Longitudinal assessment of functional connectivity impairment in rat brains

(1)

Longitudinal assessment of functional

connectivity impairment in rat brains

JOHANNES WENNBERG

(2)

Sammanfattning

För att bätte först˚a hur en ensidig hörselg˚angsatresi p˚averkar funktionella kop-plingar i en växande hjärna har en djurmodell använts. rs-fMRI har samlats in fr˚an en grupp av 13 r˚attor, b˚ade patienter med inducerad hörselg˚angsatresi och kontroller under olika stadier av deras utveckling. En modifierad process som använder ICA-AROMA har använts för att avlägsna störningar, men har inte gett trovärdiga resultat.

(3)

Abstract

To better understand how unilateral ear canal atresia affects the functional con-nectivity in the developing brain an animal model has been used. rs-fMRI data have been compiled from a group of 13 rats, both patients with induced unilat-eral ear canal atresia and controls during different stages of their development. A modified pipeline using ICA-AROMA has been used for noise removal, but has not yielded trustworthy results.

(4)

ACKNOWLEDGEMENTS

Thank you, to everyone who has helped and supported this project. To my ex-aminer Anna Burvall and supervisor Rodrigo Moreno for all their time, energy and support throughout this project. To Daniel J¨orgens and Fabian Sinzinger for all their help with server-, computer- and data-questions. To Magdalena Remppis and Androula Savva for all their help and advice, and for answer-ing questions well after their own respective projects were finished. To Peter Damberg for all the work and help with supplying and explaining the data. To the all the people at Medical Imaging, KTH who have been nothing but wel-coming and helpful throughout this project.

Without all your support this would not have been possible.

(5)

List of Figures

1 An illustration of functional connectivity between 3 different vox-els, A, B, and C, of a rodent brain. The smaller diagrams repre-senting activity over time. In this example the time series of voxel A and voxel B are more strongly correlated than those of voxel A and voxel C. They are by definition therefore more strongly func-tionally connected, regardless of any potential anatomical con-nections. Note: This is not supposed to illustrate how an actual time series of brain activity behaves. . . 9 2 Image depicting the general outline of the pipeline used to prepare

the data from analysis, including removing noise and registering the functional data to an anatomical template. . . 12 3 An example of the skullstripping process. (A) shows a sagittal

view of the brain of a subject which was tilted unusually. The anterior part of the brain, to the right in the image, is higher up in the image, compared to the other subject in (B) which is more representative a normal orientation. In (C) the tilted subject has had its brain automatically segmented using Mialite. In (D) the segmentation from (C) has been improved upon manually. . . 14 4 Example of, from left to right on the first row: A slice of a

skull-stripped brain, the segmented edge of the brain, and the seg-mented cerebrospinal fluid. On the row below are zoomed in segments for the images above. . . 15 5 Overview of the registration process to standard space. The

func-tional data is first registered to the high-resolution anatomical data from the same session. The anatomical data is in turn reg-istered to a standard space in the form of the anatomical data of one subject in the cohort. Note: While this example only shows the 1-month cohort, and the 3-month cohort, an equivalent reg-istration was to be performed using the 6-month data as well. Also, since the 3-month data and the 6-month data was provided already registered, only the potential registrations to the main space was performed. . . 17 6 Intermediate result from the group ICA-analysis. One slice

show-ing one component from MELODIC. To the left is the result of the combined 3-month and 6-month cohort, and to the right is a similar slice from just the 1-month cohort. . . 20 7 Result for component 6, all patients compared to all controls,

stronger connection to controls. . . 20 8 Result for component 14, 3-month patients compared to 6-month

patients, stronger connection to 3-month patients. . . 21 B.1 Result for component 3, all patients compared to all controls,

stronger connection to controls. . . 28 B.2 Result for component 2, all patients compared to all controls,

stronger connection to patients. . . 28 B.3 Result for component 4, all patients compared to all controls,

stronger connection to controls. . . 29 B.4 Result for component 10, 3-month patients compared to 6-month

(8)

