Title: Investigating the Impact of fMRI Pre-Processing Methods on Candidate Biomarkers for Neurological Disorders
Authors: Sadie Walter, OMSII; Ian Bledsoe, MD; Sarah Wang, PhD; Anthony Lee, MD, PhD; Doris Wang, MD, PhD; Melanie Morrison, PhD
Introduction
Functional MRI (fMRI) measures and maps brain activity via small increases in blood oxygen level dependent (BOLD) signals that occur in response to neural impulses. In resting state (rs) fMRI, spontaneous fluctuations in neuronal activity recorded during rest allow for the “functional connectivity” between two or more distinct brain network regions to be computed. Scanner noise, artifact, and physiological fluctuations can introduce non-neuronal noise into BOLD signal fluctuation data, potentially impacting the reliability of fMRI metrics. This investigation aims to evaluate the effects of data preprocessing on fMRI-derived measures of network variability (stability) and frequency (activation), which may serve as promising biomarkers for the management of Parkinson’s disease (PD) and other neurological disorders.
Methods
Twenty-nine patients with advanced PD underwent 3T fMRI prior to deep brain stimulation implantation surgery. The CONN toolbox was used to analyze the data by means of standard preprocessing including spatial smoothing and denoising (via band-pass filtering) followed by independent component analysis to extract common brain networks (default mode, visual, dorsal attention, cerebellar, sensorimotor, and basal ganglia) and calculate variability (standard deviation) and frequency (average power) from each network’s unique fMRI signal. Denoising removes non-neuronal signal, isolating neuronal activity. Spatial smoothing removes high-frequency signal while enhancing low-frequency signal, isolating the synchronous activity between brain regions. The effects of four band-pass filters: 1) 008-.09 Hz, 2) .008-.09 Hz with an additional noise removal technique, 3) .008-.06 Hz, and 4) .016-.09 Hz, and three spatial smoothing kernels: 0mm, 4mm, and 8mm were examined.
Results
Band-pass filter manipulation demonstrated less pronounced effects on the variability metric with each filter application in comparison to the frequency metric, as expected with the filtering frequency manipulation performed. Band-pass filter alteration resulted in default mode network instability, relative stability in the cerebellar, basal ganglia, and sensorimotor networks, and a downward trend in visual network variability. Spatial smoothing manipulation resulted in similar trends, but with fewer statistically significant results and weaker effects on both variability and frequency.
Discussion
These results illustrate the data-altering effects of fMRI pre-processing methods, especially band-pass filtering, and their impact on the reproducibility of brain variability and frequency metrics. fMRI variability and frequency are candidate biomarkers for PD management, however pre-processing steps used to derive these metrics must be reported and standardized across studies. Two major limitations in this study to be addressed in future research include small sample size and the novelty of fMRI metrics as clinical measures.
This judge has two questions: 1) The filters that you used for the first pre-processing methods have frequencies that overlap. Is there a method to be sure that the pre-processing step is reproducible, using the very same frequencies? An example might be a coefficient of variation for the method. 2) The spacial smoothing technique looks a good deal like pixelation (oftentimes used to intentionally blur images beyond recognition), in that much of the detail is lost. Is that part of the reason that there were fewer significant differences seen using this technique, compared to band-pass filtering?
Hi Dr. Elliott! These are great and important questions regarding our data analysis.
In terms of the reproducibility of the data, including during band pass manipulation, I feel it is best to begin with a more detailed explanation of the process of fMRI analysis in an effort to clarify my answer. With each manipulation of the two variables in this study, a separate independent component analysis (ICA) was performed with individual results consisting of both variability and frequency metrics for each. In other words, each parameter (.008-.06 hz, .008-.09 hz, 0mm spatial smoothing, etc) all get their own independent analysis. The parameters are manipulated simply by imputing the desired parameter into a computer program and subsequently running the analysis each time. This is demonstrated in figures 4, 5, and 6, showing the results of interest (variability and frequency for EACH brain region, separately) produced from each indicated (representing the different ICAs performed on each variable parameter. Thus, a total of 7 ICAs were performed, one for each manipulation of the variables. Now, to finally answer your question, the variability metric itself (one of the results of interest) is precisely what you described. fMRI variability is a measure of the relative standard deviation, or coefficient of variation, of the BOLD signal for each brain region. My understanding is that this is what makes variability is such an attractive candidate to apply as a biomarker for a variety of diseases, allowing its application across a variety of patients and groups while maintaining its integrity as a reproducible metric. One of the overarching purposes of this research is to define the “best” pre-processing parameters to apply during analysis in order to standardize these steps across studies, adding another layer of reproducibility- a vital aspect of its ability to serve as a biomarker for disease.
