Denoising approach with deep learning-based reconstruction for neuromelanin-sensitive MRI: image quality and diagnostic performance

Purpose Neuromelanin-sensitive MRI (NM-MRI) has proven useful for diagnosing Parkinson’s disease (PD) by showing reduced signals in the substantia nigra (SN) and locus coeruleus (LC), but requires a long scan time. The aim of this study was to assess the image quality and diagnostic performance of NM-MRI with a shortened scan time using a denoising approach with deep learning-based reconstruction (dDLR). Materials and methods We enrolled 22 healthy volunteers, 22 non-PD patients and 22 patients with PD who underwent NM-MRI, and performed manual ROI-based analysis. Signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) in ten healthy volunteers were compared among images with a number of excitations (NEX) of 1 (NEX1), NEX1 images with dDLR (NEX1 + dDLR) and 5-NEX images (NEX5). Acquisition times for NEX1 and NEX5 were 3 min 12 s and 15 min 58 s, respectively. Diagnostic performances using the contrast ratio (CR) of the SN (CR_SN) and LC (CR_LC) and those by visual assessment for differentiating PD from non-PD were also compared between NEX1 and NEX1 + dDLR. Results Image quality analyses revealed that SNRs and CNRs of the SN and LC in NEX1 + dDLR were significantly higher than in NEX1, and comparable to those in NEX5. In diagnostic performance analysis, areas under the receiver operating characteristic curve (AUC) using CR_SN and CR_LC of NEX1 + dDLR were 0.87 and 0.75, respectively, which had no significant difference with those of NEX1. Visual assessment showed improvement of diagnostic performance by applying dDLR. Conclusion Image quality for NEX1 + dDLR was comparable to that of NEX5. dDLR has the potential to reduce scan time of NM-MRI without degrading image quality. Both 1-NEX NM-MRI with and without dDLR showed high AUCs for diagnosing PD by CR. The results of visual assessment suggest advantages of dDLR. Further tuning of dDLR would be expected to provide clinical merits in diagnosing PD. Supplementary Information The online version contains supplementary material available at 10.1007/s11604-023-01452-9.


Introduction
Parkinson's disease (PD) is a neurodegenerative disorder involving progressive loss of dopaminergic neurons in the substantia nigra (SN) and noradrenergic neurons in the locus coeruleus (LC), both of which contain pigments called neuromelanin [1,2].Neuromelanin is a strong chelator of heavy metals, particularly iron, and plays important roles in protecting against neurotoxicity caused by free iron [3,4].Symptoms of PD are thought to appear after 50-60% of dopamine neurons have degenerated, and the presymptomatic phase often spans more than 20 years [5,6].
A denoising approach with deep learning-based reconstruction (dDLR) has been applied for MRI recently [24][25][26][27].The dDLR is trained using vast amounts of high-quality image data, and makes use of a deep learning neural network to remove image noise and produce clear images in clinical practice.We hypothesized that short-scan time NM-MRI of sufficient image quality would be achievable by applying dDLR to NM-MRI with a fewer NEX.To shorten the scan time as much as possible, we used NEX-1 NM-MRI as source images.To the best of our knowledge, no previous studies have examined the image quality or diagnostic accuracy of NM-MRI with dDLR.
The purposes of this study were thus: (1) to compare image quality between NEX-1 NM-MRI without dDLR, NEX-1 NM-MRI with dDLR, and a reference standard of NEX-5 NM-MRI; and (2) to compare the diagnostic capability of NEX-1 NM-MRI with and without dDLR for differentiating patients with PD from non-PD patients.

Study population
This prospective study was approved by the institutional ethics committee.Written informed consent was obtained from both healthy volunteers and patients prior to enrolment.For the image quality study, we prospectively enrolled 22 healthy volunteers.We recruited relatively young volunteers, because we needed them to stay still for about 16 min to acquire images from NEX-5 NM-MRI.For the diagnostic performance study, we enrolled 22 patients with PD who agreed to participate in this study, all of whom fulfilled the Movement Disorder Society PD Criteria for the diagnosis of PD [28], and 22 age-and sex-matched non-PD patients.Underlying diseases in the non-PD patients were brain aneurysm (n = 13), old brain infarction or ischemic change (n = 6) and cerebral artery stenosis (n = 3).All the lesions were located outside of the brainstem and no apparent brainstem abnormalities associated with old brain infarction such as Wallerian degeneration were observed.All participants underwent NM-MRI at our hospital between September 2019 and March 2021.No participants were excluded due to insufficient image quality or large brainstem lesions.Figure 1 summarizes the participant inclusion process and image analysis.

