Modified Diffusion Tensor Image Processing Pipeline for Archived Studies of Patients With Leukoencephalopathy

Abstract

Background:

In archived diffusion tensor imaging (DTI) studies, a reversed-phase encoding (PE) scan required to correct the distortion in single-shot echo-planar imaging (EPI) may not have been acquired. Furthermore, DTI tractography is adversely affected by incorrect white matter segmentation due to leukoencephalopathy (LE). All these issues need to be addressed.

Purpose:

To propose and evaluate a modified DTI processing pipeline with DIstortion COrrection using pseudo T₂-weighted images (DICOT) to overcome limitations in existing acquisition protocols.

Study Type:

Retrospective feasibility.

Subjects:

DICOT was assessed in simulated data and 84 acute lymphoblastic leukemia (ALL) patients with reversed PE acquired. The pipeline was then tested in 522 scans from 261 ALL patients without a reversed PE acquired.

Field Strength/Sequence:

A 3 T; diffusion-weighted EPI; 3D magnetization prepared rapid acquisition gradient echo (MPRAGE).

Statistical Tests:

Repeated measures analysis of variance and Tukey post hoc tests were performed to compare fractional anisotropy (FA) values obtained by different methods.

Assessment:

FA and corresponding absolute error maps were obtained using TOPUP, DICOT, INVERSION (Inverse contrast Normalization for VERy Simple registratION) and NO CORR (no correction). Each method was assessed by comparing to TOPUP. The pipeline in the ALL patients was evaluated based on the failure rate of the distortion correction using the global correlation values.

Results:

Using DICOT reduced the mean absolute errors by an average of 32% in FA in simulation datasets. In 84 patients, the error reductions were approximately 15% in FA with DICOT, while it was 5% with INVERSION. No significant differences between the TOPUP and DICOT were observed in FA with P = 0.090/0.894(AP/PA). Only 15 of 516 examinations requiring any additional manual intervention.

Conclusion:

This modified pipeline produced better results than the INVERSION. Furthermore, robust performance was demonstrated in archived patient scans acquired without an inverse PE necessary for TOPUP correction.

Evidence Level:

3

Technical Efficacy:

Stage 2

Archived studies contain diffusion-weighted images (DWIs) that were collected previously for research or clinical purposes. Investigators frequently want to reuse these existing/archived images for new research goals. The benefit of this is to achieve the research goal quickly and cost effectively. Previous retrospective studies have utilized existing DWI data sets to investigate late radiation-induced damages after focal radiotherapy,¹ hemispheric structural network changes as a function of location and size of glioma,² and even the relationship between reading performance and white matter (WM) alteration and reorganization in neurosurgical patients.³ In addition, as more and more archived data become publicly available, research can be conducted efficiently if the data can be processed in an appropriate way. However, one of the challenges to using data from an archived study is that the acquisition protocols might not be optimized for the novel study goal. Diffusion tensor imaging (DTI) studies almost exclusively use single-shot echo-planar imaging (EPI) or its variants such as readout segmentation of long variable echo-trains (RESOLVE).⁴ EPI is sensitive to static field inhomogeneities that create geometric image distortion, primarily along the phase encoding (PE) direction.⁵ Such artifacts are particularly prominent in the inferior temporal and frontal lobes because of magnetic field microscopic gradients/variations at the interface of tissues with different magnetic susceptibility.⁵

Different techniques have been proposed to correct this image distortion. Among these methods, the field map-based approach is a popular method, in which the B₀ field inhomogeneity (offset) is obtained from phase images acquired at different echo times.^6,7 A reversed-PE gradient approach,^8,9 implemented in FMRIB Software Library FSL,¹⁰ called TOPUP, has been widely used, including in the Human Connectome Project¹¹ and MRtrix3.¹² In this approach, images are acquired twice with opposite phase encoding polarities, resulting in pairs of images with distortions in opposite directions. From these pairs, the susceptibility-induced off-resonance field is estimated, and the distortion is corrected.

