Automated workflow for volumetric assessment of signal intensity ratio on T1-weighted MR images after multiple gadolinium administrations

Abstract.

Purpose: Repeated injections of linear gadolinium-based contrast agent (GBCA) have shown correlations with increased signal intensities (SI) on unenhanced T1-weighted (T1w) images. Assessment is usually performed manually on a single slice and the SI as an average of a freehand region-of-interest is reported. We aim to develop a fully automated software that segments and computes SI ratio of dentate nucleus (DN) to pons (DN/P) and globus pallidus (GP) to thalamus (GP/T) for the assessment of gadolinium presence in the brain after a serial GBCA administrations.

Approach: All patients ($N=113$) underwent at least eight GBCA enhanced scans. The modal SI in the DN, GP, pons, and thalamus were measured volumetrically on unenhanced T1w images and corrected based on the reference protocol (measurement 1) and compared to the SI-uncorrected-modal-volume (measurement 2), SI-corrected-mean-volume (measurement 3), as well as SI-corrected-modal-single slice (measurement 4) approaches.

Results: Automatic processing worked on all 2119 studies (1150 at 1.5 T and 969 at 3 T). DN/P were $1.085\pm 0.048$ (1.5 T) and $0.979\pm 0.061$ (3 T). GP/T were $1.084\pm 0.039$ (1.5 T) and $1.069\pm 0.042$ (3 T). Modal DN/P ratios from volumetric assessment at 1.5 T failed to show a statistical difference with or without SI corrections ($p=0.71$). All other $t$-tests demonstrated significant differences (measurement 2, 3, 4 compared to 1, $p<0.001$).

Conclusion: The fully automatic method is an effective powerful tool to streamline the analysis of SI ratios in the deep brain tissues. Divergent SI ratios using different approaches reinforces the need to standardize the measurement for the research in this field.

1. Introduction

Gadolinium-based contrast agents (GBCA) have been widely used in clinical magnetic resonance imaging procedures since its introduction over 30 years ago. Although the benefit-risk profiles of GBCA are favorable in the vast majority of patients, the usage of GBCA has been associated with the occurrence of nephrogenic systemic fibrosis in patients with renal dysfunction. Recently, repeated injections of linear GBCA was also correlated with increased signal intensities (SI) on unenhanced T1-weighted (T1w) images in the globus pallidus (GP) and dentate nucleus (DN) in patients with normal renal function.^1–3 Such SI increase was attributed to the presence of gadolinium in brain tissue and confirmed in several postmortem studies.^4–6 Existing data suggest SI increase varies among the classes of GBCAs with robust evidence of SI increase seen only in patients who received injections of linear GBCAs.^7–11 While the knowledge gaps including the molecular mechanism, the long-term biodistribution, and the toxic potential of chronic retention remain to be elucidated, the methodology in the measurement and validation of gadolinium presence should also be developed, as per a group of experts/opinion leaders in this area.^12,13

On the technical aspects, many of such studies require the collection and analysis of large amount of clinical data on this topic. Although designing studies and collecting clinical data can be daunting, the analysis of such data is no trivial task either, especially when the scope of the study involves a large number of patients and/or multiple timepoints. Most published studies used SI ratios of dentate nucleus to pons (DN/P) and/or globus pallidus to thalamus (GP/T) to assess SI changes. Typical measurement usually involves an operator-defined region of interest (ROI) placed on one side of DN or GP. Mean SI is drawn from a single slice of images. The measurement seems to be simple; nonetheless, it is tedious, and generally only the first and the last examinations of each patient are analyzed. The inability to analyze the SI changes after each GBCA exposure might dilute the scientific evidence and result in debatable or even misleading conclusions. A rigorous analysis of the consecutive serial assessment would be essential to predict the evolution of DN and GP hyperintensities. To help address this challenge, a fully automated software tool was developed and tested in this study. The tool also made it easy to exploit various approaches in obtaining SI ratios to demonstrate the urgent unmet need in the standardization of analysis.