B.5 Result for component 8, 3-month patients compared to 6-month patients, stronger connection to 3-month patients. . . 31 B.6 Result for component 4, 3-month patients compared to 6-month

patients, stronger connection to 6-month patients. . . 32 B.7 Result for component 3, 3-month patients compared to 6-month

(9)

1 Introduction

Aural atresia is a condition where the ear canal is either closed or missing. Hav-ing unilateral hearHav-ing loss as a child can negatively affect academic performance and changes in the default mode network (DMN), a functional connectivity net-work, may be related to that effect [18]. To better understand how unilateral aural atresia affects the brain functional connectivity, this project aims to pro-cess fMRI data from rats with induced aural atresia, as well as healthy controls. fMRI data is very noisy, and successful studies require a pipeline of processes for data adaptation and denoising. This project is a continuation of 2 earlier projects by Androula Savva [19] and Magdalena Remppis [17] which both were about creating and refining a pipeline for rs-fMRI data from rats. The projects were also about applying said pipeline to fMRI data acquired from rats with induced damage to one ear canal as well as from healthy controls at different time points. This could hopefully be used as an animal model in a larger project investigating the equivalent condition in humans.

In this project anatomical data has been combined with rs-fMRI BOLD data to gain information on the activity of different parts of the rodent brain. Specif-ically, to gain information on the connections between different areas related to the auditory system.

This report will start with a general introduction of the projects and its key concepts in section 1. Then follows an introduction of the pipeline, the methods used in the pipeline and the subsequent data analysis as well as how they relate to the earlier projects in section 2. Then the results are presented in section 3. Finally the methods, the results, and future ideas are discussed in section 4.

1.1 MRI

MRI is a technique combining many different ideas and techniques from many different areas of science and technology into a fast, non-invasive and versatile tool for medicine and science. By placing a subject in a strong magnetic field it is possible to create 3D images of the subject using the spin of hydrogen atoms in the subject. Through clever adjustments an MRI machine can investigate morphology of different organs, the activity of organs and even, through e.g. diffusion MRI, give information about the small scale structure of organs [3, 10].

1.2 fMRI

(10)

Figure 1: An illustration of functional connectivity between 3 different voxels, A, B, and C, of a rodent brain. The smaller diagrams representing activity over time. In this example the time series of voxel A and voxel B are more strongly correlated than those of voxel A and voxel C. They are by definition therefore more strongly functionally connected, regardless of any potential anatomical connections. Note: This is not supposed to illustrate how an actual time series of brain activity behaves.

1.3 rs-fMRI

As opposed to having the subject performing a specific action or reacting to some stimuli, resting-state fMRI (rs-fMRI) has the subject resting in the scanner as the name suggests. This can be used with many different subjects in many conditions at many different stages of their life making it a good standard state to evaluate [3].

1.4 Functional connectivity-network

This project has focused on connectivity, and more specifically functional con-nectivity (FC) where one tries to find separated areas of the brain which behave similarly, i.e. whose time series are correlated, without considering the poten-tial physical link, direct or indirect, between the areas. There are other similar but distinct forms of connectivity [3]. For a mock illustration of functional connectivity, see Figure 1.

1.5 Animal model

(11)

The methods used to remove noise and prepare the data is referred to as the ”pipeline”. This project has, with some additions, continued using the pipeline on more data from the same data set used in the previous projects. It has also combined the processed data from the earlier project to investigate how ear canal atresia affects the developing rat brain.

1.6 ICA-method

Independent component analysis (ICA) is a method to divide a complex signal with many components into a few distinct, independent components [3]. This method can be used in many different ways. Some examples are given below, to separate signal from noise, but also to divide signal into smaller separate components [3].

It is a way to solve the ”cocktail party problem” due to similarities with the ability to concentrate and hear one speaker in a noisy cocktail party [11]. A different analogy is that of a noisy market. The goal of the spatial ICA-method is then to identify that the loudspeakers belong to the same network, despite being spatially separated, being affected by different noise, and having defects and distortions. The signals are not identical, but there is still a similarity which identifies them as belonging to the same system or network.

