Усовершенствованная оценка юкстакортичных лезионов в множестве склерозов с использованием высокоускоренного метода двойного инверсионного восстановления с высоким разрешением, MR Imaging with Deep Learning by Leveled Restruction.

Introduction

Multiple sclerosis (MS) is a chronic inflammatory demyelinating disease of the central nervous system (CNS) that primarily affects young adults and leads to progressive neurological decline. MRI is essential for diagnosing MS, understanding its pathophysiology, and assessing treatment response. Histopathological studies revealed that a substantial portion of cerebral lesions in patients with MS is found in the gray matter (GM) or at the interface between the cortical and subcortical white matter (WM). 1 – 3 This recognition of GM involvement is integral to current MS diagnostic criteria, highlighting the importance of detecting juxtacortical lesions. 4 , 5 Despite the known association between GM lesion load and neurological impairment, conventional MRI sequences, such as T2-weighted imaging (T2WI), fluid-attenuated inversion recovery (FLAIR), and T1-weighted imaging (T1WI), both with and without gadolinium-based contrast enhancement, are often inadequate for the detection of GM pathology.

3D double inversion recovery (DIR) imaging, which suppresses both WM and cerebrospinal fluid (CSF) signals, offers enhanced sensitivity for detecting juxtacortical MS lesions compared to conventional MRI sequences. 6 – 8 In a previous study on the detection of juxtacortical lesions in MS using 3D DIR, Wattjes et al. found that DIR had higher sensitivity for mixed gray matter-white matter lesions compared to T2WI or FLAIR. 9 In a postmortem MRI study, Seewann et al. reported that 1.6 times more cortical lesions were detected by using DIR instead of FLAIR, with a specificity of 90%. 10 These findings consistently highlight the utility of DIR for the detection of juxtacortical lesions in MS. However, 3D DIR typically requires extended TRs and prolonged scan durations, which makes it difficult to achieve high spatial resolution and optimal SNR within a clinically practical timeframe. 11 Geurts et al. reported improved detection of intracortical lesions using 1.8 mm slice thickness 3D DIR (acquisition time: 15.3 mins). 12 Previous studies have utilized 3D DIR imaging with voxel dimensions of 1.2 × 1.2 × 1.3 mm 3 (acquisition time: 6 mins) 13 and 1.0 × 1.0 × 3.0 mm 3 (acquisition time: 3.5 mins). 11 These studies highlight the need for an innovative, rapid imaging technique that balances high spatial resolution with adequate image quality.

The challenge in achieving high spatial resolution MRI lies in the trade-off between SNR, resolution, and scan time. Recent studies have shown that deep learning (DL)-based denoising methods allows for effective noise reduction while preserving structural details and avoiding image blurring. 14 – 17 More recently, a novel DL-based technique for reconstructing highly undersampled MRI data (DL-Speed, DLS) has been developed, which demonstrated superior performance compared to compressed sensing. 18 DLS employs an unrolled iterative neural network architecture, incorporating a data fidelity module to ensure consistency between the reconstructed image and the acquired k-space data, along with dense skip connections across unrolled modules to enhance reconstruction accuracy. DLS-DIR uses a variable-density Poisson disc (VDPD) undersampling strategy to accelerate DIR scans and a DL-based reconstruction algorithm to process these undersampled k-space data into high-quality DIR images. The model, consisting of 28 iterations and trained on multichannel complex data, was integrated into the standard reconstruction pipeline to generate DICOM images for clinical evaluation. A previous study on healthy individuals 18 demonstrated that DLS can achieve up to a 10-fold acceleration in 3D MRI scans, with 2- to 5-fold acceleration compared to conventional parallel imaging, while preserving diagnostic image quality. However, the usefulness of DLS has not yet been evaluated in clinical studies. In the present study, we performed whole-brain 3D DIR imaging using DLS with a high spatial resolution of 0.7 mm isotropic voxels with an acquisition time of 4 mins 23s. This study aimed to evaluate the diagnostic efficacy of DLS-DIR compared to 3D FLAIR and conventional DIR (C-DIR) in the detection of juxtacortical MS lesions.

Materials and Methods

This retrospective observational study was carried out in accordance with the principles embodied in the Declaration of Helsinki and approved by the Institutional Review Board of Hirosaki University (approval no. 2022-052), and the requirement for informed consent was waived by the board.

MRI protocol

All MRI scans were performed using a 3T MRI system (48-channel Head Coil, Signa Premier 3.0T; GE Healthcare, Milwaukee, WI, USA). The acquisition parameters for the sequences used are detailed in Table 1 .