2. Materials and Methods

2.1. Magnetic Resonance Imaging Protocol

Unenhanced T1w images were collected retrospectively from an institutional review board–approved study with the following inclusion criteria: (1) patients must have undergone at least eight GBCA-enhanced MR scans (i.e., at least eight sections per patient) at Cedars Sinai Medical Center in the 2008–2016 period (Optimark, MultiHance, Omniscan or Gadavist); (2) at least one basal ganglion and the posterior fossa were spared tumor or radiation effects; and (3) subjects had normal renal function, defined as a glomerular filtration rate of $>30$. Informed consent was waived given the retrospective nature of the study. T1w images (3D MPRAGE) were acquired at 1.5 T (Symphony; Siemens, Erlangen, Germany) and 3 T (Verio; Siemens) scanners (1.5 T: TR 1330 ms; TE 4.8 ms; TI 800 ms; flip angle 15 deg; section thickness 12.5 mm; matrix size: $256\times 192$; echo-train length 1 | 3 T: TR 2100 ms; TE 3.0 ms; TI 900 ms; flip angle 9 deg; section thickness 11 mm; matrix size $256\times 256$; echo train length 1). These protocols were used in majority of the studies and were referred to as the reference protocols.

2.2. Data Processing

The workflow for the automated ROI segmentation and SI parameter measurements was developed and implemented within the QMENTA cloud platform (QMENTA Medical Imaging Platform. QMENTA Inc., Boston, Massachusetts),¹⁴ which allows parallel automated processing with ample amounts of images. The workflow considered a series of studies (i.e., sessions) of each patient. From each session which included all sequences acquired, the unenhanced T1w image was automatically selected through a classifier using DICOM tags and a machine learning recognition system based on image intensities.¹⁵ If more than one unenhanced T1w series were available, the session with the best resolution was used. Images with big brain deformations were excluded as they can result in inaccuracies in the segmentation of brain tissues. The unenhanced T1w images of the most recent session were identified, and a skull-stripping algorithm¹⁶ was performed to detect the brain boundary. We performed the tissue segmentation using the Atropos tool from the Advanced Normalization Tools (ANTs) toolkit.¹⁷ First we obtained the gray matter parcellation by nonlinearly registering the dataset into the brain and propagating the Desikan–Killiany–Tourville atlas labels¹⁸ to segment the GP and thalamus. The same method (nonlinear registration of template brain and mask propagation) was conducted using the spatially unbiased atlas template of the cerebellum and brainstem for labeling the DN.¹⁹ Similarly, using the Montreal Neurological Institute template, a label of the pons was acquired. Hence, segmentations of all four ROIs (GP, thalamus, DN, and pons) of the most recent session were obtained for each patient. The next step was to propagate the segmentations to all sessions of a given patient. The T1w images of the most recent session were nonlinearly registered to the T1w images of each session using ANTs to correct for tumor-induced deformations.²⁰ The warp fields and affine transformation matrix generated from this process were used to propagate the segmentations to each session. Figure 1 shows the principles of the image processing workflow.

Fig. 1.

Diagram illustrating the fully automated workflow for volumetric assessment of signal changes on unenhanced T1w MR images after multiple gadolinium administrations per patient. KDE, kernel density estimator used to evaluate the modal SI of each ROI.

2.3. Modal Value Extraction

Instead of obtaining the mean SI of the unenhanced T1w images for each ROI, as the most commonly used method, we used the maximum value of the kernel density estimator (KDE), which is a way to estimate the mode (most frequently occurring value) of the SI distribution within the ROI.²¹ Using the ROI masks previously extracted, we collected the SIs of the voxels contained in each ROI and saved them as an array (from both right and left sides in the case of DN, GP, and thalamus). Then, we extracted the mode as the maximum of the KDEs^22,23 for each ROI of each session. Examples of probability density functions and the associated modal and mean values are shown in Fig. 2. As can be appreciated in the figure, the mode is a more robust metric than the mean in the case of skewed distributions [Figs. 2(b) and 2(c)].

Fig. 2.

Examples of histograms and KDE-curves of SIs in selected ROI. The $x$ axis represents the intensity values in the ROI, the $y$ axis is the frequency of that intensity. (a) There is minimal difference between modal and mean values if SI has a simple normal distribution. (b), (c) Other types of distributions such as multiple normal peaks or skewed distributions lead to greater differences.

