Evaluation of Software-Optimized Protocols for Acoustic Noise Reduction During Brain MRI at 7 Tesla

1 Introduction

During an MRI exam, both acoustic noise and potential tissue stimulation are unwanted, yet inherent, side effects of utilizing the gradient coils [1]. During scanning, the interaction between the alternating gradient coil currents and the static magnetic field (B ₀) generates vibrational forces on the gradient-coil conductors, causing loud hammering sounds. Routine scanning typically generates sound pressure levels (SPLs) that exceed the threshold for possible hearing impairment, 85 dB (dB), and some pulse sequences even reach 110–130 dB [1-4]. Hearing protection is therefore required to minimize the risk of hearing damage [1, 2].

MR-generated acoustic noise contributes to patient discomfort and anxiety [5, 6], decreased exam success rates [7, 8], and impedes patient-radiographer communication [9, 10]. It has also been associated with stress and health-related concerns among MR personnel [10, 11]. In addition, the swiftly alternating gradient fields during scanning induce currents that may stimulate conductive tissues in the exposed person, leading to peripheral nerve stimulation (PNS). PNS can manifest as uncontrolled muscle contractions, twitching, and tingling sensations in the limbs or torso, causing discomfort or even pain [1, 12, 13].

Decreasing the gradient activity can reduce both the acoustic noise and the gradient field exposure that induces PNS [14]. In practice, this reduction can be achieved through software-optimized gradient-restricting pulse sequences, which are typically provided by MR vendors as acoustic noise reduction (ANR) sequences—for example, “Whisper Mode” or “SofTone” [15, 16]. These techniques aim to smooth steep gradient current alterations, for example, by increasing the gradient ramp time or reducing its maximum amplitude (i.e., reducing the gradient slew rate), thereby decreasing the mechanical output of the gradient coil. However, these changes can restrict sequence parameters such as echo time (TE) and repetition time (TR), potentially reducing the image quality or extending the scan time [17, 18]. As ANR methods often require comprehensive user optimization [17, 18], they need to be validated and compared to conventional gradient output sequences before they are implemented for routine imaging [19].

In a previous study [14] we evaluated two ANR software solutions, which significantly decreased peak SPLs by up to 84% and gradient field exposure by 48%–66%, without reducing image quality. That study was performed using a 1.5-Tesla (T) system. However, the demand for higher field MR systems is continuously increasing; a higher B₀ provides an increased signal-to-noise ratio (SNR) and may enable improved image quality and diagnostic utility [20, 21]. These MR systems are typically utilized for high spatial or temporal resolution, which generally pushes the gradient system further and can cause higher vibration amplitudes and louder scans [22]. Therefore, hearing safety may be of particular concern with ultra-high-field MR systems, that is, ≥ 7-T [4]. For instance, T₂-weighted turbo spin echo (TSE) and T₁-weighted magnetization-prepared rapid gradient echo (MPRAGE) were recently studied in a 7-T system, with reported peaks at 121.6 and 120.4 dB, respectively [4]. In another study, 63% of participants reported PNS to some degree during 7-T MRI [23].

The aim of this study was to evaluate image quality, SPLs, and perceived noise levels when using the ANR technique SofTone for T₂-weighted fast spin echo (T₂W FSE) and three-dimensional T₁-weighted turbo field echo (3D T₁W TFE), and to compare this to conventional imaging during 7-T brain MRI.

2 Materials and Methods

This prospective single-center study was approved by the Swedish Ethical Review Authority (DNR 2022-05816-01). Participation was voluntary, with written informed consent obtained, and could be withdrawn at any time. We employed a within-subject design to compare software-optimized ANR sequences for brain MRI with conventional counterparts. This study is part of a larger research project that aims to improve our understanding of software-based ANR in ultra-high-field (7-T) MRI, including prospective data on induced PNS, muscle activity, and other stress responses in the sympathetic nervous system. Evaluations of the non-auditory aspects will be reported in future publications.