Image quality evaluation in healthy participants

We included 5 participants (4 men and 1 woman; median age, 38 years; age range, 16–54 years) with normal neurological examinations, no history of neurological disorders or malignancies, and normal brain MRI findings. All participants underwent brain MRI, including 1 mm T2WI, T1WI, 3D FLAIR, DWI, and C-DIR, in addition to DLS-DIR. Two neuroradiologists (T.S. and Sh.K.) evaluated the quality of DLS-DIR images and whether it could potentially obscure the detection of juxtacortical lesions. The neuroradiologists documented the location and characteristics of any findings including artifacts identified in previous studies. 10 , 19 , 20 Common artifacts (false positives) identified in previous studies on DIR, juxtacortical high signal intensities caused by cortical vessels or fluid (CSF) flow within the sulci in combination with slice profile inhomogeneities. When further information to evaluate suspected artifacts was required, the neuroradiologists examined the coronal or sagittal view of the appropriate cross-section images.

MS lesion evaluation

Patients

The study cohort comprised 25 patients with MS (6 men and 19 women; median age, 44 years; age range, 25–65 years) who underwent a comprehensive brain MRI protocol, including 1 mm T2WI, T1WI, 3D FLAIR, DWI, C-DIR, and contrast-enhanced T1WI, in addition to DLS-DIR. The patients had chronic, clinically definite MS (2 with secondary progressive MS and 23 with relapsing-remitting MS; mean disease duration, 8.98 ± 7.54 years), diagnosed by the authors (H.Nishijima, C.S., and M.T., with 15, 22, and 30 years of experience in neurology, respectively). 21

Qualitative evaluation

An observer performance study was conducted to compare the interpretations of C-DIR and DLS-DIR images. Two experienced neuroradiologists (Radiologist A: S.I. and Radiologist B: K.W., each with 16 years of neuroradiology experience) participated in this analysis.

In the first reading session, conventional MR images (1 mm T2WI, T1WI, 3D FLAIR and contrast-enhanced spin-echo imaging) were analyzed. Next, the neuroradiologists were provided the C-DIR as well as the conventional MR images. In a subsequent session, the neuroradiologists reviewed DLS-DIR and conventional MR images, but not C-DIR images. This second reading session was conducted 8 weeks after the initial session to minimize recall bias. For the first and second reading sessions, the neuroradiologists identified and classified juxtacortical MS lesions based on their anatomical locations.

Before the reading sessions, the radiologists were briefed on (a) the focus of the study on patients with MS, (b) the task of detecting and classifying juxtacortical MS lesions by anatomical location, (c) the objective of counting lesions >1 mm in diameter, and (d) the objective of evaluating hyperintense lesions on T2WI, FLAIR, DIR, and hypointense lesions on T1WI. The neuroradiologists were familiarized with common previously reported artifacts on DIR images 10 , 19 , 20 and coronal or sagittal cross-section images were examined when required, as it was the case in the evaluation of healthy participants. Juxtacortical MS lesions 22 were classified into (a) subcortical WM lesions (isolated subcortical WM lesions), (b) intracortical lesions (isolated GM lesions), and (c) mixed lesions involving both GM and subcortical WM. Intracortical lesions and mixed GM/subcortical WM lesions are collectively referred to as cortical lesions. To exclude false positives, each detected lesion was retrospectively validated by 2 additional neuroradiologists (S.K. and T.S., with 29 and 7 years of experience, respectively).

Quantitative evaluation

Two neuroradiologists (T.S. and H.Nagaya, each with 7 years of experience) measured the contrast ratio (CR), contrast to noise ratio (CNR) and SNR by consensus. The CR, CNR, and SNR were measured for C-DIR and DLS-DIR. CR was calculated as C R = S I 1 − S I 2 S I 2 , where SI1 represents the signal intensity of the cortical lesion, and SI2 denotes the signal intensity of the GM. CNR was defined as C N R = S I 1 − S I 2 S D n o i s e , where SD noise represents the standard deviation (SD) of background noise. Background noise was measured by placing a ROI in the air within the FOV. CNR between GM and cortical lesions (intracortical or mixed GM/subcortical WM) was calculated as previously described. 12 For each patient with cortical lesions, the largest cortical lesion was selected by radiologist Sh.K for measurement. SNR was determined as the ratio of the signal intensity within the ROI to the SD of the signal within the same ROI. The SNR measurements were conducted in specific ROI located within CSF, WM, and GM. Specifically, the ROI for CSF was placed within the body of the lateral ventricles, that for WM in the centrum semiovale, and that for GM in regions devoid of MS lesions. Quantitative evaluations were performed both on C-DIR and DLS-DIR images.