Because additional acquisitions are needed for the B₀ field inhomogeneity distortions, they cannot be performed in archived studies in which these extra acquisitions were not acquired. This shortcoming motivated researchers to develop a method to achieve a degree of distortion correction using only these archived images. Registration-based (RB) methods correct the distortion by aligning distorted DWI images to undistorted 3D structural images such as T₁-weighed (T1w) or T₂-weighted (T2w) images.^13–17 A T2w image acquired with the same echo time (TE) as the b0 image is ideal for RB distortion correction when used with appropriately modified histogram match methods.¹³ However, 3D T2w images are not always acquired in clinical protocols and are very unlikely to be acquired with a TE that matches that of the b0 images. The benefit of using T1w images is that they are more likely to be collected in clinical research scans. However, the contrast of the diffusion images and that of the T1w images are very different. Inverse contrast Normalization for VERy Simple registratION (INVERSION)¹⁷ inverts the contrast relationship between T1w and T2w images to make these two images similar enough for registration. However, it is relatively complicated and requires considerable mathematic analysis and calculations. In addition, the approach to globally match the histograms between the two images cannot always guarantee a good local voxel by voxel match. Hence, we proposed a new way to convert a T1w images to a pseudo T2w image by following the physics principles of MR signal formation. The distortion was corrected by using the pseudo T2w image, and we call the method DIstortion COrrection using a pseudo T2w image (DICOT). Recently, deep neural network methods have also been applied to correct the distortion in EPI images¹⁸ but are beyond the scope of this project.

The goal of distortion correction is to achieve a distortion-free parameter map or tractography. The diffusion parameter map achieved with a new method should be as close as possible to the gold standard or the reference method. Direct metrics, such as fractional anisotropy (FA) or mean diffusivity (MD) error maps relative to the gold standard/reference, are preferable for evaluating a new method. However, this has only been done in a limited number of simulation data and human subject data.^18–20

Leukoencephalopathy (LE), expressed as T₂ hyperintensities in the WM, is sometimes seen in children treated for cancer and is often attributed to chemotherapy, radiation, or other drugs.²¹ These regions appear as hypointense WM on a T1w image and can cause incorrect WM segmentation.²² LE can be manually or semi-automatically segmented²³ and then combined with a WM mask to alleviate the problem. However, it is time consuming to perform manual segmentation. Semi-automated methods can be difficult to implement, and the computation cost can be high.²³

The aim of this study is to develop and test a new DTI processing pipeline for archived DTI studies. The proposed DICOT method was directly compared with TOPUP and INVERSION in two sets simulation data and 84 patients by using direct metrics (FA and MD maps). The pipeline was also tested in 261 patients with examinations that did not contain the reversed PE scan.

Methods

Pseudo T2w Images and Distortion Correction Using DICOT

Because T1w and b0 images have similar anatomical structures, there should exist a relationship between these two sets of images. This relationship can be nominally found with some assumptions based on examining the signal formation equations. The signal of T1w images can be represented as²⁴

$$S1=p_1-p_2\exp \left(-\frac{TI}{T_1}\right),$$ (1)

where TI is the inversion time, assuming a magnetization prepared rapid acquisition gradient echo (MPRAGE) acquisition is used.²⁵ T₁ is the longitudinal relaxation time. If the image is acquired with other T1w sequences, such as a gradient echo sequence, then the equation may need to be modified accordingly. p₁ and p₂ are the constants depending on the measurement parameters (see eq. 7 in the literature²⁴). The signal of T2w images is given by²⁶

$$S2=S{2}_0\exp \left(-\frac{TE}{T_2}\right)$$ (2)

where S2₀ is the S2 value at TE = 0 or proton density. By combining Eqs. 1 and 2, pseudo T2w images can be generated from T1w images as follows:

1. Mean signal values of WM and gray matter (GM) were first measured for the T1w images and p₁ and p₂ can then be solved from Eq. 1 using T₁ values of WM and GM.