Image acquisition
We acquired images from NM-MRI using a 2-dimensional gradient echo (2D-GRE) pulse sequence with MT contrast (MTC) preparation on a 3-T scanner (Vantage Centurian; Canon Medical Systems, Otawara, Japan) with a 32-channel head coil.1-NEX NM-MRI was acquired from all healthy volunteers, and 5-NEX NM-MRI was acquired from 10 out of 22 volunteers (Fig. 1).For patients with PD, only 1-NEX NM-MRI was performed in this study.Brain MRI for screening had been finished on another day, and no specific abnormalities were found.For non-PD patients, 1

Post-imaging procedure
Denoising was applied to 1-NEX images (NEX1) using a commercially available deep learning-based reconstruction algorithm (Advanced intelligent Clear-IQ Engine [AiCE]; Canon Medical Systems) to create NEX1 + dDLR.AiCE is a whole reconstruction pipeline from raw complex data to final image generation.The complex images are processed for denoising in this pipeline which incorporates convolutional neural network.Its details were described in the previous literature [24].Online Resource 1 displays its architecture.CNN architecture consists of multiple layers: the feature extraction, feature conversion and image generation layers.In the feature extraction layer, the input noisy image is convolved by the 7 × 7 discrete cosine transformation to derive 49 components, which is divided into 48 highfrequency components and a zero-frequency component.A soft-shrinkage activation function is applied to 48 high-frequency components.Next, the 48 high-frequency components undergo repeated 3 × 3 convolution and soft shrinkage in the feature conversion layers.Finally, in the image generation layer, the denoised output image is generated by the 7 × 7 inverse discrete cosine transform convolution of both the output data from the feature conversion layers and the bypassed zero-frequency component.The soft-shrinkage activation function enables adaptive noise removal using a threshold calculated from the noise level and a coefficient.Thus, there are two parameters to be trained: the 3 × 3 convolution kernels in the feature conversion layer and the coefficient of the soft-shrinkage activation function in the feature extraction and feature conversion layers.These parameters have been determined to minimize the differences between the training data and the output denoised image through the training process of AiCE.

Image analysis
All images were analyzed using ImageJ software (National Institutes of Health) as the consensus decisions of 2 boardcertified radiologists (S.O. and Y.F. with 9 and 22 years of experience in neuroradiology, respectively).Regions of interest (ROIs) of the SN, decussation of superior cerebellar peduncle (SCP), LC and pons were manually placed on the slice where the SN or LC was most clearly delineated (Online Resource 2).As for the ROIs of the SN, three circles were placed for right and left, respectively, and the signal intensities of the six ROIs were averaged [29].The SCP and pons were used as background areas for the SN and LC, respectively.The shape and size of ROIs were the same in all images.
1. Image quality Image quality was assessed quantitatively and qualitatively using images from 10 healthy volunteers and from 22 patients with PD.
For quantitative assessment, we calculated SNR of the SCP (SNR_SCP), SNR of the pons (SNR_pons), contrast-to-noise ratio (CNR) between SN and background SCP (CNR_SN) and CNR between LC and background pons (CNR_LC).We measured signal intensity (SI) in each ROI (SI SCP , SI pons , SI SN , and SI LC ).SNR and CNR were defined as follows [30]: SNR_SCP = mean SI SCP / SD of SI SCP , SNR_pons = mean SI pons / SD of SI pons , CNR_SN = (mean SI SN − mean SI SCP ) / SD of SI SCP and CNR_LC = (mean SI LC − mean SI pons ) / SD of SI pons , where SD is the standard deviation.
For qualitative assessment of image quality, three neuroradiologists (S.N., S.O. and S.O. with 15, 13 and 10 years of experience in neuroradiology, respectively) visually evaluated overall image quality, artifacts, structural conspicuity and noise of the images at the SN and LC level by consensus using a 5-point Likert scale.The criteria for image assessment on the 5-point Likert scale are presented in Online Resource 3.

Diagnostic performance by contrast ratio
To assess diagnostic performance, we calculated contrast ratios (CRs) of the SN and LC using images from 22 non-PD patients, 22 patients with PD and 22 healthy volunteers.CRs were defined as follows: CR_SN = mean SI SN / mean SI SCP and CR_LC = mean SI LC / mean SI pons 3. Diagnostic performance by visual assessment Three neuroradiologists (S.N., S.O. and S.O. with 15, 13 and 10 years of experience in neuroradiology, respectively) visually assessed NEX1 and NEX1 + dDLR images at the level of the SN and LC, respectively, to differentiate PD from non-PD.Raters selected "PD", "non-PD" and "difficult to diagnose" in accordance with the following criteria.As for SN, raters focused on whether the lateral part of SN is conspicuous or obscure.As for LC, raters focused on whether bilateral high intensities suggesting LC are well defined or not.If it was difficult to determine, "difficult to diagnose" was selected.

