MEASURING IMPACT OF SUPER-RESOLUTION ON SPINAL CORD MRI SCANS: LESION DETECTION SENSITIVITY, VARIABILITY, AND CLINICAL IMPACT

Abstract

The efficacy and potential for MRI-derived spinal cord (SC) information is an area of great interest for the study of multiple sclerosis (MS). Though the presence of SC lesions aid in the diagnosis of MS or other disease/injury, there is much debate as to whether lesions can help to predict symptoms or disability. The correlation between spinal cord lesions and MS disability is weak, even when observed at a higher resolution. A current drawback in the collection of SC magnetic resonance imaging (MRI) scans is the inability to collect high-resolution images both through-plane and in-plane. Super-resolution offers the opportunity to transform these anisotropic MRIs into high-resolution, isotropic images offering a view not previously possible. Here, we investigate how artificially altering the resolution of SC MRIs (either through super-resolution or linear interpolation) might enhance our ability to discern clinically relevant structures, including lesion load and its relation to several clinical variables, such as EDSS, across 53 patients with varying MS severity. Artificially altering the MRIs to varying levels of isotropic resolution increased sensitivity for lesion segmentation using open-source deep learning tools, but no significant association between lesion load/volume and EDSS disability measurement was found.

1. INTRODUCTION

Magnetic Resonance Imaging (MRI) is a non-invasive imaging technology that produces detailed three-dimensional anatomical images [1]. Spinal cord MRI is much more challenging than brain MRI because the spinal cord is long and thin with some degree of mobility; breathing, heartbeat, CSF flow, and aortic flow introduce artifacts that further complicate spinal cord scans [2]. For patients that are unable to sit still for long periods of time (i.e., young children or geriatric patients), MRI scans can be corrupted by movement. This necessitates the need for both faster and higher resolution scans that will enable radiologists to better detect abnormalities and disease through spinal cord anatomy. One such disease of great interest is multiple sclerosis (MS). Characterized by perivenular and spinal cord inflammatory lesions, MS can result in fatigue, vision problems (such as decreased visual acuity, color vision, contrast sensitivity, and depth perception), pain, and cognitive impairment [3]. However, without MRI, it can be very difficult to distinguish between differential diagnoses and MS given a set of symptoms.

The cross-sectional area (CSA) of the entire SC is known to be a sensitive bio marker for MS, among other various diseases and traumatic/non-traumatic spinal cord injuries [4]. Additionally, axial views of SC MRIs allow radiologists to quantify the extent of MS lesions relative to the SC gray matter [4]. However, the correlation between SC lesions and MS disability markers is notoriously weak, even when observed at higher resolution [5].

Little work has been done in the field of super-resolution on the SC, especially as it relates to the improvement of MS lesion quantification and qualification. Super-resolution offers the opportunity to take low-resolution scans obtained at faster speeds and lower costs and transform them into high resolution images. A self-supervised, super-resolution (self-SR) technique, Synthetic Multi-Orientation Resolution Enhancement (SMORE), has been proposed to improve brain MRI resolution [6]. Inspired by SMORE, a new self-SR method (denoted self-SR* from here on) was recently proposed which achieves more robust performance through modeling slice gap with ESPRESO and residual image prediction [7], [8]. However, the efficacy of self-SR in the spinal cord is an active area of research. Recent studies showed that super-resolution of artificially degraded scans using SMORE resulted in similar SC atrophy measurements of the CSA, mean upper cervical cord area, and associated clinical outcome correlations to the high-resolution ground truth [9], [10]. Though these studies provide evidence that SMORE/self-SR* can effectively estimate high-resolution images from artificial, low-resolution acquisitions, they do not investigate lesion information between the ground truth and super-resolved images. Additionally, these studies do not offer insight into further clinical information gained through performing other resolution enhancement methods, such as linear interpolation.

As such, the utility of super-resolved images in clinical and research workflows has not yet been fully investigated. In this work, we explore the efficacy of super-resolution toward MS lesion quantification using self-SR* [7]. Specifically, we investigate lesion load per cervical level and SC gray matter/white matter volume differences between original scans, linearly resampled scans, and super-resolved scans of 53 people with MS (PwMS) (Fig. 1). Quantitative spinal cord volume measurements for varying scan resolutions are made to ensure consistency between scans. The goal of this study is to determine how artificially improving the resolution of spinal cord MRIs might enhance our ability to discern clinically relevant structures, including lesion load and relation to physiological biomarkers.

Fig. 1.

Self-SR* performed on spinal cord MRIs before lesion segmentation using the Spinal Cord Toolbox [11]. Performance of super-resolution can be compared against original and linearly resampled scans.

2. METHODS

2.1. Data collection

Scans from 53 PwMS were obtained from the Vanderbilt University Medical Center (VUMC). This dataset includes high-resolution, axial, multi-slice, multi-echo gradient echo (mFFE) anatomical images (0.3125 x 0.3125 x 5 mm), sagittal T₂-weighted (T2) anatomical images (0.4878 x 0.4878 x 2 mm), and axial T₂*-weighted (T2*) anatomical images (0.2539 x 0.2539 x 5 mm) for all patients.