Statistical analyses

Data analysis was performed using EZR (Saitama Medical Center, Jichi Medical University, Saitama, Japan). 23 The Wilcoxon Signed-Rank Test was used to assess the differences in lesion counts between C-DIR and DLS-DIR. Two analyses were conducted: (i) a global assessment to compare overall lesion detection using C-DIR vs. DLS-DIR across all patients and (ii) a patient-specific analysis to compare individual lesion counts per patient between C-DIR and DLS-DIR. The number of individuals in whom MS lesions were detected was analyzed using the McNemar test. The significance level was set at P  < 0.05. Interobserver agreement was evaluated using Cohen’s kappa coefficient, with agreement strength categorized as fair (Kappa: 0.21–0.40), moderate (Kappa: 0.4–0.60), good (Kappa: 0.61–0.80), and excellent (Kappa:>0.80). The SNR and CNR between C-DIR and DLS-DIR were statistically compared utilizing a paired t-test.

Results

Image quality evaluation in healthy participants

The neuroradiologists observed regions with high signal intensity in the recurved horn of the cortex, which were identified as actual signals (not artifacts) on both C-DIR and DLS-DIR sequences in 3 out of 5 healthy participants (60%). A total of 11 regions with high signal intensity were detected with conventional MRI in 3 healthy participants (1, 4, and 6 regions in each participant, respectively). The neuroradiologists noted that many of these regions were connected to regions with high signal intensity along the superficial cerebral cortex ( Fig. 1 ). Although the exact origin of these regions with high signal intensity regions was indeterminate, the neuroradiologists concluded that these findings on DIR image were unlikely to lead to misdiagnosis due to their locations.

Qualitative evaluation

Comparison of MS lesion detection between C-DIR and DLS-DIR

The mean longest diameters were 3.1 mm (95% confidence interval [CI]: 2.6–3.6 mm) for cortical lesions and 4.2 mm (95% CI: 3.4–5.0 mm) for subcortical lesions. The findings on MS lesion detection are detailed in Table 2 . In the first reading session, Radiologists A and B detected 97 and 99 cortical lesions, respectively. Both radiologists detected a significantly higher number of juxtacortical MS lesions using DLS-DIR compared to C-DIR (Radiologist A: total lesion counts of 376 vs. 261, P  < 0.05; Radiologist B: total lesion counts of 384 vs. 256, P  < 0.05; Fig. 2 ). The use of DLS-DIR notably improved the detection of intracortical lesions (Radiologist A: 22 additional lesions, Radiologist B: 34 additional lesions, P  < 0.05) and mixed GM/subcortical WM lesions (Radiologist A: 46 additional lesions, Radiologist B: 42 additional lesions, P  < 0.05; Fig. 3 ).

In the case-based analysis, out of 21 patients with cortical lesions (intracortical or mixed GM), 3 (14%) and 4 (19%) patients were identified by Radiologists A and B, respectively, using DLS-DIR imaging, but not using C-DIR images. However, only the intracortical lesion detection by Radiologist B showed a significant difference, and no other significant differences were seen. As measured by Cohen’s kappa coefficient, interobserver agreement was 0.6232 for the initial reading session with C-DIR and 0.7270 for the subsequent session with DLS-DIR, indicating “good” agreement in both sessions.

Quantitative evaluation

Among the 25 patients, 15 had cortical lesions (intracortical or mixed GM/subcortical WM lesions) detectable using both DIR imaging techniques. The results of the CR and CNR measurements for these 15 patients are summarized in Table 3 . No significant differences were found between C-DIR and DLS-DIR in CR and CNR measurements for GM and cortical lesions (CR: P  = 0.24; CNR: P  = 0.34). However, the SNR for CSF and WM was significantly higher with DLS-DIR compared to C-DIR ( P  < 0.05), while the SNR for GM did not differ significantly between the 2 imaging techniques ( P  = 0.42).

Discussion

A significantly higher number of cortical lesions were detected using DLS-DIR compared to C-DIR. To the best of our knowledge, this is the first clinical study to evaluate the utility of 3D imaging with DLS. Achieving high-spatial-resolution DIR imaging in routine clinical practice presents challenges due to the inherent trade-offs between imaging time, SNR, and spatial resolution. In this study, we utilized a DL-based acceleration technique (DLS) with a high acceleration factor of 10, enabling whole-brain DIR imaging at a high spatial resolution of 0.7 mm isotropic voxels with an acquisition time of 4 mins 23s. DLS-DIR improved diagnostic performance, as it allowed the identification of more juxtacortical MS lesions compared to the conventional imaging protocol (C-DIR).

