Introduction

Intracranial atherosclerotic disease (ICAD) is one of the most common causes of ischemic stroke, and serves as an important indication for the intracranial vessel wall imaging (VWI)1. The currently recommended VWI protocol in ICAD patients includes 3D T1-weighted images (T1WI) with and without contrast-enhancement2, which can detect imaging features of the culprit atherosclerotic plaques, such as contrast enhancement and intraplaque hemorrhage (IPH)1,3,4. However, due to the practical limitation of signal-to-noise ratio (SNR) in VWI, it is often challenging to identify these features or even detect the lesion itself, especially in small distal arteries5,6. Considering the size of intracranial arteries and the pathologic lesions, the recommended VWI protocols require both submillimeter in-plane resolution and sufficient SNR within a clinically acceptable scan time5,6. Many acceleration techniques, such as parallel imaging and compressed sensing, have been applied in VWI to achieve this goal5,7,8,9. Nevertheless, obtaining VWI with desirable image quality remains a challenging task.

Recently, image quality improvement of MRI is a promising application for deep learning (DL) in radiology10. VWI is an ideal candidate for DL-based image quality improvements, and a few studies have considered the benefit of DL-based noise reduction of VWI11,12. Meanwhile, super-resolution is another approach for image quality improvement that is commonly used in nonmedical images13. As recent reports have suggested the use of DL-based super-resolution on brain and neurovascular MRI14,15, super-resolution would be another well-suited method for improving the image quality of VWI.

In this study, we hypothesize that the improvement of image quality and resolution can increase the clinical value of VWI, particularly for the detection and characterization of small vessel wall lesions. To achieve this goal, we employed a DL-based method that simultaneously performs super-resolution and noise reduction on VWI. The purpose of this study is to assess the image quality of DL-augmented intracranial VWI and to explore the clinical value of the proposed method in detection and characterization of vessel wall lesions.

Methods

Study subjects

The institutional review board of The Catholic University of Korea, Seoul St. Mary’s Hospital, College of Medicine (KC22RISI0764), approved this single-center study and waived the requirement for informed consent due to its retrospective nature. The entire process of the research was performed in accordance with the declaration of Helsinki, and with relevant guidelines and regulations of our Institutional Ethics Committee.

A total of 143 patients underwent MRI with a high-resolution VWI protocol from January to July 2022 at our institution. The exclusion criteria were as follows: (1) deviant VWI protocols (n = 26) and (2) loss of unprocessed DICOM images (n = 0). Clinical parameters, including age, sex, indication of the VWI, presence of acute and chronic ischemic lesions, and cerebrovascular risk factors, were recorded. For the participants with acute ischemic strokes, we also collected the date of the image-confirmed infarct.

MRI protocols

All MRI examinations were performed on one of four 3.0-T scanners (MAGNETOM Vida; Siemens Healthineers, Erlangen, Germany) with 64-channel coils. The protocol for acquiring VWI included acquisition of 3D turbo spin-echo T1WI with and without contrast enhancement and proton density images (PDI). Oblique coronal scans were performed to cover the circle of Willis and both carotid bifurcations at once. Both sequences were scanned with voxel size of 0.51 × 0.51 × 0.45 mm3 within 6 min. Further detailed scan parameters are provided in Supplemental Materials. Contrast-enhanced T1-weighted images (CE-T1WI) were acquired 4 min after intravenous infusion of gadolinium-based contrast media with the same protocol as nonenhanced T1-weighted images (NE-T1WI).

Deep learning-based image quality improvement

This study utilized SwiftMR v2.2.0.0 (AIRS Medical, Seoul, Korea)16, an MR image reconstruction software incorporating a deep neural network (DNN) based on the U-Net architecture17. The model includes 18 convolutional blocks complemented by layers for max-pooling, up-sampling, and feature concatenation, culminating in three convolutional layers. This design aims to improve spatial resolution and contrast while minimizing noise.

In generating the training inputs, k-space sampling patterns were applied through multi-dimensional degradation of k-space data. Various undersampling patterns (uniform, random, kmax, partial Fourier, and elliptical) and noise addition were simulated and incorporated. Auxiliary contextual data, defining k-space sampling and expected noise reduction ratios, were prepared to assist the network in handling diverse learning tasks. To effectively utilize this auxiliary input, the U-Net architecture was modified to include a dynamic modulation pathway, resulting in a context-enhanced U-Net (CE U-Net).