2. A pseudo T₁ map can be calculated by using the following equation:

   $$T_1=-\frac{TI}{\ln \left[\frac{p_1-S1}{p_2}\right]}$$ (3)
3. Convert the pseudo T₁ map to a pseudo T₂ map by using the following equation:

   $$T_2=26+0.04T_1$$ (4)

   assuming the linear relationship between T₁ and T₂ if only GM and WM are considered.
4. Convert the pseudo T₁ map to a proton density map by using the following equation:

   $$S{2}_0=46.24+0.011T_1$$ (5)

   assuming a linear relationship between T₁ and S2₀.

5. A pseudo T2w image can be obtained by using Eq. 2 and then smoothed by using a Gaussian kernel with an full-width half maximum (FWHM) value of 2.35 mm.

The purpose of the conversion is to generate a pseudo T2w image for coregistration, instead of a real T2w image, so we can reasonably make more assumptions in the conversion. Parameters p₁ and p₂ are assumed to be voxel independent. Only WM and GM were considered in the conversion because cerebrospinal fluid (CSF) signal is well separated from the WM and GM. The T₁ values were 1084/1820 msec (WM/GM) and T₂ = 70/100 msec (WM/GM).²⁷ The proton density ratio of WM and GM was 70/80.²⁸ The script to convert a T1w image to a pseudo T2w image has been made publicly available (https://github.com/ruitians/distcorr_pipeline).

LE Patching Method

To include LE in the WM segmentation, we patched the hypointense LE region to match normal-appearing WM by implementing the following steps:

1. Generate a segmentation and parcellate labeled image (aparc+aseg.mgz) from a T1w image by using FreeSurfer (http://surfer.nmr.mgh.harvard.edu/).

2. Calculate the mean WM value of the T1w image, S_{w_mean}, over a WM mask obtained from aparc+aseg.mgz file.

3. In the WM region of the T1w image, identify signals below S_{w_mean} voxels and patch them to S_{w_mean} to eliminate hypointense WM regions (LE regions).

4. The patched T1w image is then used with MRtrix3 to achieve a correct five tissue type (i.e. cortical grey matter, sub-cortical grey matter, WM, CSF, and pathological tissue [5TT]) segmentation.¹²

DTI Processing

DTI tractography was processed by using MRtrix3 (version:3.0_RC3, www.mrtrix.org /),¹² FSL (version 6.0, www.fmrib.ox.ac.uk/),¹⁰ and FreeSurfer (version: 5.3.0, www.surfer.nmr.mgh.harvard.edu).²⁹ The pipeline is demonstrated in Fig. 1. After DWI images are preprocessed via denoising, removing Gibbs ringing artifact, and eddy current correction, the b0 image is extracted from the DWI image. Then the brain extraction is performed on T1w images by using FreeSurfer before creating a pseudo T2w image. The extraction using Freesurfer failed in one patient, and ROBEX (https://www.nitrc.org/projects/robex)³⁰ was used for that patient. The pseudo T2w image is then linearly registered to the b0 image using the FSL flirt coregistration tool. In the diffusion space, the distortion of the b0 image is corrected by nonlinearly registering it to the pseudo T2w image along the phase encoding direction. A symmetric normalization (SyN) transformation model in advanced normalization tools (ANTs)³¹ was chosen for the nonlinear coregistration. The correction is then applied to all DWI volumes.

FIGURE 1:

Flow chart of the modified DTI processing pipeline. A 5TT (five tissue type) image is created from a patched T1w image as described in the methods section. Then the 5TT image is transformed to the diffusion-weighted image space. A pseudo T2w image, converted from the T1w image, is rigid body coregistered to the b0 image extracted from the preprocessed DWI image. The distortion of the DWI images is then corrected by a nonlinear registration. The DTI tractography can then be generated from fiber orientation distributions (FODs) with the aid of anatomically constrained tractography (ACT).

Other parameters used in this registration are a Hamming windowed sinc interpolation, mutual information metric for rigid and affine transformation, and correlation coefficient for deformable transformation. The DTI tractography can finally be generated from fiber orientation distribution functions with the aid of anatomically constrained tractography (ACT).³² This processing pipeline has been made publicly available (https://github.com/ruitians/distcorr_pipeline).