1.6.1 ICA as noise removal

As mentioned in section 1.6, the ICA method is used to divide a complex signal in smaller independent components. By using some sort of analysis one can identify some sources as noise, for example caused by movement of the head during the MRI-scanning. Having identified the component as noise it can then be regressed out of the data. Then the residual should have a higher signal-to-noise ratio, compared to the original signal [3].

In this project ICA-based Automatic Removal Of Motion Artifacts (ICA-AROMA) is used to remove sources of noise from the data. ICA-AROMA is a pre-trained classifier of noise in fMRI data which despite being pre-trained and using a small number of features to classify components is robust and yields comparable results to other methods [14, 15].

1.6.2 Group ICA

This group ICA is a multi-subject analysis preformed on many subjects per-formed on all the subjects in the cohort. One performs an ICA-based investiga-tion on all the data taken at a specific age, hoping to find statistically significant differences between control and patient group. This was done in both the ear-lier projects [19, 17]. This is to find statistically significant differences in the rs-fMRI data. With several investigations repeated at different subject ages, one can also gain information about at which age these differences start to appear by comparing the non-longitudinal results between ages. This information could be useful in investigating when the equivalent difference occurs in humans. 1.6.3 Longitudinal ICA

(12)

the processed results from each cohort. While a longitudinal investigation could give insight into the development of the brain, one must consider both the sources of variation inherent in an fMRI study and the added sources of variation due to the study being longitudinal [6, 21].

In this project the FMRIB Software Library (FSL) has been used, but there are many different models and packages for longitudinal analysis of fMRI data such as Statistical Parametric Mapping (SPM), Mixed effect modeling, and Neuropointillist. They all have different advantages and disadvantages, such as being able to handle missing data or not being appropriate for estimating within-individual changes [21, 12].

2 Methods and Material

2.1 Data acquisition

The data set is produced from the same set of rats as the earlier projects by Magdalena Remppis and Androula Savva [17, 19]. 13 rats, 5 controls and 8 patients with surgically induced monaural canal atresia were scanned using a 9.4 T Varian scanner at Karolinska Experimental Research and Imaging Centre (KERIC). They were scanned at the ages of 1, 3, 6, and 12 months. The data of the 3-month old cohort and the 6-month old cohort was analyzed in the earlier projects using the same pipeline as in this project. In this project the 1-month data was analyzed and processed.

Out of the 13 rats, 2 patients had to be excluded. One patient due to missing data, and one patient due to some sort of distortion/corruption in the high-resolution anatomical T2 data necessary for registration in later steps.

2.2 Overview

In general, and especially for data pre-processing and denoising, the methods and pipeline used are the same as in the earlier project by Androula Savva [19] and the project by Magdalena Remppis [17], with only minor changes. Some parts have been modified and will be mentioned when describing the relevant method. This has been done, partly to improve the pipeline and partly for easier comparison with the earlier projects. For example in this project the bias field correction was done before the automated part of the skullstripping as opposed to earlier projects. However, this should not prevent results from being compared as the automatically generated masks were later on manually inspected and corrected in both this and the earlier projects.

(13)

Figure 2: Image depicting the general outline of the pipeline used to prepare the data from analysis, including removing noise and registering the functional data to an anatomical template.

Not pictured in figure 2 is the subsequent analysis of the data which was per-formed using a group ICA analysis for the 1-month cohort and then in different combinations for a combination of cohorts.

2.3 Pre-processing

(14)

2.3.1 Data checks

As a first step, but also continuously though the project, the data was checked for anomalies of different sorts. The number of time points recorded for the functional scan was checked, and also their file sizes. Easier checks, such as comparing file- or directory sizes between subjects were performed more often than more in-depth checks. More in-depth checks were usually preformed when anomalies were found, but also for a few random subjects between steps. When errors were found, the results from earlier steps were investigated, in order to see which steps needed to be repeated.

2.3.2 Nifti conversion

The anatomical data T2 was given in the nifti (.nii) format and did not need any conversion. The functional MRI data was given in Forms Data Format (.fdf) with each time point as a separate file, and had to be converted to the nifti format. This was done using the ”Multi FDF Opener and Multi VFF Opener”-plugin [24] in Fiji [20]. Using Fiji, the time points of the functional data were combined into the 4-dimensional nifti format.