L1-loss ensured reconstructed images closely matched high-resolution reference images and maintained anatomical precision. The learning rate, initially set at 0.001, was reduced by a factor of 0.1 after 10 epochs, followed by 3 additional epochs of training. The model was optimized using Adam18 and four GPUs (Tesla V100, NVIDIA Corporation, Santa Clara, CA, USA), each with 32 GB of memory. The training and validation datasets included approximately 3 million image slices (1 million 2D slices and 2 million 3D slices), covering a broad range of MR imaging sequences, anatomical locations, and imaging parameters.

Vessel wall images were part of both training and validation datasets described above, and they are mutually exclusive from the dataset used during this study to maintain the study’s analytical validity. The trained model transforms conventional vessel wall images (Conv-VWI) into deep learning-augmented vessel wall images (DL-VWI) by performing super-resolution (0.28 × 0.28 × 0.45 mm3) and noise reduction (Fig. 1).

Figure 1
figure 1

An example of DL-based image quality improvement on nonenhanced T1-weighted VWI. Conv-VWI (A) displays acceptable image quality and good wall-to-lumen contrast. DL-VWI (B) show excellent suppression of noise with well-preserved contrast. Especially, vessel walls are more clearly visualized on DL-VWI than on Conv-VWI (magnified views, bifurcation of internal carotid arteries).

Full size image

Image quality and vessel wall lesion assessments

Two neuroradiologists (with 18 and 7 years of experience, respectively) independently assessed overall image quality, visually perceived SNR, conspicuity of vessels, and degree of artifacts, in nonenhanced T1-weighted Conv-VWI and DL-VWI. The four parameters were rated on a 5-grade Likert scale (Supplemental Table 1). We referred to the image quality evaluation method from a previous study with our modifications9.

The two reviewers also independently assessed every delineable vessel wall lesion, regarding the features of confidence of detection, conspicuity, presence of IPH, and contrast enhancement (Fig. 2, Supplemental Table 1). IPH was considered present when a plaque exhibited a focal or segmental region with precontrast T1WI hyperintensity (> 150% of the normal wall) within the eccentric thickened wall lesion1,4. Lesion enhancement was scored as 0 (none), 1 (less than the pituitary stalk) and 2 (equal to or greater than the pituitary stalk); based on previous studies, only score 2 was considered positive for plaque classification1,19,20,21. The images were randomly divided into two complementary sets, each consisting of either Conv-VWI or DL-VWI per patient. Both reviewers were blinded to the patient order and the types of randomly shuffled images. A four-week interval was maintained between the assessments of the two image sets to avoid recall bias.

Figure 2
figure 2

An example of a magnified vessel wall lesion on Conv-VWI (A, coronal; C, axial) and DL-VWI (B, coronal; D, axial). The eccentricity of the plaque and the presence of internal T1WI hyperintensity compared to the adjacent normal vessel wall are more clearly depicted on DL-VWI.

Full size image

Quantitative image quality assessment was based on a 3D semiautomatic segmentation of the normal vessel walls and lumens of the 1st segments of the middle cerebral arteries (M1s) and basilar arteries (BAs) in NE-T1WI, supported by PDI (Fig. 3). During the SNR and contrast-to-noise ratio (CNR) calculation, noise was measured in the vessel lumen. All measured values were normalized across subjects based on the reference tissue (central pons). The same procedure was performed for the included vessel wall lesions, both in NE-T1WI and CE-T1WI (Fig. 3). The entire assessment was performed using dedicated software (3D Slicer, version 5.0.3, http://slicer.org) by an experienced neuroradiologist, and the further detailed methods are provided in Supplemental Materials.

Figure 3
figure 3

An example of the segmentation process for the vessel wall, lumen, and vessel wall lesions. After coregistration of NE-T1WI (A, B) and PDI (C), the regions of interest for a normal vessel wall and lumen are drawn semiautomatically (outer margin, yellow; inner margin, orange). The region of interest for a delineable middle cerebral arterial wall lesion is shaded in yellow on NE-T1WI (D, E) and CE-T1WI (F).

