with the research protocol approved by the Institutional Review Board. This gender imbalance limits the generalizability of our findings to female patients. Future studies including gender-balanced cohorts are warranted to assess potential sex-related differences in trochlear nerve visualization and to validate these findings in diverse patient populations. Second, no pathological evaluations were performed. However, the clinical evaluation was conducted by consensus between 2 experienced neuroradiologists, and the significance of our study results was not compromised. Third, there are several potential sources of reader bias in this study. Qualitative image quality analysis may have been influenced by unconscious reader bias because of the different appearances of each image sequence. Furthermore, the reading order was fixed (3D-FIESTA, followed by T2-CUBE without DLR, then T2-CUBE with DLR). Although the case presentation order within each sequence was randomized and a 10-day washout period was implemented, this fixed reading order may not have completely eliminated potential learning effects or expectation bias. Fourth, although ideally spatial resolution should be matched when comparing image quality, the spatial resolution of 3D-FIESTA was inferior to that of T2-CUBE in this study. This discrepancy was unavoidable because our pilot study demonstrated that higher-resolution 3D-FIESTA induced peripheral nerve stimulation symptoms, precluding its use from a patient safety perspective. Finally, because ARDL is not yet available for 3D-FIESTA, its application to this sequence could not be evaluated.

Conclusion

T2-CUBE with DLR at 3T MRI demonstrated significantly better visualization of the trochlear nerve than 3D-FIESTA and T2-CUBE without DLR, achieving 100% visualization across all segments. This technique has the potential to replace conventional 3D-FIESTA for cisternal imaging and may be extended to other cranial nerves, including the facial, trigeminal, and glossopharyngeal nerves. Further research is required to validate its clinical utility in pathological conditions.