2.3.3 BIDS

The data was reorganized in order to follow the Brain Imaging Data Structure (BIDS) as close as possible [5]. This was done using a script made by [19] for this purpose and for data in this specific format. With the data in the BIDS-format data and meta-data could easily be accessed using PyBIDS (version 0.8.0), a Python module made for working with data in BIDS-format [16]. For more information of how BIDS-format and PyBIDS benefit this kind of project see [19, 5].

2.4 Denoising and registration

Denoising is important for any MRI-analysis, but the indirect and weak signal of the fMRI data makes denoising perhaps even more important. This project uses a mix of automated and manual methods in several steps to remove sources of noise based on their location or behaviour. ICA-AROMA, despite using fewer variables compared to the trainable ICA-FIX yields comparable results depending on the data set investigated [14].

2.4.1 Bias field removal

To achieve the automated step described in section 2.4.2, more uniform data and to make it easier to evaluate the results of the skullstripping, bias field removal was performed before skullstripping in this version of the pipeline. This was done before skullstripping in closer accordance with [4] as opposed to earlier versions of this project [19, 17].

(15)

Figure 3: An example of the skullstripping process. (A) shows a sagittal view of the brain of a subject which was tilted unusually. The anterior part of the brain, to the right in the image, is higher up in the image, compared to the other subject in (B) which is more representative a normal orientation. In (C) the tilted subject has had its brain automatically segmented using Mialite. In (D) the segmentation from (C) has been improved upon manually.

2.4.2 Skullstripping

Using Mialite [2], the volumes containing brain was automatically segmented. These segmentations were then inspected and improved by hand using ITK-SNAP [23]. The main priority was to include all volumes containing brain tissue, with the secondary goal of including as few volumes as possible. This segmentation was then used to cut away all volumes not in the segmentation, keeping as much signal and as little noise as possible. For an illustration of part of the skullstripping process, see figure 3.

2.4.3 Tissue segmentation

(16)

Figure 4: Example of, from left to right on the first row: A slice of a skullstripped brain, the segmented edge of the brain, and the segmented cerebrospinal fluid. On the row below are zoomed in segments for the images above.

areas containing only cerebrospinal fluid, the area just around the brain, as well as a reverse of the mask created for the skullstripping in the earlier step respectively, was created. Instead of segmenting the whole region of interest, which is very time consuming, small patches of a few voxels each was segmented. This process took on the order of 1 hour/subject. See figure 4.

2.4.4 Distortion field correcting

(17)

2.4.5 Registration

The distortion field corrected functional data was then registered using a linear transformation to the high-resolution anatomical T2 data using the FMRIB’s Linear Image Registration Tool (FLIRT) [7, 8] functionality belonging to FSL [9]. This registration was performed with 7 degrees of freedom.

The high resolution anatomical data of each subject was then registered to a chosen subjects anatomical space. This was first performed using FLIRT in order to get a starting point for the non-linear registration FNIRT, also belonging to FSL [9]. The same subject was used as standard space as in Magdalena Remppis project [17]. The subject number was found in internal documentation and was subject 11, a female control rat.

The main output of this step was the linear and non-linear transformations as well as the the transformed tissue segmentations used in later steps. See part of figure 5 for an overview of the relations between different types of data in the registration.

2.4.6 Spatial smoothing

The same distortion field corrected functional data used in the registration was then spatially smoothed using a Gaussian kernel with a full-width-half-maximum of 0.8 mm. This is done to improve the signal-to-noise ratio. As a rule, the width of the kernel is set to around 2 times the voxel-dimensions, but the criteria can vary depending on the number of time points and the size of the structure to be studied [3].

2.4.7 ICA-AROMA

The spatially smoothed functional data along with the relevant subjects masks and warpings were passed into ICA-AROMA in order to produce 15 components which were automatically classified by ICA-AROMA as signal or noise, the noise being removed from the data. To divide the data into 15 components specifically was chosen to be comparable to the earlier projects.