Statistical analysis
For quantitative image quality analysis, SNRs and CNRs were compared among NEX1, NEX1 + dDLR and 5-NEX images without dDLR (NEX5) using analysis of variance (ANOVA) with Bonferroni correction for healthy volunteers, and among NEX1 and NEX1 + dDLR using paired t-test for patients with PD.For qualitative analysis, each qualitative index for the three types of images was compared using the Wilcoxon signed-rank test.
For diagnostic performance analysis by contrast ratio, we evaluated difference in CR_SN and CR_LC between ageand sex-matched non-PD and PD groups using Student's t-test.We then performed receiver operating characteristic curve analyses for differentiating patients with PD from non-PD patients and compared the area under the receiver operating characteristic curve (AUC) between NEX1 and NEX1 + dDLR images using the DeLong test.In addition, differences in CR_SN and CR_LC between healthy and PD groups were assessed and receiver operating characteristic curve analyses were performed.
For diagnostic performance analysis by visual assessment, accuracy was calculated by dividing the number of cases with correct diagnosis by the total number of cases.
MedCalc version 20.009 software (MedCalc Software, Ostend Belgium) was used for statistical analyses, with differences of p < 0.05 considered significant.

Results
The characteristics of participants are listed in Table 1

Image quality
Representative images of the SN and LC from a 30-year-old healthy female volunteer are shown in Fig. 2. NEX1 + dDLR and NEX5 images visualize the SN and LC more clearly than NEX1 images.Results for SNR and CNR of healthy volunteers are presented in Fig. 3 and Table 2. P values are shown in Online Resource 4. SNR and CNR were significantly higher for NEX1 + dDLR than for NEX1 (p < 0.001) at the SN and LC.SNR and CNR from NEX1 + dDLR did not Fig. 3 Box-and-whisker plots and scatter plots for SNR and CNR from NEX1, NEX1 + dDLR and NEX5 images of healthy volunteers.SNR and CNR from NEX1 + dDLR were significantly higher than those from NEX1 and showed no significant difference from those of NEX5.SNR signal-to-noise ratio; CNR contrast-to-noise ratio, NEX number of excitations, dDLR denoising approach with deep learningbased reconstruction show any significant difference from those of NEX5.For patients with PD, SNR and CNR at the SN and LC were significantly higher for NEX1 + dDLR than for NEX1 (p < 0.001) (Table 2).
The results of qualitative assessment are shown in Online Resource 5. Scores for overall image quality, structural conspicuity and noise were significantly better for NEX5 among the three types of images (p < 0.01), and those for NEX1 + dDLR were significantly better than those of NEX1 (p < 0.001) for both the SN and LC.No significant differences in scores for artifacts were seen among the three images for both the SN and LC.

Diagnostic performance by visual assessment
The results of diagnosis by visual assessment are shown in Table 4. Accuracy of each rater for NEX1 and NEX1 + dDLR was 0.45-0.59and 0.59-0.64 for the SN, while 0.32-0.39 and 0.39-0.41for the LC, respectively.

Discussion
In this study, we applied dDLR to NM-MRI with an NEX of 1, which offers a much shorter scan time than conventional NM-MRI, and examined the resulting image quality and diagnostic performance.Image quality analyses showed approximately 1.5-fold improvement in SNR and CNR by applying dDLR to NEX1 images.The diagnostic performance using CR_SN and CR_LC in NEX1 + dDLR NEX number of excitations, dDLR denoising approach with deep learning-based reconstruction, SNR_SCP signal-to-noise ratio of the decussation of superior cerebellar peduncle, SNR_pons signal-to-noise ratio of the pons, CNR_SN contrast-to-noise ratio between the substantia nigra and decussation of superior cerebellar peduncle, CNR_LC contrast-to-noise ratio between the locus coeruleus and pons