For comparison, patient data were also processed by using the INVERSION method, which is integrated into the DTI processing pipeline in BrainSuite (http://brainsuite.org/). DWI images were preprocessed in the same way as they were for the DICOT method. The BrainSuite processing pipeline was performed on the preprocessed DWI images with masks of the b0 and T1w images. All FA and MD were obtained with the corrected DWI images by using MRtrix3.

Simulation Data

The proposed method was first tested in two simulation data sets, signal-to-noise ratio (SNR) = 20 (A) and SNR = 40 (B), of one virtual human subject with noise and susceptibility distortion downloaded from https://www.nitrc.org/projects/diffusionsim/.¹⁹ In the simulation data, susceptibility distortion was introduced into the DWI images by considering susceptibility and PE direction. The parameters used in the simulation are TE = 109 msec, b = 5, 1000 s/mm² with eight directions; voxel size: 2 × 2 × 2 mm³, matrix size: 90 × 106 × 68; two averages. In each dataset, the images were simulated with both blipup and blipdown PE directions, resulting in two subsets of data (blipup and blipdown). FA and MD maps were obtained with the DICOT, TOPUP, no distortion correction, and ground truth data.

The correction effectiveness was evaluated by comparing FA maps to the ground truth FA map by using an absolute error map and mean absolute error values. The error map was generated by calculating the absolute difference between the estimated FA and the ground truth FA on a voxel-by-voxel basis. MD was also examined in the same way as FA. The mean absolute error was obtained by calculating the mean of the absolute difference over a brain mask eroded by two voxels.¹⁹

Patient Datasets (With Reversed PE Acquisitions)

All measurements were performed on a 3.0 T Siemens whole-body system (TrioTim, Siemens Medical Systems, Iselin, NJ) with an 8 element HeadMatrix coil, a 20 or 64 element HeadNeck coil. All patient data used in this study were acquired with imaging protocols approved by the local Institutional Review Board, and written informed consent was obtained from all the patient or guardian of each patient as appropriate. The TOPUP method was chosen as a reference method to validate the proposed DICOT method in 84 patient scans, including three patients with LE, acquired with both AP (anterior to posterior) and PA (posterior to anterior) phase encoding directions. The acquisition parameters were as follows: single average; b = 0, 1500 s/mm²; 64 diffusion gradient directions; spatial resolution = 1.8 × 1.8 × 1.8 mm³; TR/TE = 4000/79 msec. Two subsets of data were obtained for each patient: the images acquired with AP phase encoding and those acquired with PA phase encoding. As part of standard preprocessing practices, DWI images were reviewed by experienced signal processing technicians for excessive motion and removed from further processing.

For each subset of data, FA maps were obtained after TOPUP, DICOT, INVERSION, or no distortion correction. Mean FA was calculated for each subject over a brain mask eroded by two voxels as described previously.¹⁹ The average and standard deviation of the mean FAs over all subjects were calculated and used to compare the different methods. Absolute error maps for each corrected or uncorrected image were also calculated relative to the TOPUP correction, which served as the gold standard. The reduction of the absolute error was used to score the improvement, defined as

$$\text{IMP}=\frac{\text{RB}-\text{NO\_CORR}}{\text{NO\_CORR}}$$ (6)

where RB was either DICOT or INVERSION. If IMP < −10%, then we considered the method to have failed to perform distortion correction. MD was also examined in the same way as FA.

Patient Datasets (Without Reversed PE Acquisitions)

The proposed modified pipeline was also validated in 261 patients treated for acute lymphoblastic leukemia (ALL) who underwent imaging twice, resulting in a total of 522 scans. Three patients were excluded due to insufficient brain coverage on DWI images. Excluding both scans from these three patients resulted in a total of 516 examinations, including 87 with LE.

One patient’s imaging was re-processed using a better cropped brain image to ensure a good 5TT image. For each patient, the first scan was acquired at reinduction I, approximately 6 months into therapy, and the second scan was acquired approximately 2 years later at completion of therapy. The DTI streamlines integrity was examined automatically using the warning message generated in MRtrix3 as part of the process outputs. A warning of lack of streamlines in parcellation was also checked in the MRtrix3 processing output. Quality of the distortion correction was examined using global correlation (GC) between pseudo T2w images and distortion corrected b0 images using ANTs. The mean and standard deviation were calculated, and a threshold of 2 standard deviations below the mean was used to identify outliers where the correction may not be optimal and needed further manual evaluation.

