Quality Control Considerations for the Effective Integration of Neuroimaging Data

Abstract

Ensuring image quality control (QC) for data acquired in a multi-modality context offers substantial advantages for both multi-level and multi-point analysis. Although a variety of neuroimage analysis algorithms exist, the tasks of multimodal neuroimaging data QC and integration remain challenging because image quality can be affected by numerous factors. Here, we discuss the challenges of the QC and integration of neuroimaging data and provide two examples of often-neglected and potentially under-appreciated problems related to the QC of diffusion tensor imaging (DTI) data and to their integration with other modalities. Specifically, we illustrate the challenges of (1) DTI/MRI co-registration and (2) scanner vibration artifacts, both being representative examples of difficulties involving both data QC and its integration. Additionally, we highlight the need for automatic methods which can address neuroimaging data QC which allows for its successful integration.

1 Introduction

The heterogeneity of neuroimaging data can render their integration challenging, particularly in the case of diffusion tensor imaging (DTI), where sophisticated post-acquisition processing can be required for such data to become amenable to integration with other imaging data types, such as T ₁ -weighted magnetic resonance imaging (MRI), functional MRI (fMRI) and positron emission tomography (PET). Figure 1 illustrates several common neuroimaging artifacts affecting study quality which (a) indicate that data would be a challenge to integrate, and (b) likely imply the necessity of discarding experimental subjects so affected from further analyses. What is more, during the image acquisition process, under-appreciated and occasionally-neglected issues may arise. During DTI scans, artifacts due to scanner vibration may occur and these can translate into undesired tractography results, with certain types of MRI scanners being more prone to vibration artifacts than others. The purpose of this paper is to discuss the potential challenges of multimodal neuroimaging data integration and quality control (QC).

Fig. 1.

Commonly encountered examples of artifacts affecting image quality (in transverse, saggital, and coronal planes): (A) a dental appliance resulting in T1 image distortion, (B) ghosting artifact, (C) echo-planar susceptibility artifacts, and (D) T ₂ -weighted image volume slice misalignment.

2 QC as a Requirement for Multimodal Integration

Beyond the obvious QC issues shown in Fig. 1, two representative examples of often unappreciated challenges are provided for illustration purposes. The first is the scenario where volumes acquired using distinct modalities have different spatial resolutions, which suggests the need for implementing three-dimensional (3D) interpolation during co-registration prior to data integration and analysis. The application of such operations is often accompanied by numerical errors which can affect the quality of DTI tractography, and the detailed effects of co-registration/interpolation methods should be quantified carefully during QC and data integration. Secondly, we illustrate the under-appreciated problem related to QC and to data integration, which involves the occasional presence of systematic vibration artifacts in DTI. Specifically, because large gradient lobes are employed during data acquisition, the accompanying vibrations of the patient table can lead to substantial disruption of diffusion measures, particularly in occipital areas. QC metrics which can identify and quantify such effects automatically would be very useful, as would their resolution using image processing methods which can be applied prior to integrating DTI data with other neuroimaging modalities.

2.1 Accounting for Differences in Spatial Resolution

Neuroimaging data can be acquired using a variety of different modalities which often have distinct spatial resolutions. For example, T ₁ -weighted MRI volumes are often acquired at relatively high spatial resolution because they are used to quantify structural properties of the brain, whereas DTI and PET volumes may have lower resolutions than structural MRI due to the various challenges of acquiring high-resolution volumes using these modalities for water diffusion and metabolic measurements, respectively. For multimodal neuroimaging data integration, however, it is often necessary to scale data to the same resolution because many analysis are performed at the voxel level.

Various methods for the co-registration of DTI volumes to MRI volumes exist. In a very common approach, the DTI B ₀ volume is registered and interpolated to the resolution of the T ₁ -weighted volume, whereafter each DTI gradient volume is registered and interpolated to the already-interpolated B ₀ volume. Though common in practice, this method requires two registration and interpolation steps and appreciable differences in the end result can exist depending on which interpolation and registration algorithms are used, possibly resulting in propagated errors.