2.2. Pre-processing

A pre-processing pipeline used Spinal Cord Toolbox (SCT) open-source functions [11]. In order to compare spinal cord measurements across varying resolutions of an image, four total groups of scans were prepared: original (original scan resolution), resampled (linear interpolation to 1 x 1 x 1 mm), super-resampled (linear interpolation to the highest resolution in the volume), and super-resolved (self-SR*, also to the highest resolution in the volume) (Table 1).

Table 1.

Summary of scan modalities and associated resolutions/dimensions analyzed.

Cohort	Modality	Scan Dimensions
Original	mFFE	0.3125 x 0.3125 x 5 mm
	T2*	0.2539 x 0.2539 x 5 mm
	T2	0.4878 x 0.4878 x 2 mm
Resampled	mFFE	1 x 1 x 1 mm
	T2*	1 x 1 x 1 mm
	T2	1 x 1 x 1 mm
Super-Resampled	mFFE	0.3125 x 0.3125 x 0.3125 mm
	T2*	0.2539 x 0.2539 x 0.2539 mm
	T2	0.4878 x 0.4878 x 0.4878 mm
Super-Resolved	mFFE	0.3125 x 0.3125 x 0.3125 mm
	T2*	0.2539 x 0.2539 x 0.2539 mm
	T2	0.4878 x 0.4878 x 0.4878 mm

For both linear interpolation resampling procedures, all scans were set to a common orientation before being resampled to an isotropic resolution of 1 x 1 x 1 mm (denoted “resampled”) or the highest resolution in the volume (denoted “super-resampled”). Because self-SR* learns a mapping between low-resolution to high-resolution patches based on the high-resolution 2D slices within an image, super-resolved scans have isotropic resolutions that match that of the highest resolution 2D slice [7].

SCT’s deep-learning-based SC segmentation algorithm (sct_deepseg_sc) was used on all scan types to obtain whole SC segmentations and centerlines. T2 cervical levels were segmented before the scan was registered (using sct_register_to_template) to the PAM50 template [12]. Conversely, mFFE and T2* axial scans were registered using sct_register_multimodal, initializing a warping and inverse warping field with the T2 sagittal warping fields already obtained from the same patient. Any scans with extensive motion artifacts or those that failed any step of pre-processing were removed from further analysis.

2.3. Lesion and CSA quantification

Lesion segmentation was obtained for each scan using sct_deepseg_lesion. Finally, lesion volume and CSA across the C2-C4 vertebral levels were obtained through multiplying the lesion mask with the desired atlas of interest (from PAM50) before computing per-slice/per-level statistics with sct_process_segmentation. Multiplication of the lesion masks and the associated atlases was achieved by using sct_maths. We analyzed total spinal cord volume and lesion load in the gray matter, white matter, (GM/WM) and in the corticospinal tract (CST), between each scan resolution and modality for all patients included after pre-processing. Friedman tests to check for significant differences among cohorts were utilized, followed by post-hoc Wilcoxon signed-rank test with Benjamini-Hochberg correction for all pairwise comparisons. Only vertebral levels C2-C4 were included because the protocol covered through C4.

3. RESULTS

Qualitative differences between cohorts can be viewed in Fig.2

Fig. 2.

Visualized differences between resolution cohorts in axial slices of T2 sagittal scans. The upper two rows show an example with lesions with and without segmentation (lesion segmentations shown in red). The bottom row shows an example without lesions.

Resampling and super-resampling did not have a significant effect on total SC volume measured by sct_deepseg_sc in the GM/WM when compared to the original scans (p = 0.143 and p = 0.272, respectively) and to each other (p = 0.462). However, super-resolving did have a significant effect on SC volume measurement in the GM/WM when compared to all other cohorts (p < 0.001 for all comparisons), increasing the average SC volume from 3145.13 mm³ in the original cohort to 3239.58 mm³. Restricting analysis to the CST, total SC volume varied widely between cohorts, with original scans having the highest average volume (584.54 mm³) and super-resampled having the lowest (524.82 mm³). In the CST case, there were statistically significant differences between all cohorts except for original and resampled scans (Table 2).

Table 2.

Summary of Wilcoxon signed-rank tests for SC volume between cohorts. Significant values are in bold.

SC Volume Comparison	GM/WM p-value	CST p-value
Original vs. Resampled	0.143	0.258
Original vs. Super-Resampled	0.272	8.59e-9
Original vs. Super-Resolved	7.16e-8	1.76e-5
Resampled vs. Super-Resampled	0.462	4.53e-12
Resampled vs. Super-Resolved	1.13e-9	3.39e-7
Super-Resampled vs. Super-Resolved	5.79e-10	2.45e-7

To measure how consistently a given technique affected SC segmentation, we measured the average standard deviation in SC volume across modalities/scan types for each given patient. Original scans had the lowest standard deviation (272.48 mm³), whereas resampled had the highest (347.41 mm³). Super-resampled and super-resolved had a similar standard deviation (299.34 mm³ and 299.38 mm³, respectively). Within the CST, super-resolved had the lowest (58.56 mm³) and original had the highest (115.27 mm³); as a consistent trend, standard deviation decreased from original to super-resolved. Within the GM/WM, there were no significant differences between the standard deviations of SC volume per patient across all cohorts. In the CST, there were significant differences only between the original cohort and all others (p << 0.05 for all comparisons).