LE was identified based on radiological reports of the clinical examinations. After LE patching, all LE cases were visually inspected by three observers (R.S., J.O.G, and W.E.R. with 18, 26, and 30 years of MRI research experience) to ensure any holes in the WM mask were filled by the patching algorithm. The acquisition parameters were 12 directions, 4 averages, and a spatial resolution of 1.5 × 1.5 × 3.0 mm³ (TR/TE = 6500/120 msec; b = 0.700 s/mm²). 3D T1w images were acquired with an MPRAGE sequence (TR/TE/TI = 1560/2.75/900 msec) having 1 mm isotropic resolution.

Statistical Analysis

One-way repeated measures analysis of variance (ANOVA) and Tukey HSD post hoc tests were performed by using R (version 4.0.3) on the mean FA values for the 84 patient scans acquired with both AP and PA phase encoding directions. The P values were adjusted for multiple comparison. All studied cases, including the failed cases of DICOT and INVERSION, were included in the statistical analysis and comparison. The significance threshold was set to P < 0.05.

Results

Simulation Dataset

Figure 2 shows FA maps estimated from the ground truth and the corresponding absolute error maps of each method for simulation dataset B. This figure demonstrates that the TOPUP method has the smallest mean absolute error about 0.031. However, the improvement by the DICOT method can also be clearly appreciated with a mean error of 0.039, which was much smaller than that of no distortion correction with a mean error of 0.063.

FIGURE 2:

Comparison of different distortion methods to the ground truth in the simulation data. FA map for the ground truth before addition of distortion is shown in the left column. Absolute error maps, defined as absolute value of estimated FA maps minus ground truth FA on a voxel-by-voxel basis, are also shown for FSL TOPUP, DICOT, and no distortion correction (NO_CORR) methods. The scale bars at the bottom are the scales for FA and error maps (both unitless).

Table 1 lists the mean and standard deviation of the FA values over a brain mask along with the mean absolute error of each method for two data sets (A and B) with both blipup and blipdown; Although the mean FA values of the different methods were similar, the mean absolute errors were not zero for all methods. The smallest errors were observed in the TOPUP method. However, the errors produced by the DICOT method were also lower than no correction in all cases and are comparable to those produced by the TOPUP method for the blipdown in dataset B. It can also be observed that the distortion correction performed better in higher SNR for both TOPUP and DICOT.

TABLE 1.

TOPUP, DICOT, and No Correction (NO_CORR) Methods With Ground Truth (GT) in Two Sets of Simulation Data (A and B) With Blipup (bup) and Blipdown (bdown) Phase Encoding Direction

Data	FA	GT	TOPUP	DICOT	No_Corr
A/bup	Mean ± SD	0.23 ± 0.17	0.23 ± 0.17	0.22 ± 0.17	0.23 ± 0.17
	Abs. Error		0.046	0.055	0.074
	IMP(%)		38	26
A/bdown	Mean ± SD	0.223 ± 0.173	0.24 ± 0.16	0.23 ± 0.17	0.23 ± 0.17
	Abs. Error		0.046	0.054	0.073
	IMP		37	26
B/bup	Mean ± SD	0.23 ± 0.17	0.23 ± 0.17	0.23 ± 0.17	0.23 ± 0.17
	Abs. Error		0.026	0.040	0.063
	IMP		59	37
B/bdown	Mean ± SD	0.22 ± 0.17	0.24 ± 0.16	0.23 ± 0.17	0.23 ± 0.17
	Abs. Error		0.036	0.038	0.062
	IMP(%)		42	39

FA = fractional anisotropy; Mean and SD = the mean and standard deviation of the FA values; Abs. Error = mean absolute error defined as absolute difference between two related images, and IMP is improvement defined in Eq. (6).