2.4.8 Temporal filtering

The output data from ICA-AROMA was filtered through a high-pass filter using a cut-off period of 100 second. This cut-off period was chosen to make the process consistent with the earlier projects [19, 17]. Applying a high-pass filter is a normal procedure for fMRI data and is done to remove the effects of scanner drift over time [3]. This step is the final noise-removal step of the functional stream and the output is the functional data in functional space.

2.5 Transformation to standard space

(18)

(19)

For comparisons between cohorts in the longitudinal analysis, the last regis-trations was repeated from the respective cohorts standard space to the 3-month cohorts standard space. All this is illustrated in figure 5.

2.6 Group ICA

The Group ICA and the longitudinal ICA both use Multivariate Exploratory Linear Optimized Decomposition into Independent Components (MELODIC) and Dual Regression, both tools from FSL [9]. The functional data in standard space for the cohort was concatenated into what resembles one long time series of only one subject. The different part of the long time series are marked as belonging to either patient or control. MELODIC splits the data into 15 chosen components which are then fed to a Dual Regression which performs a spatial regression from the group components to the subject, and then a temporal regression on the subject level in order to find the connectivity maps for individual subjects which can then be used to compare sub-groups against each other.

The output of the dual regression are heat maps of p-values for the voxels where group A has a stronger connection to the component being investigated than group B has. The tests were also done in reverse to find where group B has a stronger connection than group A. In the Dual Regression a permutation testing is performed, and 5000 permutations are performed, which according to [3] is ”safe” as a lower limit.

2.7 Longitudinal ICA

The longitudinal ICA analysis is essentially performed in the same way as the group ICA. All the processed data of the different cohorts are converted into the standard space of the 3-month cohort and concatenated into one long time series. All the available data has been used, even if that subjects data is missing from one of the cohorts. Two sets of test have been done, all of the patients from the 3-month cohort and the 6-month cohort have been compared with all of the controls in the 3-month cohort and the 6-month cohort, also the patients from the 3-month cohort has been compared with the patients of the 6-month cohort.

(20)

3 Results

The 1-month cohort has not yielded final results, but intermediate results and results of improving the pipeline are presented here. Chosen results of the longitudinal analysis are presented here, with additional images in appendix B.

3.1 Pre-processing

The pre-processing went as expected, converting the data into correct formats and generating a file structure fulfilling the criteria according to the BIDS-validator program from pyBIDS [16]. However, the main script putting the data into the BIDS-format generated temporary files which where not deleted after program completion. This seems to have been caused by a program to parallelize the process. When the process was repeated, the build-up of files caused memory issues. Once the problem was found, a temporary solution was created.

3.2 Improving process

Most of the code has been updated to work with PyBIDS version 0.8.0. Small improvements in user friendliness have been done and further improvements are planned. Earlier methods of parallelizing steps of the process have been improved on. Some steps which were earlier done in series, despite every subject being done independently, has been manually made parallel. For steps where one subject could take up to 10 hours to compute, this has streamlined the process considerably. Some, originally manual, steps have gotten small bash-scripts to speed up those steps and to make them more reliable. There are plans to automate or partially automate further steps.

3.3 Group-ICA

Example results of the analysis using ICA-AROMA are presented in figure 6. The output of the group ICA is at this step 15 extracted components in standard space. No results from the dual regression building on these results have been finished successfully and are not presented her.

3.4 Longitudinal ICA

(21)

Figure 6: Intermediate result from the group ICA-analysis. One slice showing one component from MELODIC. To the left is the result of the combined 3-month and 6-3-month cohort, and to the right is a similar slice from just the 1-month cohort.

(a) Slice 37 (b) Slice 35 (c) Slice 33

(d) Slice 31 (e) Slice 29

Figure 7: Result for component 6, all patients compared to all controls, stronger connection to controls.

(22)

(d) Slice 29 (e) Slice 27 (f) Slice 25

(g) Slice 23

Figure 8: Result for component 14, 3-month patients compared to 6-month patients, stronger connection to 3-month patients.