Full size image

Atherosclerotic plaque classification

The reference standard was determined after completing the entire image analysis procedure, which was blinded to the previous image analysis. All detected atherosclerotic plaques were deliberately classified by the consensus of two experienced neuroradiologists into potentially culprit and nonculprit plaques. A plaque was considered potentially culprit if it met one of the following criteria: (a) it was the only lesion or (b) it was the most stenotic lesion in the vascular territory associated with a concurrent image-proven acute ischemic stroke1,19,20. All plaques located in other vascular territories, and the indeterminate plaques other than the most stenotic lesion within the stroke vascular territory, were regarded as nonculprit plaques. The presence and accurate location of these lesions were confirmed by consensus between two neuroradiologists, using not only Conv-VWI and DL-VWI but also all other available clinical information and radiologic modalities, including MR, CT or digital subtraction angiographies.

Statistical analysis

The clinical characteristics of the patients are provided as descriptive statistics. Cohen’s quadratically weighted Kappa coefficients (κ) were calculated based on the quality ratings of the two reviewers. The criteria of Landis and Koch were used to interpret these values, with κ values of 0.00–0.20, 0.21–0.40, 0.41–0.60, 0.61–0.80, and 0.81–1.00 indicating slight, fair, moderate, substantial, and almost perfect agreement, respectively22. The Conv-VWI and DL-VWI scores were compared using the Wilcoxon signed-rank test. Values of the quantitative assessment, i.e., SNR, CNR, normalized mean signal intensity (SI) and standard deviation (SD) of vessel walls, lumens and lesions were compared with the Wilcoxon signed-rank test or paired t-test, according to their normality of distribution. The diagnostic performance of plaque enhancement and IPH in identifying potentially culprit plaques was assessed using sensitivity, specificity and area under the receiver operating curve (AUROC). Differences between the AUROCs of Conv-VWI and DL-VWI were compared using DeLong’s test. All statistical analyses were performed using R statistical software (version 4.0.3; R Foundation, Vienna, Austria).

Results

Subject characteristics

A total of 117 eligible patients were included in our study, and their overall clinical characteristics are shown in Table 1. Among the eligible patients, 15 patients showed markedly degraded images in the vessel wall segmentation step and were excluded from the subsequent quantitative analysis (Supplemental Fig. 1). The most frequent indication on VWI was workup for acute ischemic stroke (74/117, 63.2%), followed by assessment of steno-occlusive intracranial arterial lesions (23/117, 19.7%) and suspicion of dissection (11/117, 9.4%). Approximately two-thirds of the patients had acute infarcts, with a mean interval of 5.41 days (range: 1–21 days) between the image-confirmed infarct and VWI.

Table 1 Clinical characteristics of the included patients and vessel wall lesions.
Full size table

A total of 86 vessel wall lesions were delineated from the VWI of 56 patients. Among them, 70 lesions were atherosclerotic plaques, and 11 lesions were suspected to be dissections. Four lesions from 4 patients were excluded in lesion assessment using CE-T1WI due to lack of contrast-enhanced images.

Image quality and vessel wall lesion assessments

Table 2 displays the results of the image quality and lesion assessments. Overall image quality (4.18 ± 0.72 vs. 2.94 ± 0.62), visual SNR (4.32 ± 0.68 vs. 2.59 ± 0.55) and vessel wall conspicuity (4.26 ± 0.69 vs. 3.12 ± 0.65) were significantly higher in DL-VWI than in Conv-VWI. While the interrater agreement was fair to substantial for both Conv-VWI and DL-VWI, it was generally higher in DL-VWI than in Conv-VWI (Table 2).

Table 2 Comparison of image quality and vessel wall lesion assessments between Conv-VWI and DL-VWI.
Full size table

In the vessel wall lesion assessment, confidence of detection (4.54 ± 0.56 vs. 2.59 ± 0.79) and conspicuity (4.30 ± 0.67 vs. 3.00 ± 0.83) were significantly higher in DL-VWI compared to Conv-VWI with moderate to substantial agreement. The two reviewers made more concordant judgments (discordant results 5 vs. 11; Kappa 0.85 vs. 0.59) and reported a higher presence of IPH (19 vs. 11) in DL-VWI compared to Conv-VWI. The visual degree of lesion enhancement was slightly but significantly higher in DL-VWI (1.40 ± 0.65 vs. 1.26 ± 0.76) with substantial to almost perfect interrater agreement.