Patient Datasets (With Reversed PE Acquisitions)

Figure 3 shows T1w (a), pseudo T2w (b), and original b0 (c) images. The contrast of the pseudo T2w (GC = 0.794) is observed to be more similar to that of the original b0 image than was the T1w (GC = 0.230). The WM region on both pseudo T2w and b0 images are dark, but it is bright in the T1w image. The CSF shows bright in both pseudo T2w and b0 but dark in the T1w image. In the supplement section, pseudo proton density and T₁ and T₂ maps are also shown (Supporting Information Fig. S1). It took 47 seconds to convert a T1w image with a matrix size of 190 × 200 × 170 on a Linux workstation with 64 GB of memory.

FIGURE 3:

Conversion of a T1w image to a pseudo T2w image for a typical patient. T1w (a), pseudo T2w (b), and original b0 (c) images.

The comparison of different methods to TOPUP for a representative subject is shown (Fig. 4). The left column contains the FA maps obtained by using the TOPUP method for AP and PA images. The mean absolute errors relative to the TOPUP correction are greatest with NO_CORR (0.067/0.063; AP/PA), improved with the INVERSION method (0.056/0.052; AP/PA), and are best for the DICOT method (0.048/0.042; AP/PA).

FIGURE 4:

Evaluation of DICOT method by comparing the DICOT, INVERSION and no distortion correction (NO_CORR) FA maps to TOPUP FA maps. The left column contains the FA maps obtained by using the TOPUP from AP and PA acquisitions. The second column from left contains absolute error maps made by using the DICOT, the third column from left contains absolute error maps made by using INVERSION, and the right column contains the absolute error maps without correction. The scale bars at the bottom are the scales for FA and absolute error maps (both unitless).

The results from the cohort of 84 patients are summarized in Table 2. The INVERSION method demonstrated improvements of 2.8% ± 20.2% in AP acquisition and 7.5%± 14.6% in PA acquisition. DICOT performed better, with improvements of around 14.5% ± 10.7% in AP and 15.8% ± 10.8% in PA. The INVERSION method failed to substantially improve distortion correction (<−10% IMP) for 25 of the AP acquisitions and 10 of the PA acquisitions. The DICOT approach also failed to correct the distortion in a few cases. For the AP acquisitions, one acquisition failed with a large negative IMP (−10.7%), and six other cases of smaller negative improvements from −0.3% to −3.6% were observed. For the PA acquisitions, there was also one failed case with a large negative IMP (−11.4%) and five other cases of minor negative improvement from −1.3% to −4.4%. The two outliers were not from the same patient, and the reversed PE acquisition for each was successful. Inspection of these two cases revealed unusually large distortions in the corresponding PE direction, and nonlinear registrations between b0 and pseudo T2w failed in these two cases. One failed case is shown in Fig. S2 of the in the Supplemental Materia. The comparison of MD values between different methods was summarized in Table S2 of the Supplemental Material. Like the FA results, the mean absolute error maps using MD in NO CORR was reduced by about 23% with DICOT, compared to about 13% with INVERSION.

TABLE 2.

Calculated FA Values From DICOT, INVERSION or No Correction (NO_CORR) Methods With AP or PA Phase Encoding (PE) Directions Relative to Those From The TOPUP Method Evaluated in 84 Patients

PE	FA	TOPUP	DICOT	INVERSION	NO_CORR
AP	Mean ± SD	0.22 ± 0.013	0.22 ± 0.013	0.22 ± 0.01	0.23 ± 0.013
	Abs. Error		0.055 ± 0.008	0.062 ± 0.010	0.065 ± 0.008
	IMP(%)		14.5 ± 10.7	2.8 ± 20.2
PA	Mean ± SD	0.22 ± 0.012	0.22 ± 0.013	0.22 ± 0.013	0.23 ± 0.012
	Abs. Error		0.051 ± 0.007	0.056 ± 0.008	0.061 ± 0.007
	IMP(%)		15.8 ± 10.8	7.5 ± 14.6

Mean and SD are the mean and standard deviation of the FAs across the population of the patients.