3.4.1 All patients compared to all controls

The component deemed most interesting was component 6 with stronger con-nection in the combined control group. The regions shown in figure 7 were estimated to be related to the primary somatosensory cortex, forelimb region and the primary somatosensory cortex, hindlimb region, as well as simple lob-ule A and simple loblob-ule B respectively. All other components were deemed not interesting due to size or placement outside brain.

3.4.2 3-month patients compared to 6-month patients

(23)

Table 1: The number of significant test results from the 2 longitudinal dual regressions based on minimum p-value.

Experiment Significant results Not significant results All patients and all controls 4 26

3-month, and 6-month patients 10 20

4 Discussion

While the lack of results makes the evaluation of the modified pipeline difficult, focus have been given to the methods used, intermediate results, and future prospects. Meanwhile, the numerous results from the longitudinal analysis have been time consuming to categorize and difficult interpret, however some of the most interesting results are highlighted and discussed.

4.1 Methods

As of now, there seems to be a flaw in the pipeline, either some error introduced in this project, or some quality check missing to catch a potential error or irregularity in the data causing this problem. Given the results from the earlier projects [19, 17] it seems possible that some error has been introduced.

4.2 Results

4.2.1 1-month cohort investigation

As of now, there are no complete results from the 1-month cohort to discuss. The results closest to a finished analysis would be the 15 components of the group ICA in both the 1-month cohort, an example of which is shown in figure 6.

The very blurry background in the 1-month cohort analysis indicates that something has gone wrong and this calls the results into question. The back-ground is supposed to be a combination of all the registered functional data creating the somewhat cloudy background seen in the left image. The results from the dual regression builds on these components and is therefore as unreli-able the results from ICA-AROMA and are not presented in the cases the dual regression finished.

Attempts to find the problem at several earlier steps in the pipeline have been done but did not yield any results in time.

4.2.2 Longitudinal analysis

(24)

seem to be a persisting difference between patient and control over time in the rats functional connectivity, which supports the conclusion that unilateral aural atresia affects the functional connectivity of the brain.

In the longitudinal analysis of the 3-month patients compared to the 6-month patients there were many significant results, even if some were deemed uninteresting. The result of component 14, testing for stronger connectivity with the 3-month cohort has a region which seems to be related to different auditory parts of the brain. Interestingly enough, this could indicate that the connections related to the hearing are changing asymmetrically as late in the rats development as between ages 3 months and 6 months. It is very difficult to draw conclusions from this, but it would be very interesting to complement this analysis with a similar comparison between additional cohorts. However, effort would perhaps be better spent investigating some of the alternatives to FSL in order to perform a more sophisticated analysis with the data already acquired. Some of the other component test would also be interesting to investigate more closely. For example component 2 has significant results for stronger con-nectivity in both directions, and an investigation into how the concon-nectivity belonging to one component is developing could be very interesting.

The varying quality of the results from the longitudinal ICA-analysis could be a sign that some other tool than FSL would be better suited for an analysis. The feature of being able to subtract the mean of the controls from the patients in their respective cohort would be an interesting feature and would lend more support to the results from such an analysis. Also, it would be interesting to investigate what results a seed-based analysis, as opposed to an ICA based one, as used in [19] would yield if used longitudinally. Worth noting is that both the segmentations and the identification of the relevant regions were done by the author with the help of an anatomical atlas [13]. Due to very limited experience of the author, results could probably be much improved if an expert did the segmentation and identification.

4.3 Future projects, improvements and problems

Apart from general improvements which applies to almost any investigation, such as having more data, measured more often, there are some more specific possible improvements and future projects for this project.

4.3.1 Automation

Both the pre-processing and the pipeline itself has steps which at the moment are manual, but which could be automated. The pre-processing involved manually checking the number of files in each subject’s data set and check the files for abnormalities. This could at the very least be made easier through full or partial automation.

(25)

4.3.2 Tissue segmentation

As mentioned in section 2.4.3, tissue segmentation, this version of the project uses a different method for the tissue segmentation compared to [19] and [17]. The earlier projects used a time-consuming method of completely marking the features to be segmented on 1 of the subjects which were then transformed into standard space. This project did instead to a rough ”spotty” segmentation of the different features for every subject. The idea is then that with much of the noise represented in those spots, ICA-AROMA would then regress out that contribution.