On quantitative analysis the SNR and CNR of normal vessel walls in DL-VWI were significantly higher than those in Conv-VWI (SNR M1 4.04 ± 1.05 vs. 3.38 ± 0.81, BA 4.83 ± 1.23 vs. 3.02 ± 0.59; CNR M1 2.21 ± 0.83 vs. 2.00 ± 0.75, BA 1.80 ± 0.75 vs. 1.32 ± 0.53), and this difference was larger in BAs than M1s (Supplemental Fig. 2). The decrease of SD in DL-VWI, which represents decreased image noise, was more prominent in the central area of images (BA) than in their periphery (M1) (Supplemental Table 2).

The vessel wall lesions demonstrated significantly higher SNR and CNR in DL-VWI compared to Conv-VWI, both in CE-T1WI (SNR 22.04 ± 11.62 vs. 8.30 ± 3.25, CNR 16.67 ± 11.05 vs. 6.49 ± 3.27) and NE-T1WI (SNR 7.56 ± 3.15 vs. 4.76 ± 1.34, CNR 4.68 ± 2.60 vs. 3.05 ± 1.26, Supplemental Fig. 3). The lesions in CE-T1WI demonstrated slightly but significantly higher SI in DL-VWI (Supplemental Table 2), in line with the qualitative lesion assessment. The decrease of SD in wall lesions was weaker compared to normal vessel walls and lumens, especially in CE-T1WI (normal walls and lumens, all P < 0.001; lesions in CE-T1WI, P = 0.10; Supplemental Table 2). Notably, culprit plaques in CE-T1WI showed slightly higher SD in DL-VWI than in Conv-VWI (44.13 ± 12.43 vs. 42.44 ± 10.66), but this difference was not statistically significant (P = 0.10).

Atherosclerotic plaque classification using VWI

The diagnostic performance of Conv-VWI and DL-VWI in plaque classification is demonstrated in Table 3. Plaque enhancement showed similar sensitivity (88.2% vs. 88.2%) and specificity (83.7% vs. 81.6%) between Conv-VWI and DL-VWI (AUROC 0.86 vs. 0.85, P = 0.32). IPH exhibited only 23.5% sensitivity in Conv-VWI, which has significantly increased to 70.6% in DL-VWI, while maintaining excellent specificity in both VWIs (94.3% and 94.3%, respectively). DL-VWI was significantly better for detection of potentially culprit plaque using IPH (AUROC 0.59 vs. 0.82). When combination of enhancement and IPH was used for culprit plaque classification, DL-VWI showed higher AUROC (0.91; sensitivity 100%, specificity 81.6%) in diagnosing culprit plaques than Conv-VWI (0.84; sensitivity 88.2%, specificity 79.6%), with marginal significance (P = 0.10).

Table 3 Comparison of diagnostic performance in identifying culprit plaques between Conv-VWI and DL-VWI.
Full size table

Representative cases

Figure 4 demonstrates better depiction of the culprit plaque characteristics on DL-VWI. DL-VWI was also helpful in detecting subtle plaque enhancement (Fig. 5) and lesions in severely motion-degraded images (Supplemental Fig. 4). Identification of dissection, especially differentiating it from atherosclerosis, was more confidently performed when nonenhanced and contrast-enhanced DL-VWI were available (Supplemental Figs. 5, 6). Another example culprit lesion containing IPH was used to measure signal profiles, both on NE-T1WI and CE-T1WI (Fig. 6).