Healthy volunteers
Patients   was comparable to that of NEX1.These results may suggest that NEX1 images are sufficient to differentiate between PD and non-PD patients and no apparent benefit for diagnostic performance by dDLR.However, diagnosis by visual assessment showed slight improvement of diagnostic performance by applying dDLR, which suggests the advantages of dDLR in diagnostic performance.Further tuning of dDLR would be expected to provide clinical merits in diagnosing PD.
Our study showed a lower diagnostic capability of the LC for PD compared with that of the SN, which was consistent with previous studies [9,31,32].The lower diagnostic performance of the LC than SN and the lower diagnostic performance of NEX1 + dDLR than NEX1 for the LC in our results may be because the LC is so small a structure that limited resolution of MR imaging can make it difficult to quantify the signal intensity accurately.Also, a previous study demonstrated that the optimal flip angle for LC imaging is different from that for SN [33].Optimization of flip angle to increase contrast of LC may be required for stable measurement of LC and for taking full advantage of dDLR for LC images.Several articles have already applied deep learning methods to MRI for diagnosing PD, such as for the creation of diagnostic biomarkers for PD [34], automatic segmentation of the SN on NM-MRI [35][36][37], and interpretation of nigrosome 1 on susceptibility map-weighted imaging [38].However, to the best of our knowledge, no previous studies have applied deep learning-based denoising methods to NM-MRI.Our study demonstrated the utility of using dDLR for NM-MRI to reduce examination times without degrading image quality.Acquisition times for NEX5 and NEX1 are 15 min 58 s and 3 min 12 s, respectively; so, using NEX1 + dDLR instead of NEX5 would achieve a time reduction of around 12 min.This is quite advantageous, particularly for evaluating patients with PD who have tremors or involuntary movements.Furthermore, there is a possibility that denoising by dDLR can be beneficial for diagnosis by visual assessment.Considering that there have not been established criteria for visual diagnosis of PD by NM-MRI, further studies are required in the future.
Several limitations to this study should be considered.First, the number of participants was relatively small.Second, diagnostic performance using NEX5 images was not evaluated.This was because the scan time for NEX5 (15 min 58 s) was too long and uncomfortable for patients with PD, who suffer from tremors or involuntary movements with tolerate.Although we did not compare diagnostic performance between NEX1 + dDLR and NEX5, NEX1 + dDLR showed sufficiently high AUCs (0.87 for CR_SN, 0.75 for CR_LC).Third, age-and sex-matched healthy volunteers were not enrolled, because we recruited relatively young volunteers who could stay still during the scan of about 16 min.Diagnostic performance in this study was therefore evaluated between patients with PD and age-and sex-matched non-PD patients.Fourth, we used 2D NM-MRI in this study and we did not perform voxel-wise analysis as in previous papers [13].Further investigations should be performed to assess the usefulness of dDLR in the application to 3D NM-MRI.Fifth, various 2D and 3D image sequences including turbo spin echo and GRE have been used for NM-MRI other than the 2D-GRE sequence we used.A future comprehensive study of NM-MRI using these sequences for healthy volunteers and patients with PD is required to determine the most appropriate NM-MRI.Lastly, our study used a vendorsupplied DLR algorithm to assess its clinical feasibility at only one institution.A multicenter study with various MRI scanners is therefore required.
In conclusion, 1-NEX NM-MRI with dDLR provided comparable image quality to 5-NEX NM-MRI, which represented the reference standard in this study.Our study demonstrated the potential of dDLR to reduce scan time of NM-MRI without degrading image quality.The diagnostic performance of 1-NEX NM-MRI using contrast ratio of the SN and that of LC was sufficiently good enough that dDLR did not further improve diagnostic accuracy in this study.However, the results of diagnosis by visual assessment suggest advantages of dDLR.Further tuning of dDLR would be expected not only to reduce scan time of NM-MRI without degrading image quality, but to provide clinical merits in diagnosing PD.

Fig. 1
Fig. 1 Flowchart of study participants and image analysis.PD Parkinson's disease, NEX number of excitations, dDLR denoising approach with deep learning-based reconstruction

Fig. 2
Fig. 2 Images of the SN (arrowheads in the upper row) and LC (arrows in the lower row) from NEX1, NEX1 + dDLR and NEX5 for a 30-year-old healthy female participant.NEX, number of excitations; dDLR, denoising approach with deep learning-based reconstruction

Fig. 4 Fig. 5
Fig. 4 Images of the SN (arrowheads in the upper row) and LC (arrows in the lower row) from NEX1 + dDLR for a 73-year-old female non-PD patient (left column) and a 54-year-old female patient with PD (right column).PD Parkinson's disease

Table 4
Diagnostic performance by visual assessment of NEX1 and NEX1 + dDLR for differentiation between healthy volunteers and patients with PD SN substantia nigra,

Table 1
Characteristics of volunteers and patients *Mean age ± standard deviation (years), with range shown in parentheses **Healthy volunteers vs patients with PD ***Non-PD patients vs patients with PD PD Parkinson's disease, HY Hoehn and Yahr, UPDRS Unified Parkinson's disease rating scale

Table 3
NEX number of excitations, dDLR denoising approach with deep learning-based reconstruction, PD Parkinson's disease, AUC area under the curve, CR_SN contrast ratio of the substantia nigra, CR_LC contrast ratio of the locus coeruleus