Some sort of investigation into how robust and effective this method of tissue segmentation would be interesting in a future project. The complete segmen-tation of the earlier projects only has to be done once, no matter how many other subjects are used, while the rough method of this project takes a shorter amount of time for every subject. Therefore it would be interesting to com-pare the effectiveness of the respective methods in order to make an informed decision of which method to chose in the future.

Performing the investigation without artificial data where one knows the ground truth could be challenging. One relatively easy idea to perform would be to randomly remove parts of the segmentations in order to make them even more spotty. Then use ICA-AROMA with the new masks. Then one can redo the experiment several times with the same data in order to measure how consistent the results are. Varying how much of the segmentation is removed could also perhaps give information of how much of the tissue needs to be ”sampled” in order to produce consistent and reliable results.

4.3.3 Standardization

A standardization of the input-output of the different scripts in the project could also be a great way of making the code controllable from another script, potentially automating many parts of the pre-processing and the pipeline. 4.3.4 Printouts and logging

Somewhat related to standardization is that many of the scripts lack a consistent printout of the process in the terminal or in a log file. Many of the tools such as ICA-AROMA has this sort of printout already, but making sure every script has such a system would facilitate debugging and determining how long a process has run.

4.3.5 Parallelization

(26)

5 Conclusions

The pipeline created by Androula Savva [19], and then improved upon and converted to work on a Linux server by Magdalena Remppis [17] and improved in this project seems very promising for processing rodent data. While there are steps which could be taken to improve the pipeline in terms of user friendliness and efficiency, it seems to remove enough noise from the fMRI data to produce results in the later analysis using methods such as ICA. Having a fast and standardized method for processing could simplify later analysis of rodent data and allow researchers to focus more time and energy on analysis.

The 1-month data has not been completely processed without error. Thus it is not possible from this part of the project to draw any additional conclusions about the effects of unilateral ear canal atresia on the functional connectivity in rat brains.

Hopefully the changes done to the implementation of the pipeline and docu-mentation done in this project can help future projects using this pipeline. Also should the problem in the authors treatment of the 1-month cohort be found, this data could be added to the longitudinal analysis of the functional connectiv-ity. However, since the processing of the data failed, the changes to the pipeline need to be viewed with some suspicion, even if it is not the only possible source for errors. The masks, segmentations and many of the improvements to the pipeline will at the very least speed up the process in the future.

(27)

A

Software details

(28)

B

Longitudinal ICA, additional results

Table B.1: All longitudinal tests with significant result and an evaluation. Sorted by increasing minimum p-value.

Component, test Interesting Reason for decision All patients versus

all controls

3, control > patient No Mostly outside brain 6, control > patient Yes

2, patient > control No Too small 4, control > patient No Too small 3-month patients versus

6-month patients

14, 3-month > 6-month Yes

10, 6-month > 3-month No Component excluded 8, 3-month > 6-month No Result is hard to interpret 4, 6-month > 3-month No Result is hard to interpret 3, 3-month > 6-month Yes

(29)

B.1 All patients compared to all controls

Figure B.1: Result for component 3, all patients compared to all controls, stronger connection to controls.

(a) Slice 29

(30)

(a) Slice 32

(31)

B.2 3-month patients compared to 6-month patients

(32)

(g) Slice 25 (h) Slice 23 (i) Slice 21

(j) Slice 19 (k) Slice 17 (l) Slice 15

(m) Slice 13 (n) Slice 11 (o) Slice 9

(33)

(34)

(35)

(36)

(j) Slice 16

(37)

(38)

(39)

(j) Slice 20 (k) Slice 19

(40)

References

[1] Jesper L.R. Andersson, Stefan Skare, and John Ashburner. “How to cor-rect susceptibility distortions in spin-echo echo-planar images: application to diffusion tensor imaging”. NeuroImage 20.2 (2003), pp. 870–888. issn: 1053-8119.

[2] Simone Bendazzoli et al. “Automatic rat brain segmentation from MRI using statistical shape models and random forest”. Medical Imaging 2019: Image Processing. Vol. 10949. International Society for Optics and Pho-tonics. 2019, 109492O.