PE = phase encoding direction; AP = anterior to posterior; PA = posterior to anterior; Abs. Error = mean absolute error defined as absolute difference between two related images, and IMP: improvement defined in Eq. 6.

Results of the ANOVA statistical analyses of the correction methods are documented in Table 3. Significant differences were found among TOPUP, DICOT, INVERSION and NO_CORR with adjusted P < 0.001, except between TOPUP and DICOT with adjusted P values of 0.090 (AP) and 0.894 (PA).

TABLE 3.

Repeated Measure ANOVA Test of Mean FA Values With Distortion Correction by TOPUP, DICOT, INVERSION, or No Correction (NO_CORR) in 84 Patients

PE	Reference	DICOT	INVERSION	NO_CORR
PA	TOPUP	0.849	<0.001*	<0.001*
	DICOT		<0.001*	<0.001*
	INVERSION			<0.001*
AP	TOPUP	0.090	<0.001*	<0.001*
	DICOT		<0.001*	<0.001*
	INVERSION			<0.001*

PE = phase encoding direction; AP = anterior to posterior; PA = posterior to anterior;

The adjusted P value is reported and marked accordingly if the difference between the two methods is statistically significant (* P < 0.05) or not significant (no mark).

Patient Datasets (Without Reversed PE Acquisitions)

The mean and standard deviation of GC between the pseudo T2w and distortion corrected b0 were 0.803 and 0.048, respectively, so the threshold was set to 0.706 (mean – 2 * STD). Fifteen patient scans had GC values less than the threshold scans and required manual inspection and intervention. Distorted (original image, Fig. 5(a)) and undistorted (corrected image, Fig. 5(b)) b0 images are shown in Fig. 5 for a representative patient scan. The distortion on the frontal pole region is effectively corrected by comparing Fig. 5(a,b).

FIGURE 5:

Performing distortion correction by using the proposed method. The distortion can be seen in the original b0 image (a) as indicated by the arrows. After correction, the distortion is much improved (b). The images are overlaid with a white and GM boundary obtained from the T1w image (c).

Without distortion correction, based on the MRtrix log files, nine patient scans had one or more cortical regions that contained no successful streamlines because of imperfect coregistration between the T1w and DWI images. After the DICOT correction, only one patient scan had a single cortical region with no successful streamlines. DTI tractography overlaid on T1w images is shown in Fig. 6(a) (without distortion correction) and Fig. 6(b) (with distortion correction). Without distortion correction, no streamlines were assigned to the parcellation labeled L_10v. Streamlines were successfully assigned to this area with the help of the proposed distortion correction technique.

FIGURE 6:

DTI tractography overlaid on a T1w images using a distorted b0 image (a) and a DICOT distortion corrected b0 image (b). Using this distortion correction method, the streamlines are greatly improved within the distortion region (red circled).

LE patching was applied to 87 of the patient scans in the validation cohort (16.9%). As an example, Fig. 7 shows T1w (a, d), WM segmentation (b, e), and DTI tractography (c, f) with and without LE patching. The holes observed in the periventricular WM in Fig. 7(b,c) (circled areas) were effectively filled by the LE patching algorithm as shown in Fig. 7(e,f). As can also be seen in this representative case, tractography was able to be performed through this region after patching. Upon inspection by the three observers, all cases with severe LE identified by the radiological reports had been successfully corrected by the patching algorithm automatically.

FIGURE 7:

Hypointense leukoencephalopathy (LE) (circled) is observed in a T1w image (a) and caused holes in the WM mask (b) and DTI tractography (c). The hypointense LE is patched to the mean value of WM as shown on the patched T1w (d). This method corrected the false WM segmentation (e) and achieved improved tracks (f).

Discussion

The proposed modified pipeline includes image distortion correction and LE patching for improved WM segmentation. Pseudo T2w images generated from T1w images ensures robust deformable transformation between distorted b0 images and distortion-free T1w images and eliminates the need for additional T2w images. In addition, no further registration from T2w to diffusion is required because pseudo T2w and T1w are in the same space. We used an LE patching method to achieve improved WM segmentation. DICOT demonstrated better performance to correct image distortion than INVERSION. Compared to other LE segmentation methods,²³ our method was easy to implement with low computation cost, taking advantage of pre-existing FreeSurfer parcellations needed for connectome analysis. DICOT was tested in two sets of simulation data and 84 patients by comparing FA and MD maps. Retrospective application of the DICOT pipeline was successfully tested for automated processing of a large cohort of scans in patients treated for ALL.