2.4. Correction of Signal Intensity Ratio

The obtained modal values were corrected using simplified inversion-recovery spoiled gradient echo signal equation, ignoring T2* decay:

$$S=\rho ·(1-2·e^{-\mathrm{TI}/T1}+e^{-\mathrm{TR}/T1})\mathrm{Sin}(\alpha ),$$

where $\rho$ is a factor proportional to the equilibrium longitudinal magnetization and $\alpha$ is the flip angle. Since only the SI ratios were compared, both $\rho$ and $\mathrm{Sin}(\alpha )$ were canceled out in the ratios. SI ratio of GP and thalamus (GP/T) was also not affected by the correction because both tissues are within gray matter and presumed to have the same T1. T1w images acquired with the above-mentioned reference protocols did not require SI correction. However, for those with different TI or TR, SI ratios were normalized with respect to the reference protocols. Based on a pool of literatures, gray matter T1 of 1086 ms (1.5 T) and 1470 ms (3 T) were used for DN, and white matter T1 of 778 ms (1.5 T) and 1110 ms (3 T) were used for pons.²⁴

2.5. Statistical Analysis

We compared only SI ratios of DN to pons and GP to thalamus. Continuous variables were expressed as $\text{mean}\pm \text{standard}$ deviation (SD). Studies were divided into two groups based on the image acquisitions at 1.5 T or 3 T. Paired $t$-test was employed for tests within the same field strength. SI ratios of SI-corrected modal values from volumetric assessment were used as the standard of reference (SI-corrected-modal-volume, measurement 1) and compared against SI ratios of SI-uncorrected-modal-volume (measurement 2, uncorrected SI ratio), SI-corrected-mean-volume (measurement 3, mean SI ratio instead of modal), as well as SI-corrected-modal-single slice (measurement 4) approaches. The slice used for measurement 4 was selected from the centroid of the DN or GP. SI-corrected-modal-volume ratios of right and left sides of the brain were also compared using paired $t$-test. To evaluate the repeatability of the software, measurements were repeated in a subset of randomly selected patients ($N=8$ with a total of 199 studies). Repeatability was tested by Bland–Altman analysis. Intraclass correlation coefficient (ICC) and 95% confidence interval (CI) were used to estimate the agreement between measurements; $p<0.05$ (two-tailed) was considered statistically significant. Statistical analyses were performed using Microsoft Excel 2010 for Windows 10.

3. Results

Excluding the images with artefacts or brain necrosis in the ROI, 113 patients with a total of 2119 studies (1150 studies at 1.5 T and 969 studies at 3 T) were included in the final analysis. Figure 3 shows unenhanced MPRAGE images with ROI contours. 307 studies at 1.5 T (27%) with protocol deviation had TR ranged between 1210 and 1520 ms and $\mathrm{TI}=800\mathrm{ms}$, among which only three studies with different TI. At 3 T, there were only 71 studies (7%) with protocol deviation ($\mathrm{TR}=1720\text{-}2300\mathrm{ms}$) but with the TI at the reference value ($\mathrm{TI}=900\mathrm{ms}$). SI-corrected-modal-volume DN/P were $1.085\pm 0.048$ at 1.5 T and $0.979\pm 0.061$ at 3 T. SI-corrected-modal-volume GP/T were $1.084\pm 0.039$ at 1.5 T and $1.069\pm 0.042$ at 3 T. 1.5 T and 3 T results were analyzed separately.

Fig. 3.

Automatic (a) DN, (b) pons, (c) GP, and (d) thalamus segmentation. Images were from the same patient.

Modal DN/P ratios from volumetric assessment at 1.5 T failed to show a statistical difference with or without SI corrections ($p=0.71$). Notice that GP/T ratios remained the same with or without SI correction because the same T1 value was used in the signal equation for both GP and thalamus. All other paired t-tests demonstrated significant differences (measurement 2, 3, 4 compared to measurement 1, $p<0.001$). All correlations were also significant ($p<0.001$). Tables 1 and 2 summarize the results at 1.5 T and 3 T, respectively. Both DN/P and GP/T ratios were significantly different between right and left sides of brain (Table 3). The Bland–Altman plot for repeatability is shown in Fig. 4. Points of the plots were grouped and close to the mean, indicating a good agreement between the measurements ($-0.0007\pm 0.0037$, 95% CI, $-0.0012\text{-}-0.00018$, $\mathrm{ICC}=0.999$).

Table 1.

Results of paired $t$-test and Pearson correlation ($R$) at 1.5T.