As for spatial smoothing, two important purposes of this step are as follows:
1) No two brains are alike, and the ability to reproduce these analyses across subjects requires normalizing these anatomical variations. During analysis, the individual brain images are overlayed so that spatial correlation can be established between them. Brain regions among individuals rarely occupy the same voxels in space, so spatial smoothing increases the overlap between these voxels and accompanying brain regions. This results in a visually blurring effect, and of course sacrifices some of the precision in brain network locations depending on the degree of manipulation. Balance is key!
2) It attempts to isolate low frequency data involved in brain activity and exclude high frequency data resulting from physiological noise which we don’t want. This essentially reduces the rate of type I errors. Thus, it may reduce the statistical significance, but in a way that gives us a more accurate representation of neuronal activity. Without spatial smoothing, we would be looking at data that includes unwanted BOLD signal, possibly resulting in more significant, but also less accurate results. Our results did show that spatial smoothing had a significant effect on both variability and frequency, but it seems to depend on which brain network is analyzed. The reason for this would be an interesting concept to explore in future research.
I hope this answered your questions, please let me know if further clarification is warranted. Biomedical imaging analysis is complex and technical, making its understanding problematic at times without being provided proper background information. This was difficult to achieve in our time constrained presentation, but I hope this explanation provides some simplification.
The judge has this question: The presenters note the novelty of using fMRI as clinical measures. Please comments on the prior work in this area. What does your study add to what we already know about the impact of pre-processing on data quality and our ability to make clinical inferences. Thanks!
Hi Dr. Fritz! Thank you for your inquiry, as our impact with this project is a very important aspect in our motivation to continue this research.
As you and I pointed out, research involving the reproducibility of variability and frequency metrics with varying pre-processing methods is novel. This absence of precursory exploration is the major limitation of our study. There is scant research involving other metrics. These metrics that have been explored were not of interest to us due to my team’s involvement in neurological disease processes since the effect of pre-processing method parameters and which methods are used greatly depend on the metrics of interest. To demonstrate, the recent manuscript submission for publication referenced in slide 9 applies the results extracted here to brain tumor patients with radiotherapy-induced memory impairments and vascular injury. Since variability has only recently become a candidate biomarker for pathological neurological processes, there has not yet been standardization of these pre-processing methods. Being that significant effects were found with manipulation of both band pass filtering and spatial smoothing, standardization is vital in establishing variability as a reliable biomarker. Hopefully, future research that my team is currently working on will contribute to advancing the utility of variability as an important piece in the future management of neurological and psychiatric conditions.
As a judge, I am wondering your thoughts on the differences in variability between the brain regions. Why do you think that is and do you think ultimately each region may have a “best” standard filter? Are filters specific to disease states or clinical question being investigated?
Hi Dr. Salido! Thank you for your question as it gives me an opportunity to further dive into variability and its potential clinical utility.
Variability metrics in BOLD signal fluctuations not only serve as a representation of a given brain network’s stability- it is also thought to represent the cognitive flexibility of the brain- namely its ability to efficiently process and respond to unexpected external stimuli. Of course, cognitive demand differs across brain regions; as an example, the sensorimotor and visual networks may exhibit this phenomenon. However, these networks were chosen by us due to their increased likelihood in producing meaningful variability data to examine. Furthermore, these networks are commonly what are used to explore the application of variability as a disease biomarker because they are thought to be affected by many neurocognitive disorders. Although each brain network may have an ideal pre-processing parameter, it unfortunately wouldn’t be practical to apply different parameters to different networks or disease processes. This is because it would introduce too many variables into data analysis and ultimately have a negative impact on the reliability and reproducibility of their clinical utility if the parameters aren’t standardized. As I mentioned in my response to Dr. Elliott, balance is key with pre-processing. Each step and its degree of manipulation is associated with its own benefits and drawbacks, so the intent is to find the parameters that can be effectively applied to as many disease processes as possible while minimizing any sacrifices in analytic or statistical elements. The clinical utility of fMRI variability is being explored in a variety of disease processes such as Alzheimer’s disease, social anxiety, neurocognitive aging, Parkinson’s disease, autism spectrum disorder, depression, and OCD. The manuscript we recently submitted for publication referenced in slide 9 used variability as a tool to explore memory impairments and vascular injury in young brain tumor patients. Due to the novelty of this work, there will inevitably be more to come. I am excited about the prospect of fMRI potentially serving as a non-invasive approach to neurological disease and as the first objective measure in psychiatric disorders!