Lesion segmentations on super-resolved scans yielded the greatest total average lesion volume in the GM/WM across scans (182.08 mm³), as well as the greatest average lesion volume in the CST (25.49 mm³). As a trend, lesion volume increased with resolution enhancement from original to super-resolved (Fig. 3).

Fig. 3.

Total Lesion Volume (mm³) vs. Resolution Cohort across C2-C4 vertebral levels for gray/white matter. Lesion volume is log-transformed (using natural log) with lines between individual scan data points for better visualization. Bars located at the top of the graph represent a significant difference between the average total lesion volume between cohorts (***: p < 0.001; **: p < 0.01; *: p < 0.05).

To measure consistency of lesion segmentations, we measured the average standard deviation in lesion volume across modalities/scan types for each given patient (Fig. 4). Resampled scans resulted in the lowest average standard deviation in lesion volume across a patient when measured across all of the GM/WM (78.93 mm³) and the CST (14.60 mm³). Super-resolved scans had the highest for all GM/WM (195.60 mm³) and CST (26.90 mm³).

Fig. 4.

Total Lesion Volume STD (mm³) vs. Resolution Cohort across C2-C4 vertebral levels. Data points represent averaged standard deviations across a given patient, and lines between them provide trend visualization. Bars located at the top of the graph represent a significant difference between the average standard deviation per patient between cohorts (***: p < 0.001; **: p < 0.01; *: p < 0.05).

Artificially increasing resolution, either through linear interpolation or super-resolution, increased overall lesion volume found by the SCT deep-learning lesion segmentation tool. However, it had varying effects on producing consistent lesion segmentations depending on the given resolution and technique used. For example, linear interpolation in the resampled and super-resampled groups resulted in less variability than the original, unaltered scans. These results are summarized in Table 3.

Table 3.

Summary of Lesion Volume Statistics. Bold values represent the highest average lesion volume or lowest STD.

Cohort	Spinal Cord Region	Avg. Lesion Volume (mm3)	Avg. Lesion STD (mm3)
Original	GM/WM	101.04	144.05
	CST	14.67	20.42
Resampled	GM/WM	112.09	78.93
	CST	19.93	14.60
Super-Resampled	GM/WM	141.66	114.28
	CST	16.43	15.23
Super-Resolved	GM/WM	182.08	195.6
	CST	25.49	26.90

Total lesion volume vs. EDSS score was plotted for each of the scan types (mFFE axial, T2 sagittal, and T2* axial) and for each resolution in both GM/WM and CST. Across all types, no correlation (p > 0.05) between total lesion volume and EDSS was found for all resolutions.

4. DISCUSSION

Artificially increasing the resolution of SC MRIs increases lesion segmentation sensitivity, resulting in a greater overall detected lesion volume. This is a general trend across both gray/white matter and the corticospinal tract. Super-resolving an SC MRI using the self-SR* technique results in the greatest overall lesion volume detected. Despite improved sensitivity to lesion detection, we see no correlation to clinical disability in multiple sclerosis.

The standard deviation in lesion volume across scan types for a given patient decreased when using linear interpolation as a resampling method, with the best performance coming from a conservative resampling to 1x1x1 mm resolution. However, super-resolved scans resulted in higher standard deviation of lesion volume than their original, unaltered counterparts.

Artificially increasing MRI resolution can help automated imaging tools to detect anatomical structures, decreasing variability among scan type performance in the process. As such, this should be considered as a potentially important step in utilizing image analysis tools in clinical practice, especially when high-resolution imaging is not possible, such as in the case for SC MRIs.

Lastly, these results suggest that self-supervised super-resolution might either introduce nonexistent structures or inconsistently affect a given image. Since super-resolution increases overall SC volume and lesion load detection while increasing variability (except for SC volume in the CST), users should be cautious of using super-resolved SC MRI scans without considering other orientations and scan modalities for cross comparison.

5. CONCLUSION

This study investigated how super-resolution and basic linear interpolation affect the performance of open-source deep-learning segmentation tools for lesion detection in the spinal cord. Artificially increasing the resolution of MRI scans in the spinal cord should be considered as a possibly valuable step in real practice due to its potential to increase lesion detection sensitivity and decrease variability among scan modalities. Utilizing self-supervised super-resolution techniques such as self-SR* allows for the greatest sensitivity increase but at a cost of increased variability. In addition, this work reaffirms past findings on the clinico-radiological paradox, as no correlation between lesion volume and multiple sclerosis disease severity was found, despite increased sensitivity to lesion detection. Future work could investigate the importance of lesion location and the effects of other super-resolution techniques (especially any that are designed specifically for the spinal cord) on identifying lesions or the central vein sign in the spinal cord.