All scans took place using a 7-T system (Achieva, Philips, Best, the Netherlands). Scanning was performed in first-level controlled operating mode, with the specific absorption rate (SAR) limited to 4 W/kg for whole-body or 3.2 W/kg for head. Each participant was examined head-first in a supine position, using a 2 transmit (Tx)/32 receive (Rx) head coil (Nova Medical, Wilmington, MA, USA). To homogenize and increase the signal intensity in the cerebellum and mid-brain area, we positioned dielectric pads between the head coil and the occipitotemporal region of each participant's head [24, 25].

Our evaluation focused on T₂W FSE and 3D T₁W TFE, as both are commonly used for morphological imaging in brain protocols at the study center. One conventional (non-ANR) scan and one SofTone-optimized scan were conducted for both. SofTone can be applied in five steps (factors), where each increment places additional restrictions on the gradients, achieving further noise reduction [15]. During optimization, we adjusted the parameters as little as possible to ensure that image contrast and scanning time remained similar to the conventional sequences. The SofTone sequences required a minor increase in the minimum allowed TE, for which we applied the shortest possible settings. Additionally, we increased the receiver bandwidth, resulting in a slight reduction in SNR. The echo spacing increased by 1.8 ms, from 12.0 to 13.8 ms, using SofTone on T₂W FSE. Other parameters remained the same for the conventional and SofTone pairs (Table 1).

The protocol also included a survey scan (time; min:sec; 00:57), a coil reference scan (00:11), a SENSitivity Encoding (SENSE) reference scan (00:38), and a B ₀ shimming (00:18). In addition, the participants were examined using both conventional (05:42) and SofTone (05:42) diffusion tensor imaging (DTI); however, that data will be presented in a separate publication. The order of the T₂W FSE, 3D T₁W TFE, and DTI sequences was randomized for each participant, and participants were blinded to the specific sequences being used to mitigate bias.

Participants were healthy volunteers recruited through advertisements in the study center, hospital, and university area, and via local contacts in the researchers' networks. The exclusion criteria were non-MRI-compatible implants, being < 18 years old, pregnancy, and not meeting the clinic's MRI safety criteria. Additionally, participants were required to have functional hearing to be able to evaluate perceived noise levels and to enable verbal communication via the intercom system. Since the head coil was too narrow to fit headphones, the participants were required to use two layers of hearing protection: disposable earplugs and a sound-attenuating impression paste applied to their auricles (see Figure 5.1 in [26]). All morphological MR images were interpreted by a radiologist to ensure no incidental pathologies were missed and, where necessary, reported for further follow-up. Data were collected between March and April 2023.

3 Results

Our study included 28 healthy volunteers. We scanned an additional participant who could not complete the examination due to uncomfortable sensations at the site of a metallic orthopedic implant. For safety reasons, we stopped the examination and excluded this participant, and recruited another participant as a replacement. Among the final cohort of 28 participants, 20 identified as women and eight as men (age range 18–56; mean ± standard deviation, SD: 33 ± 11 years). The images of two participants exhibited motion artifacts throughout their examinations that noticeably degraded the image quality of at least one sequence among the same-weighted conventional and SofTone pairs. To ensure a consistent basis for the evaluation of image quality, we excluded these two participants from all image-quality assessments, resulting in n = 26 for those evaluations. However, we included these two participants in the SPL and Borg CR10 evaluations (n = 28).

3.1 Sound Pressure Level

Significant differences were observed in the peak SPLs between SofTone and conventional scanning for both T₂W FSE and 3D T₁W TFE. The recorded peak SPL during T₂W FSE conventional scanning was 116.3 ± 1.8 dBA (mean ± SD), and 97.0 ± 2.1 dBA during T₂W FSE SofTone (significant). This represents a mean difference of 19.3 dBA, corresponding to an 89% reduction in sound pressure with the application of SofTone. For 3D T₁W TFE, the mean peak SPL was 123.7 ± 1.2 dBA for conventional scanning and 101.5 ± 1.2 dBA for SofTone scanning (significant). This difference of 22.2 dBA translates to a 92% reduction in sound pressure. The SPL distributions are presented in Figure 1.