Figure 4
figure 4

Two illustrative cases of potentially culprit plaques. (A) In a 62-year-old male with abrupt dysarthria and pontine infarct, only mild distal basilar stenosis was identified on MR angiography. (B) Conv-VWI revealed the plaque with punctate T1WI hyperintensity, which was interpreted as an intraplaque hemorrhage (IPH) by only one reviewer. (C) DL-VWI more clearly depicts the apparent T1WI hyperintensity and positive remodeling. (D) Another 66-year-old male presented with abrupt left homonymous hemianopia, showing right posterior cerebral artery territory infarct with focal distal stenosis. (E) Same as Fig. 1, the signal of the plaque in Conv-VWI was misinterpreted as image noise by both reviewers. (F) DL-VWI provides a clearer depiction of the presence of IPH.

Full size image
Figure 5
figure 5

A 71-year-old female with a segmental atherosclerotic plaque along the 1st segment of the left middle cerebral artery. The shape of the wall lesion is more clearly delineated in DL-VWI (D–F) than in Conv-VWI (A–C). In contrast-enhanced Conv-VWI (B and C), contrast enhancement is not clearly seen due to the background noise of the vessel wall and brain parenchyma. However, contrast-enhanced DL-VWI (E and F) clearly depict subtle enhancement at the distal portion of the wall lesion.

Full size image
Figure 6
figure 6

An example of a potentially culprit vessel wall lesion containing intraplaque hemorrhage. The signal profile is measured using both Conv-VWI (A, NE-T1WI; D, CE-T1WI) and DL-VWI (B, NE-T1WI; E, CE-T1WI). Both signal peaks from the normal vessel wall and the lesion are of similar height when comparing nonenhanced Conv-VWI and DL-VWI (C). On the other hand, contrast-enhanced images display a slightly higher signal intensity peak in DL-VWI compared to Conv-VWI (F). Note that the background noise signal variation is more flattened in DL-VWI, both in nonenhanced (C) and contrast-enhanced images (F).

Full size image

Discussion

This study demonstrated the clinical value of deep learning-based noise reduction and super-resolution for improving the image quality of intracranial vessel wall imaging. The deep learning-based approach simultaneously improved image quality and spatial resolution, which was feasible and useful for the clinical evaluation of vessel walls. This assisted in detecting vessel wall lesions and identifying potentially culprit plaques, particularly by better characterizing intraplaque hemorrhage.

Prior approaches to acquire VWI with acceleration techniques were practically limited in terms of the acceleration factor that can be achieved while maintaining image quality7,8,9. For example, to attain DL-VWI resolution (0.28 × 0.28 × 0.45 mm3) using conventional acceleration while minimally preserving SNR (similar as Conv-VWI), it requires at least a fourfold increase in acquisition time (24 min) compared to the current protocol, which is not practical in most clinical circumstances. DL-based approaches can resolve this clinical dilemma. Another benefit of our approach is that direct application is possible without any specialized or dedicated method for acquisition or reconstruction. As shown by our results, DL-based image quality improvement can be successfully incorporated into MR images already obtained with conventional acceleration techniques.

Many reports have explored the value of DL-based MRI quality improvement for imaging of various body parts, mostly focusing on noise reduction11,12,23,24,25,26. Meanwhile, there have been reports regarding the potential of DL-based super-resolution processing for MRI14,15,27,28. In this study, we applied a model that performs both denoising and super-resolution simultaneously. This allows seamless integration of the two competing goals for VWI, aiming to achieve the recommended “ideal” resolution near the histologically suggested normal vessel wall thickness while retaining excellent image quality5. Compared to the pilot study, which used same quality ratings12, even a slight decrease in voxel size (0.56 × 0.56 × 0.5 mm3 vs. 0.51 × 0.51 × 0.45 mm3) largely degraded the visually rated SNR of the original images (3.5 vs. 2.59). The combination of DL-based denoising and super-resolution effectively compensated for the threefold decrease in voxel size (0.28 × 0.28 × 0.45 mm3), resulting in high SNR in both qualitative and quantitative assessments.