[3] Janine Bijsterbosch, Stephen M Smith, and Christian F Beckmann. Intro-duction to Resting State FMRI Functional Connectivity. Oxford Univer-sity Press, 2017.

[4] Oscar Esteban et al. “fMRIPrep: a robust preprocessing pipeline for func-tional MRI”. Nature Methods 16.1 (2019). issn: 1548-7091.

[5] Krzysztof J Gorgolewski et al. “The brain imaging data structure, a for-mat for organizing and describing outputs of neuroimaging experiments”. Scientific Data 3 (2016), p. 160044.

[6] Megan M Herting et al. “Test-retest reliability of longitudinal task-based fMRI: Implications for developmental studies”. Developmental cognitive neuroscience 33 (2018), pp. 17–26.

[7] Mark Jenkinson and Stephen Smith. “A global optimisation method for robust affine registration of brain images”. Medical Image Analysis 5.2 (2001), pp. 143–156. issn: 1361-8415.

[8] Mark Jenkinson et al. “Improved Optimization for the Robust and Accu-rate Linear Registration and Motion Correction of Brain Images”. Neu-roimage 17.2 (2002), pp. 825–841. issn: 1053-8119.

[9] _{M Jenkinson et al. “FSL”. Neuroimage 62.2 (2012), pp. 782–790. issn:} 1053-8119.

[10] Denis Le Bihan and Heidi Johansen-Berg. “Diffusion MRI at 25: Exploring brain tissue structure and function”. NeuroImage 61.2 (2012), pp. 324– 341. issn: 1053-8119.

[11] Te-Won Lee et al. “Combining time-delayed decorrelation and ICA: to-wards solving the cocktail party problem”. Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP ’98 (Cat. No.98CH36181). Vol. 2. IEEE, 1998, 1249–1252 vol.2. isbn: 0780344286.

[12] Tara Madhyastha et al. “Current methods and limitations for longitudinal fMRI analysis across development”. Developmental cognitive neuroscience 33 (2018), pp. 118–128.

[13] George Paxinos and Charles Watson. The rat brain in stereotaxic coordi-nates. 6th ed. Elsevier, 2007.

(41)

[15] Raimon H.R. Pruim et al. “ICA-AROMA: A robust ICA-based strategy for removing motion artifacts from fMRI data”. NeuroImage 112 (2015), pp. 267–277. issn: 1053-8119.

[16] _{PyBIDS. version: 0.8.0. url: https : / / github . com / bids - standard /} pybids.

[17] Magdalena Remppis. Independent Component Analysis of Resting-state Functional MRI Data of Rats with Monoaural Atresia. FH Aachen Uni-versity of Applied Sciences and KTH Royal Institute of Technology. Dis-sertation. 2019.

[18] Anna-Katharina Rohlfs et al. “Unilateral hearing loss in children: a ret-rospective study and a review of the current literature.” European jour-nal of pediatrics 176.4 (2017), pp. 475–486. issn: 1432-1076. url: http: //search.proquest.com/docview/1862936816/.

[19] Androula Savva. Assessment of Functional Connectivity Impairment in Rat Brains. School of Engineering Sciences in Chemistry, Biotechnology and Health, KTH Royal Institute of Technology. Dissertation. 2019. url: http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-247593. [20] Johannes Schindelin et al. “Fiji: an open-source platform for

biological-image analysis”. Nature Methods 9.7 (2012). issn: 1548-7091.

[21] Eva H Telzer et al. “Methodological considerations for developmental lon-gitudinal fMRI research”. Developmental cognitive neuroscience 33 (2018), pp. 149–160.

[22] Nicholas J Tustison et al. “N4ITK: Improved N3 Bias Correction”. IEEE Transactions on Medical Imaging 29.6 (2010), pp. 1310–1320. issn: 0278-0062.

[23] Paul A Yushkevich et al. “User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability”. Neuroimage 31.3 (2006), pp. 1116–1128.

Longitudinal assessment of functional connectivity impairment in rat brains