3.3 Image Quality

For T₂W FSE, no significant differences were observed in general image quality, ability to differentiate between white matter and gray matter, and ability to differentiate CSF from adjacent soft tissues (p ranging from 0.21 to 1.00, see Table 2). Interestingly, Reader 2 rated SofTone as having significantly fewer interfering artifacts than the conventional sequence, although the mean score only differed by 0.16. Reader 1 and Reader 3 did not observe any significant differences in artifacts between SofTone and conventional scanning (Reader 1: p = 1.00; Reader 3: p = 0.74).

TABLE 2. Image-quality assessments for T₂W FSE (n = 26).

	Sequence		p
	T2W FSE Conventional	T2W FSE SofTone
Reader 1
Artifacts	3.12 ± 0.43	3.12 ± 0.52	1.00
General image quality	3.42 ± 0.58	3.35 ± 0.63	0.64
White matter from gray matter	3.50 ± 0.58	3.50 ± 0.65	0.98
CSF from soft tissues	3.96 ± 0.20	3.96 ± 0.20	1.00
Reader 2
Artifacts	2.88 ± 0.43	3.04 ± 0.45	< 0.05
General image quality	2.92 ± 0.69	3.00 ± 0.49	0.59
White matter from gray matter	3.50 ± 0.58	3.35 ± 0.63	0.32
CSF from soft tissues	3.85 ± 0.46	3.88 ± 0.33	0.74
Reader 3
Artifacts	2.77 ± 0.43	2.73 ± 0.45	0.74
General image quality	2.69 ± 0.55	2.58 ± 0.50	0.37
White matter from gray matter	2.69 ± 0.55	2.54 ± 0.51	0.21
CSF from soft tissues	2.92 ± 0.27	2.92 ± 0.27	1.00

• Note: The distribution of image assessment scores (range: 1.00–4.00) shown with mean ± standard deviation. Bold values indicate statistical significance.
• Abbreviations: CSF, cerebrospinal fluid; T₂W FSE, T₂-weighted fast spin echo.

The image quality for 3D T₁W TFE exhibited a slightly higher subjective variance among the three radiologists (Table 3). While no significant differences were observed in terms of artifacts or ability to differentiate CSF from soft tissues (p ranging from 0.05 to 0.32), Reader 1 rated the general image quality of SofTone (mean ± SD: 3.08 ± 0.63) as significantly inferior to that of conventional scanning (mean ± SD: 3.65 ± 0.56). Neither of the other two readers observed any significant differences in general image quality (Reader 2: p = 0.13; Reader 3: p = 0.08). However, Readers 2 and 3 found the differentiation between white and gray matter to be significantly worse in 3D T₁W TFE SofTone (mean ± SD, Reader 2: 3.19 ± 0.63; Reader 3: 2.27 ± 0.45) than in conventional scanning (mean ± SD, Reader 2: 3.54 ± 0.58; Reader 3: 2.81 ± 0.40), whereas Reader 1 did not (p = 0.06). An example image is shown in Figure 2.

TABLE 3. Image-quality assessments for 3D T₁W TFE (n = 26).

	Sequence		p
	3D T1W TFE Conventional	3D T1W TFE SofTone
Reader 1
Artifacts	2.92 ± 0.39	2.77 ± 0.43	0.16
General image quality	3.65 ± 0.56	3.08 ± 0.63	< 0.01
White matter from gray matter	3.85 ± 0.37	3.62 ± 0.57	0.06
CSF from soft tissues	4.00 ± 0.00	3.96 ± 0.20	0.32
Reader 2
Artifacts	3.00 ± 0.28	2.88 ± 0.43	0.18
General image quality	3.04 ± 0.34	2.85 ± 0.54	0.13
White matter from gray matter	3.54 ± 0.58	3.19 ± 0.63	0.02
CSF from soft tissues	3.77 ± 0.43	3.50 ± 0.51	0.05
Reader 3
Artifacts	2.88 ± 0.33	2.69 ± 0.55	0.13
General image quality	2.73 ± 0.45	2.50 ± 0.51	0.08
White matter from gray matter	2.81 ± 0.40	2.27 ± 0.45	< 0.001
CSF from soft tissues	3.00 ± 0.00	2.88 ± 0.33	0.08

• Note: The distribution of image assessment scores (range: 1.00–4.00) shown with mean ± standard deviation. Bold values indicate statistical significance.
• Abbreviations: CSF, cerebrospinal fluid; 3D T₁W TFE, three-dimensional T₁-weighted turbo field echo.