An alternative approach involves reconstructing the diffusion tensors in the native space of the DTI volume. Specifically, FA volume and eigenvectors are calculated first and the FA volume is then registered to the T ₁ -weighted volume. The same transformation matrix can then be applied to the other volumes, and this approach requires only one registration operation. Additionally, the operation of eigenvector normalization can avoid the cancellation of fiber directions information which may result from eigenvector interpolation.

Figure 2 illustrates DTI tractography differences which are due to the use of various interpolation methods. As the figure suggests, different interpolation methods can yield vastly different tractography results. Given these substantial differences, it results that reducing the number of times that interpolation is implemented is desirable because it reduces the amount of propagated interpolation error.

Fig. 2.

(A) Effects of applying several DTI data interpolation techniques, namely sinc, average, linear and closest neighbor can lead to substantial differences in tractography results. (B) Eigenvector normalization (B2) is found to improve data quality over the scenario where this operation is not performed (B1). (C) Correction of an artifact due to scanner table vibration. The first row illustrates the artifact, and the second row shows corrected images.

This example illustrates the fact that the use of alternative methods for co-registering and then integrating neuroimaging data may inadvertently result in substantially different results, which can also pose problems from the standpoint of QC. When the alternative approach for data integration is applied as described in the previous paragraph, the implementation of eigenvector normalization is found to result in improved quality of the interpolation due to the fact that this operation is only applied once (Fig. 2).

2.2 Accounting for Scanner Vibration Artifacts

Frequently, scanner vibration artifacts in DTI data are insufficiently appreciated. Although such artifacts have been identified in data acquired from many sites, it is not clear how widespread this problem is and whether all MRI scanners are affected. Such artifacts are due to the vibration of the patient table during gradient data acquisition and can affect DTI recordings, particularly in occipital brain white matter regions. These artifacts can disrupt DTI image quality and can be controlled via a linear correction method, which has been found to be effective (Fig. 2). This figure illustrates that, after using the linear method for the reduction of artifacts due to scanner motion, the effect of the latter is greatly reduced, confirming the improvement in image quality and suggesting that the effects of this artifact can be addressed satisfactorily. For QC purposes, it would be useful to develop methods which identify this type of artifact and which can then signal the researcher that an appropriate correction should be implemented during the process of pre-processing DTI volumes.

Figure 3 illustrates a suggested data integration workflow for MRI/DTI/PET co-registration and analysis, which can be implemented after accounting for problems such as differences in image resolutions and scanner vibration artifacts. Figure 4 shows the results of multimodal integration of MRI/DTI/PET data, as visualized simultaneously.

Fig. 3.

Workflow for multiple modality image registration. Each of the steps involved in multimodal registration can present challenges due to the presence of artifacts, resulting in the need for image QC.

Fig. 4.

DTI/PET/T1-weighted MRI data integration. (A) RGB map with tractography; (B) FA map with tractography; (C) PET image with tractography; (D): PET, FA map with tractography; (E) T1 image with tractography; (F) T1, PET image with tractography.

3 Conclusions

In this paper, we discussed several basic challenges of multimodal image QC and integration of neuroimaging data. We provide two examples of often-neglected and potentially under-appreciated problems related to the QC of diffusion tensor imaging (DTI) data and to their integration with other modalities. The usefulness of minimizing the number of interpolations when registering DTI data to structural MRI was illustrated, in addition to the need to account for scanner vibration artifacts prior to htfidelity data integration. Additionally, we discussed several challenges of multimodal image QC and of neuroimaging data integration. For vibration artifacts, MRI scanners may occasionally induce vibration artifacts during DTI scans due to the nature of pulse sequence designs for this modality.

In conclusion, we provided examples of often-neglected and potentially under-appreciated problems related to the QC of diffusion tensor imaging (DTI) data and to their integration with other modalities. The usefulness of minimizing the number of interpolations when registering DTI data to structural MRI was illustrated, in addition to the need to account for scanner vibration artifacts prior to DTI tractography and analysis. Image QC is a necessary step in advance of high-fidelity data integration.