Although significant differences were observed between TOPUP and NO_CORR in mean FA due to the distortion, no significant difference could be found between TOPUP and DICOT. The statistical analysis only compares the mean values, the negative and positive errors could be canceled out in the comparison. Consequently, the absolute errors were also used to assess the correction effectiveness. Among the two RB methods, DICOT outperformed INVERSION in correcting distortion by reducing the absolute errors by 15% (FA) compared to only 5.2% (FA) for INVERSION.

RB distortion correction is usually performed between T2w images and b0 images^13,14,16 because they are both T2w. We chose T1w images as reference images mainly because they are more frequently acquired in many brain research and clinical protocols. Although both T2w and b0 images are T₂ weighted, their contrasts were not exactly matched because the TEs used in the sequences were different (TE is around 300 msec for T2w images and around 100 msec for b0 images). In addition, brain image segmentation, such as the one used to produce the 5TT image in MRtrix3, is commonly generated from a T1w image. No further registration from the T2w image to the diffusion image is required because the pseudo T2w and T1w images are in the same space.

In converting a T1w to a pseudo T2w image, only WM and GM were considered in the conversion. T₁ of CSF was around 4000 msec,³³ and that of GM and WM were below 1500 msec at 3 T.²⁷ T₂ of CSF was about 500 msec³⁴ and that of GM and WM was less than 100 msec at 3 T.²⁷ Because the T₁ and T₂ values of CSF were far from those of GM and WM, CSF was not considered in the conversion. The CSF was still bright and well separated in the resulting pseudo T2w images.

The conversion was built on the assumption that WM T₁ < GM T₁ < CSF T₁ and WM T₂ < GM T₂ < CSF T₂. If these relationships between different tissues are valid, then the contrast of pseudo T2w should be similar to that of an actual T2w image regardless of age or sex based on Eqs. 1–5. However, because many assumptions must be made for the conversion, the pseudo T2w image is not identical to an acquired T2w image. The pseudo T2w image is for registration only. Therefore, the T₁ and T₂ values were not required to be accurate for the conversion, and standard mean T₁ and T₂ values obtained from literature²⁷ were used in this study.

Limitations

The proposed DICOT method may not always perform as expected, and extra caution should be taken when using this method. The method failed in the distortion correction of one of the 84 subjects, indicating a degradation of alignment. However, image distortion was substantially corrected in more than 90% of subjects compared to the TOPUP method. Another limitation of the method is that a robust and accurate registration tool is required. ANTs was used in the current study because of its superb performance in previous tests.³⁵ It required careful selection of the parameters to ensure sufficient performance. Although ANTs has been proven to be robust for performing nonlinear registration between images,³⁶ room for improvement exists. Brain extraction from both T1w images and b0 images is an essential component of the processing pipeline. The quality of these extractions may affect the registration, and subsequently, the performance of the DICOT method.

As demonstrated in the simulation and human studies, the DICOT method offers an improvement over no distortion correction. However, the performance is not as good as that of the TOPUP method. Consequently, the proposed pipeline is most useful for archived studies where no other advanced distortion correction methods, such as TOPUP, can be performed due to lack of additional required images. If a new research protocol is designed and developed, one should consider collecting extra images for the use of other more accurate methods, such as the TOPUP method.9 Another limitation is only three LE patients were included in the first test cohort. However, 87 patients with LE were tested in the second cohort and satisfactory WM segmentation was achieved for all patients.

Conclusions

We have proposed a modified DTI processing pipeline (DICOT) for archived studies of images possibly containing LE by adding distortion correction and LE patching. The pipeline can be easily integrated into standard DTI processing pipelines in MRtrix3 and was highly successful in processing archived studies of patients treated for ALL.