The interrater agreement and reliability scores are shown in Table 4. Overall, the agreement between the three readers was low: no percent agreement exceeded 77%, with the highest α being −0.3753, corresponding to a fair systematic disagreement in assessment.

TABLE 4. Reliability and agreement between observers (n = 26).

Image assessment	Sequence
	T2W FSE Conventional	T2W FSE SofTone	3D T1W TFE Conventional	3D T1W TFE SofTone
Artifacts	α = 0.3084	α = 0.2719	α = −0.0268	α = 0.2745
	76%	64%	77%	71%
	(20/26; 18/26; 21/26)	(16/26; 16/26; 18/26)	(20/26; 19/26; 21/26)	(17/26; 19/26; 19/26)
General image quality	α = 0.2604	α = 0.0453	α = 0.0027	α = 0.2093
	53%	46%	40%	49%
	(15/26; 8/26; 18/26)	(12/26; 8/26; 15/26)	(8/26; 4/26; 18/26)	(10/26; 10/26; 18/26)
White matter from gray matter	α = 0.1473	α = −0.1223	α = −0.1477	α = −0.1515
	40%	33%	32%	29%
	(19/26; 7/26; 5/26)	(10/26; 7/26; 8/26)	(15/26; 2/26; 8/26)	(16/26; 3/26; 3/26)
CSF from soft tissues	α = −0.2446	α = −0.3753	α = −0.3601	α = −0.3086
	33%	32%	33%	31%
	(23/26; 0/26; 2/26)	(22/26; 1/26; 2/26)	(20/26; 0/26; 6/26)	(12/26; 1/26; 11/26)

• Note: The interrater agreement between qualitative assessments made by the three radiologists: Reader 1 (R1), Reader 2 (R2), and Reader 3 (R3). Data shown in top row, Krippendorff's alpha (α); in middle row, percent agreement (%); in bottom row, total number of agreements out of a maximum n = 26 between R1/R2; R1/R3; R2/R3.
• Abbreviations: CSF, cerebrospinal fluid; 3D T₁W TFE, three-dimensional T₁-weighted turbo field echo; T₂W FSE, T₂-weighted fast spin echo.

Artifacts were identified in all images; these were typically related to 7-T-specific characteristics such as B ₁ inhomogeneity, susceptibility artifacts (e.g., sinus nasale), or Gibbs' artifacts [25]. These artifacts, while noticeable, were not necessarily deemed to hinder the overall diagnostic utility of the morphological information. However, they did affect certain areas of the image, reducing the general image quality assessment scores given by all three readers, as well as the differentiation of CSF from soft tissues. Additionally, some assessments revealed the presence of motion, chemical shift, and/or flow artifacts. Figure 3 illustrates an example where increased signal loss, motion, and image blurring were noticed on 3D T₁W TFE SofTone compared to the conventional scanning, along with a decreased ability to differentiate between white and gray matter.

4 Discussion

The use of SofTone in 7-T brain MRI was found to significantly reduce both peak SPLs and perceived noise levels while maintaining acceptable subjective image quality.

No significant differences in image quality were observed between conventional and SofTone T₂W FSE, although one of the radiologists noted slightly less artifact interference with SofTone. This minor improvement could possibly be attributed to the increased receiver bandwidth, potentially reducing chemical shift and susceptibility artifacts [33, 34]. Several studies investigating various ANR software and B₀ field strengths have shown preserved image quality with T₂W FSE/TSE [14, 27, 35-38]. Some of these have evaluated algorithmic software that optimizes and smooths gradient forms, which could offer even quieter exams [14, 37]. Since not all MR systems have access to such software, it is useful to know that plain slew rate restriction is still effective in 7-T MRI.