Our results show that the use of DLS-DIR with high spatial resolution significantly improves the diagnostic accuracy for detecting cortical MS lesions, suggesting that DLS-DIR could be used to diagnose GM involvement in MS, which may be missed using conventional imaging protocols. Specifically, the improved detection of small MS lesions with DLS-DIR can be largely attributed to its higher spatial resolution, as no significant differences in CNR were observed between DLS-DIR and C-DIR. Additionally, the smaller voxel size of DLS-DIR mitigates partial volume effects and permits high-accuracy multiplanar reformations, facilitating detailed visualization of lesions in conventional and oblique planes from a single acquisition. Furthermore, the higher interobserver agreement observed with DLS-DIR compared to C-DIR suggests that the increased resolution of DLS-DIR improves the delineation of cortical lesions from subcortical WM lesions. This increased clarity contributes to more consistent and accurate diagnoses of juxtacortical MS lesions.

While higher spatial resolution can theoretically reduce CNR and SNR, our quantitative analysis revealed no significant differences in CNR between GM and cortical lesions for DLS-DIR compared to C-DIR. Our quantitative evaluation revealed that DLS allows for improved spatial resolution while maintaining image quality. These results are in line with quantitative results reported in previous studies on DL-based denoising methods. 14 – 17 Iwamura et al. reported that DL-based denoising methods improve CNR between MS lesions and brain parenchyma on T2WI. 15 The DLS method differs technically from DL-based denoising methods. Therefore, combining DLS with advanced denoising techniques could further enhance imaging performance and potentially reduce scan times.

A previous study has similarly shown that thinner-slice of 1 mm T2WI detects more MS lesions compared to conventional MRI of 5 mm T2WI. 15 In this study, numerous lesions that had not been identified in the initial reading sessions based on 1 mm T2WI and 3D FLAIR images, were subsequently detected using DLS-DIR in the second reading session. This finding underscores the fact that DLS-DIR provides additional diagnostic value for detecting juxtacortical MS lesions, complementing high-spatial-resolution T2WI and FLAIR imaging.

Image quality evaluation in healthy participants revealed regions with high signal intensity in the recurved horn of the cortex on DLS-DIR images. Their presence on both DLS-DIR and C-DIR images suggests that they are not unique to DLS but are inherent to the DIR sequence. While the exact cause of these findings is unknown, we speculate that they might have been at least partly due to the cortical vessels or fluid (CSF) flow within the sulci in combination with heterogeneity in the slice profile, as has been previously reported. 10 , 19 , 20 Thus, radiologists should be aware of the possibility of observing clinically irrelevant regions with high signal intensity on DIR images, and the need to minimize interpretation errors by examining coronal or sagittal cross-section DIR images. Additionally, to avoid confusion with artifacts and reduce the number of false positives, it is essential to evaluate these findings in conjunction with other sequences such as T2-FLAIR while assessing lesions across consecutive slices to improve diagnostic accuracy. Because these regions of high signal intensity in typical locations were easily identified on DIR images, we believe that misdiagnosis is unlikely to have occurred in the present study.

However, this study has some limitations, including its retrospective design and small sample size. Although demonstrating the relationship between cortical lesions and cognitive function is important, 24 , 25 no data on cognitive function was available at the time of MRI examination. Therefore, future prospective studies with larger cohorts are needed to validate the diagnostic value of DLS-DIR and its relationship with clinical outcomes including neurological impairment and cognitive decline. Additionally, our lesion assessment relied on radiological consensus, primarily based on DLS-DIR findings, and therefore may have included false positives. The lesions identified by DLS-DIR should have been compared to high SNR DIR as the gold standard for differentiating between MS lesions and artifacts. Nevertheless, due to scan time limitations, we were unable to acquire it in this study. However, we believe that our careful protocol for image analysis, which included interpretation of DLS-DIR data by using multiplanar reconstruction images, should have effectively minimized the possibility of false positives. The major strength of this study is that the high-resolution whole-brain DLS-DIR was completed in 4 mins 23s, demonstrating a clinically feasible imaging time.

Conclusion

The use of DLS technology enabled high-resolution DIR imaging with a clinically acceptable acquisition time of 4 mins 23s, significantly enhancing the assessment of cortical lesions in MS.