We demonstrated the clinical feasibility of DL-VWI in detecting potentially culprit plaques, aided by better identification of the IPH and contrast enhancement. In previous reports, IPH generally showed a low sensitivity (12.1‒49.5%) but high specificity (66.7‒98.3%) in detecting culprit plaques1,4,21,29,30, more similar to Conv-VWI in our results. Our DL-VWI identified an additional 8 potentially culprit plaques (47.1% out of 17 plaques; Table 3) containing IPH that Conv-VWI did not detect properly, while maintaining excellent specificity. Meanwhile, plaque enhancement was not significantly different between Conv-VWI and DL-VWI for culprit plaque detection. Similar to our results, previous studies reported a high prevalence of grade 2 enhancement among culprit plaques (50.0‒92.3%), suggesting high diagnostic sensitivity1,19,21,29,31. Our study presents that DL-VWI could be particularly beneficial for culprit lesion detection in patients unable to use contrast materials. It should also be emphasized that detection and delineation of the lesion itself were more confidently performed in nonenhanced DL-VWI.

DL-VWI better revealed the presence of IPH, despite a slightly lower precontrast mean SI of potentially culprit plaques compared to Conv-VWI (88.13 ± 20.37 vs. 89.14 ± 18.49, P < 0.001, Supplemental Table 2). In Fig. 6, the SI of the normal vessel wall and the lesion did not substantially increase in nonenhanced images. Instead, signal variation from background noise was more suppressed in DL-VWI, resulting in better contrast between the lesion and cerebrospinal fluid (CSF). On the other hand, the intrinsic signal heterogeneity of the culprit plaques, caused by various internal components32 and presented as higher SD than other lesions, was retained after DL-based denoising (Supplemental Table 2). We speculate that the DL-based image transformation may preserve the true heterogeneous appearance of culprit plaques, such as IPH, making them more noticeable (Fig. 5).

Interestingly, CE-T1WI showed slightly but significantly higher lesion enhancement in DL-VWI, both qualitatively and quantitatively. These findings address the concerns raised by a prior meta-analysis regarding the reduced visibility of small enhancing lesions in high-resolution images due to decreased partial volume effect3. DL-based processing might be a practical solution for this issue, which suppresses undesired background noise while preserving the true enhancement signal (Fig. 6). However, this increased enhancement did not improve the diagnostic performance of DL-VWI in plaque characterization significantly. More than three-quarters of potentially culprit plaques already displayed strong enhancement in Conv-VWI, indicating that the clinical benefit of DL-VWI for detecting contrast enhancement was not remarkable.

There are a few points that should be kept in mind for the clinical application of DL-VWI. Although the DL-based approach is a black box and cannot be fully understood, there were some findings that may imply signal referencing from adjacent structures. For example, a slight increase in lumen SI might be affected by the features of the adjacent vessel walls. Lower SI of the normal BA walls in DL-VWI could be associated with surrounding CSF and susceptibility-related signal loss of the skull base, whereas M1 walls, which are more frequently adjacent to gray matter, showed higher SI in DL-VWI (Supplemental Table 2). Quantitative signal increase of the lesions in contrast-enhanced DL-VWI may also be explained by this signal referencing (Fig. 6). Although this phenomenon may accentuate some signal characteristics of input images, it has not produced false positive lesions based on the observation of this study. However, further evaluation is required in real-world settings, especially for the lesions in images involving tiny structures and subtle signal changes such as VWI.

This study has several limitations. First, this is a retrospective study with a limited number of patients. The reference standard was not pathologically proven, so indeterminate lesions less stenotic than the culprit plaque were assumed to be nonculprit to avoid excluding them or designating two culprits for one infarct territory, thereby minimizing selection bias. Owing to the clinically unfeasible acquisition time as mentioned above, the reference standard scans not utilizing DL for super-resolution images could not be obtained for each patient. Additionally, during the quality rating, radiologists may have recognized apparent quality differences in some images even after shuffling, constituting a failure of image type blinding and prevention of observer bias. Coregistration of the images for quantitative measures may not be perfect in some cases, since normal vessel walls are extremely thin and misregistration of a single voxel may cause partially inadequate segmentation. Finally, SNR was calculated only on a single clinical scan per patient, which might lead to errors for images with parallel imaging.

Conclusions

In conclusion, the application of deep learning-based noise reduction and super-resolution to intracranial VWI successfully improved the quality and spatial resolution of the images. Deep learning-augmented VWI led to increased detection of the vessel wall lesions and intraplaque hemorrhage compared to conventional VWI, suggesting its clinical feasibility in identifying potentially culprit plaques.