Despite greater variance in the assessed image quality for 3D T₁W TFE, as evidenced by the two radiologists who found SofTone to provide worse white and gray matter differentiation, their ratings remained within an acceptable diagnostic range. However, the loss of contrast between white and gray matter may have been caused by the slightly longer TE, allowing more T₂*-relaxation and, thus, less T₁-contrast. Jacobs et al. [39] also found that the contrast-to-noise ratio (CNR) between gray and white matter decreased when using an ANR-optimized T₁W MPRAGE combined with a silent gradient coil. Similarly, software-based ANR fluid-attenuated inversion recovery (FLAIR) resulted in decreased CNR at 7 T [40]. The assessment of general image quality significantly differed by one radiologist in our study, where the mean score for 3D T₁W TFE SofTone was closer to “good” compared to “excellent” for the conventional sequence. While the quality is still acceptable for diagnostics, the results suggest uncertainty regarding how software-optimized ANR potentially influences subjective preferences.

Another effective ANR software approach for 3D T₁W gradient echo sequences, including TFE and MPRAGE, is the “zero echo time” (ZTE) technique [41]. ZTE utilizes sub-millisecond TEs and continuously applied gradients [42] to achieve near-ambient noise levels, and reduced SPLs by 27–29 dB as compared to conventional T₁W MPRAGE [43, 44]. However, ZTE may limit user customization, for example, requiring a cubic field of view or disabling acceleration factors [41], and it inherently struggles to provide high image quality with T₂W or diffusion weighting [42]. Notably, ZTE is not an option on all systems, including the one used in this study, which complicates direct comparisons. Nonetheless, there is a clear contrast between our study and others in terms of the sound level of the conventional sequences: our T₁W TFE peaked at 123.7 dBA, which was considerably louder than the 78 and 87 dB recorded for conventional T₁W MPRAGE in previous ZTE studies at 3-T [43, 44]. Despite SofTone reducing SPLs by up to 22.2 dBA in our study, the scanner's noise level still necessitates disposable hearing protection, as the system's head coil is too small for headphones [24]. However, since every 3 dB decrease corresponds to a halving of sound intensity [45], the SPL reduction achieved with SofTone here effectively extends permissible scan time nearly sevenfold, making it a very useful tool for hearing protection during lengthy 7-T exams.

The average peak SPL reduction of 89%–92% aligns with our participants' scores for noise perception, rating SofTone as quieter than conventional scanning. This noise reduction not only decreases the risk of hearing damage, but may enhance patient comfort [35, 36, 46]. Additionally, patients have previously experienced less discomfort and an improved ability to hear music during MRI at 1.5-T and 3-T with ANR sequences [47]. In another study, patients reported the overall experience of undergoing 7-T MRI as significantly less comfortable than healthy volunteers [23] and suggested improvements such as receiving information prior to particularly loud sequences and better hearing protection. Consequently, the Borg CR10 scores reported by the healthy volunteers in our study might underestimate the discomfort experienced in clinical contexts (i.e., by patients). ANR sequences could be especially helpful for vulnerable patient groups, such as children, claustrophobic and non-cooperative patients, those under sedation, and patients with tinnitus or other auditory sensitivities [7, 33, 46, 48].

Technical details regarding SofTone are sparse in the literature. However, our practical experience indicates that some sequence parameters (e.g., TE, TR, echo spacing, and receiver bandwidth) increase when applying the software, although the extent of this depends on the SofTone factor. Our study used SofTone factors 3 and 4, suggesting that a lower factor could still significantly reduce the SPL—albeit slightly louder than our current results show—but with even less impact on image quality. An updated implementation of SofTone is available on the vendor's MR platform, named ComforTone, in which optimization of ANR settings is automated [15]. Still, there are challenges that need to be addressed [33].

The variety of ANR software available for commercial MR systems offers alternatives, but the potential compromises in image quality and prolonged scanning time for a specific sequence introduce uncertainty. Future technological advancements, including artificial intelligence (AI), can hopefully address these issues. Notably, Swedish MR radiographers desired more user-friendly solutions to increase the clinical adoption of software-based ANR [10]. Despite its promising utility, current ANR software may influence contrast-determining parameter changes, and we recommend testing ANR on your local system before clinical implementation to ensure its viability. Given earplugs do not always fit and a second layer of hearing protection can only account for up to 5 dBA of extra dampening [45], reducing SPLs by 19–22 dBA through ANR software is a powerful way of minimizing the risk of hearing damage and increasing patient comfort.