	Measurement	DN/P $\mathrm{mean}\pm \mathrm{SD}$ ($p$)	$R$	GP/T $\mathrm{mean}\pm \mathrm{SD}$ ($p$)	$R$
1	SI corrected modal volume	$1.085\pm 0.048$	—	$1.084\pm 0.039$	—
2	SI uncorrected modal volume	$1.085\pm 0.047(0.71)$	0.983	—	—
3	SI corrected mean volume	$1.113\pm 0.055(<0.001)$	0.763	$1.08\pm 0.037(<0.001)$	0.711
4	SI corrected modal single slice	$1.083\pm 0.042(<0.001)$	0.917	$1.083\pm 0.042(<0.001)$	0.707

SI, signal intensity; SD, standard deviation; DN/P, SI ratio of DN to pons; GP/T, SI ratio of globus pallidus to thalamus. $P$ values for the paired $t$-test and Pearson correlations ($R$) of measurement 2, 3, and 4 to 1. Measurement 2 of GP/T is identical to measurement 1.

Table 2.

Results of paired $t$-test and Pearson correlation ($R$) at 3 T.

	Measurement	DN/P $\mathrm{mean}\pm \mathrm{SD}$ ($p$)	$R$	GP/T $\mathrm{mean}\pm \mathrm{SD}$ ($p$)	$R$
1	SI corrected modal volume	$0.979\pm 0.061$	—	$1.069\pm 0.042$	—
2	SI uncorrected modal volume	$0.983\pm 0.053(<0.001)$	0.893	—	—
3	SI corrected mean volume	$1.031\pm 0.071(<0.001)$	0.815	$1.064\pm 0.038(<0.001)$	0.603
4	SI corrected modal single slice	$0.974\pm 0.059(<0.001)$	0.931	$1.062\pm 0.047(<0.001)$	0.761

SI, signal intensity; SD, standard deviation; DN/P, SI ratio of DN to pons; GP/T, SI ratio of globus pallidus to thalamus. The $p$ values are for the paired $t$-test and Pearson correlations ($R$) of measurement 2, 3, and 4 to 1. Measurement 2 of GP/T was identical to measurement 1.

Table 3.

Results of paired $t$-test ($\mathrm{mean}\pm \mathrm{SD}$) and Pearson correlation ($R$) of SI-corrected modal ratios from volumetric assessments between right and left side of brain.

	1.5 T		3 T
	DN/P	GP/T	DN/P	GP/T
Right	$1.078\pm 0.044$	$1.091\pm 0.05$	$0.972\pm 0.059$	$1.064\pm 0.043$
Left	$1.091\pm 0.052$	$1.077\pm 0.039$	$0.985\pm 0.066$	$1.06\pm 0.051$
R	0.826	0.516	0.848	0.441

SI, signal intensity; SD, standard deviation; DN/P, SI ratio of DN to pons; GP/T, SI ratio of globus pallidus to thalamus.

All $p$-values in paired $t$-test between right and left side were significant ($<0.001$).

Fig. 4.

Bland–Altman plot of difference against the mean in two measurements. The solid line represents the mean of the two measurements ($-0.0007$). The dashed lines are the limits of agreement ($\mathrm{mean}\pm 1.96\times$ standard deviation).

4. Discussion

The primary goal of the current study was to develop an automatic workflow to streamline the analysis of SI ratios in the deep brain tissues on unenhanced T1w images in patients with repeat GBCA exposure. The method was deployed in 113 patients with 2119 studies. The novelty of our work includes segmentation of multiple ROIs in serial images, volumetric assessment of brain tissues, extraction of modal values, and SI corrections for protocol heterogeneity. We further demonstrated that SI ratios were divergent if obtained using different approaches.

While MRI-detected hyperintensities in the DN or GP are not a direct measurement of gadolinium presence, it is an in vivo assessment and remains to be the most feasible technique. However, MR SI is a relative parameter and is affected by several factors including system hardware, software, patients’ setting, and physiology. SI ratio could potentially compensate or normalize some factors, but the way it is measured varies even in single-center studies since ROI drawing is operator dependent and inter- or intrareader agreement is usually not assessed. Volumetric instead of single slice assessment can further circumvent the reader’s subjective selection for cross comparison. Using modal instead of mean SI can improve the sensitivity when the distribution of SI is skewed or multimodal. Furthermore, the SI ratio difference between right and left sides of brain nuclei as shown in our data reinforces the need of an automatic tool to standardize the measurement for the research in this field.

In the workshop about gadolinium retention convened by a group of experts, the statement “MRI techniques and analysis must be standardized so that results between centers are comparable” is among one of the high-priority items in the research roadmap.¹² In recognition of lack of consistent methodology across different institutions, standardized assessment of the brain MRI changes was further accentuated in European Gadolinium Retention Evaluation Consortium (GREC) Task Force position statement.¹³ We have found significant different SI ratios between 1.5 T and 3 T, and the discrepancy could not be improved by SI correction. This finding confirmed the GREC recommendation that data from different field strengths should be treated separately. Also, exclusion of data with variations of TR and TE above 15% was suggested. In our data, TR deviations were mostly within 15% of the reference TR. No significant difference in DN to pons with or without SI correction at 1.5T but SI ratio was significantly higher after SI correction at 3 T. The T1 values we employed for SI correction were selected from a pool of literature.²⁴ The ideal situation would be to use in-vivo measured T1 from each patient. Nonetheless, our results might imply that protocol deviations have stronger impact on SI ratio at 3 T than at 1.5 T.

5. Limitations

There are several limitations in our study. Despite the large number of studies available for analysis, images were all from a single center on a single vendor. While our algorithm does not presume any specific MRI platform, and therefore applicability to other vendors is expected, it should be confirmed by further studies. We analyzed only the SI ratios of DN to pons and GP to thalamus, as those are the most commonly assessed areas. The reference protocols at 1.5 T and 3 T were chosen based on the most occurrent TR/TI in the study to minimize the deliberating SI correction. If different reference protocols were used, the absolute SI ratios would be changed. We did not perform any post processing on the 3D MPRAGE images. Flow or MR intensity inhomogeneity artefacts might have some influence on the results as suggested by Saake et al.²⁵

6. Conclusions

We have developed a fully automatic and highly reproducible method to segment brain tissue volumetrically from 3D brain MR images. The workflow enables high-throughput analyses of SI ratios in the deep brain nuclei in patients with repeated GBCA exposure. In future studies on gadolinium brain retention, the use of our tool could be an asset to expedite the research in this field.

Biographies

Chia-Ying Liu received her PhD in Physics from the University of California, Riverside. She was previously employed at Bayer Radiology, National Institutes of Health, and Johns Hopkins University and is currently a senior clinical scientist at Canon medical systems.

Marc Ramos graduated in telecommunications engineering from the Universitat Politecnica Catalunya in 2013. He started an internship at QMENTA, entering the acceleration program as one of the top talented engineering students. Afterward, he fully joined the QMENTA team and has been working for more than 7 years as a scientific Python developer, focusing on architecture and process automation.

David Moreno-Dominguez holds a PhD in medical image analysis from the Max Planck Institute and has 10 years of experience in the field of MRI image processing. He is currently affiliated with Boehringer Ingelheim as a clinical innovations system analyst. At the time of this study, he worked as a medical imaging team lead at QMENTA Inc.

Vesna Prčkovska is the cofounder and chief operating officer (COO) of QMENTA. She graduated in electrical engineering at Saints Cyril and Methodius University, Macedonia. She was awarded the prestigious Marie Curie postdoctoral fellowship at Harvard Medical School, USA. As part of this grant, she also joined the Department of Radiology at Massachusetts General Hospital, USA. In her role as COO, she leads the implementation of the company’s strategy and the R&D program.

Paulo Rodrigues graduated from the Software Engineering Department at the University of Minho, Portugal, and received his PhD in neuroimaging from Eindhoven University of Technology, Netherlands. He is a cofounder and chief technology officer at QMENTA. He has published numerous scientific papers and patents. In 2013, he was selected for the MIT Technology Review’s list of top innovators under 35.

Markus Blank obtained his PhD and Master of Science degrees from University of Erlangen, Germany. He is heading the Infrastructure team at Bayer Radiology with 20+ years’ experience in medical image visualization, a PhD in human biology, master’s in computer science.

Franklin G. Moser, FACR, is a professor of imaging at the S. Mark Taper Imaging Center at Cedars Sinai Medical Center. He serves as a vice-chairman for radiology research and director of clinical and interventional neuroradiology. He studied at Yale College and the McGill University College of Medicine. He did his residency at Mt. Sinai Medical Center, followed by a fellowship at Columbia University’s College of Physicians and Surgeons. He also has received a Master of Medical Management degree from the Marshall School of Business at the University of Southern California.

Jacob Agris is a VA and PA state licensed neuroradiologist. He graduated from Baylor College of Medicine and did his residency at Wayne State University School of Medicine, University of Texas Health Science Center – Houston, University of Vermont–Fletcher Allen Health Care, and Boston University Medical Center. He was the director of global clinical development at Bayer Health Care at the time of this research and currently serves as the chief medical and innovation officer at ConvaTec.

Disclosures

The authors have no relevant financial interests.