fionaeyoung / thesis

A medical imaging PhD thesis, written in LaTeX, (hopefully) comprehensively version controlled

Fibre tract imaging with intraoperative diffusion MRI for neurosurgical navigation

A doctoral research project by Fiona Young.

Undertaken at University College London, funded by the UCL EPSRC Centre for Doctoral Training in Intelligent, Integrated Imaging In Healthcare (i4health)

Supervisors: Prof. Jon D. Clayden, Mr. Kristian Aquilina, Prof. Chris A. Clark

Abstract

Mapping and understanding the brain's structure and function is never more critical than when it suffers injury or illness. Lifesaving neurosurgical procedures may put essential neural communication pathways called white matter tracts at risk, with grave consequences for the patient, so accurately depicting their location using diffusion magnetic resonance imaging (dMRI) is becoming a key component of modern neurosurgical practice. More recently, obtaining new intraoperative MRI partway through surgery has demonstrated potential to further improve outcomes by providing updated anatomical information after the dynamic effects of intraoperative brain shift have diminished the accuracy of preoperative imaging. With the ability to sample directional water diffusivity in tissue, dMRI produces millimetre-scale maps of white matter fibre orientations which are key to reconstructing individual tracts. However, established image computational methods suffer from limitations in accuracy and practicality which restrict the wider clinical uptake of dMRI white matter imaging generally, and particularly for intraoperative MRI. After an in depth review of the state of the art in white matter imaging and image-guided neurosurgery, this thesis explores the development of a novel white matter tract mapping tool, named tractfinder, which applies a priori anatomical knowledge encoded within a statistical tract orientation and location atlas to achieve rapid tract segmentation in a patient dMRI scan. The proposed pipeline includes explicit patient-specific modelling of tumour deformation effects, an element missing from many research-oriented tract reconstruction approaches. Tractfinder's effectiveness in a range of applications is detailed through thorough quantitative evaluation, while clinical case studies demonstrate its key advantages over existing approaches. In addition, the technical and practical challenges of intraoperative imaging are explored together with their implications for effective clinical translation of advanced dMRI-based white matter imaging.

Copyright

You are free to use the contents of this repository, with appropriate attribution, under a CC-BY-NC licence.

thesis's Issues

Corrections

Post-viva minor corrections for final version, both self determined and examiner submitted.

Self-found

Frontmatter

• Correct author formatting in research paper declaration (full names in author lists in sections 1f/2d for each publication)
• Reset citation numbering after TOC (don't count from research paper declaration form).
• Reset header contents for Introduction (currently still says list of abbreviations)

Chapter 1 Neuroanatomy

• §1.4, end second paragraph, page 41: "resulted in an inevitably imperfect..."
• §1.4, page 41: Set off new paragraph from "Medium and long range..."
• §1.4, page 41: "There are 3 categories ..." -> "Tracts are grouped into three categories..."
• XKCD figure: \vfill either side, and correct hyperlink formatting. (Should display comic title)

Chapter 2 MR physics

• §2.1: Delete "residual strong force"
• eq(2.1): $B$ -> $B_0$
• §2.1, footnote: $I$ -> $\mathbf{I}$
• Fig. 2.1: "… precession of a nucleus ... around an external ... "
• §2.1, end: “… external field strength $B$ → $B_0$”
• eq(2.2): $B_0$

Chapter 6 Tractfinder atlas

• §6.2.1, first line: LaTeX artefact “[noindex=false]”

Chapter 7 Deformation modelling

• §7.2.1: add comma “from $S$ to the brain surface , and $D_t$ … “
• §7.2.1, after eq(7.1): “reaching 0 only at the brain boundary” → reference Fig. 7.5
• §7.2.1: add comma “… computable, close-form , and invertible” → add Oxford comma.
• §7.2.1: “We begin with a function of the form $k(P) \propto e^{-\lambda\frac{D_p}{D_t}}$” → denominator in the exponent should be $D_b$!

Chapter 8 Evaluation

• §8.4: Clarify thresholds in Figs. 8.10-11 (experiment in section 8.4, threshold of 0.01 applies)
• Fig. 8.12: Don’t expand “World Health Organisation” in caption

Chapter 9 Application

• Fig. 9.3: Typo “$b$b1000”
• Fig. 9.7: Sans-serif face “e.”

External examiner SJP

Chapter 1

• §1.3, page 9: "endocrine" -> "neuroendocrine"
• §1.4, page 42: Define "hodology" more clearly

Chapter 4

• §4.1, page 79: “There is also no general consensus on how these categories [GTR etc] are defined” → New RANO resection guidance, recently published: https://doi.org/10.1093/neuonc/noac193
• §4.2, page 81: Role of radiotherapy in children - different issue. RT causes significant cognitive injury, use chemotherapy to delay need for RT as long as possible. Mention this as it is critical
• §4.3, page 84: How can you identify invasion vs. compression etc. with dMRI
• §4.4, page 87: 5-ALA — which tumours does it work on? Look at paediatrics data for paediatric tumours - https://pubmed.ncbi.nlm.nih.gov/30989383/. This is important as it means 5-ALA is only useable in a limited number of tumours (certainly in paediatric practice)
• Overall: additional focus on paediatric tumours?

Chapter 5

• Highlight objectives clearly in a bullet list

Chapter 6

• 1st sentence: “The chosen approach” what other approaches were discussed? How did you get to this approach
• General: potentially put ROI figures (or abbreviated version) in the main text
• Tab. 6.1: What do the numbers mean → clarify streamline counts
• §6.6: DICE coefficients, what is the ground truth? Is 0.7

Chapter 8

• §8.2.1, page 149: Reason for poor registration in excluded 9 TractoInferno subjects
• Fig. 8.2: Explain figure (what does this mean). is DICE score sufficient for clinical use? Any idea of threshold?

Chapter 9

• Fig. 9.2: Distortions of images, issues with intraoperative imaging? Frontal lobes and midline issues esp corpus callosum. Have they tried susceptibility maps to measure distortion - probably won't affect tracts like AF and CST? (-viva comment: you explained the methods well but were too long to apply)
• General: Your 'gold standard' is direct stimulation. Have you compared accuracy of tracts with DES?

Chapter 10

• General: Have you achieved objectives -- should you outline how? Again, not clear -- put as itemised list with explanation of what done?
• General: Limitations -- Effect of steroids. Intraoperative use? Changes diffusion parameters. Is this important?

Internal examiner EP

• Chapter 1: It would be nice to have somewhere near the introduction clear aims for the thesis (so that then you can refer back to them in your bullet point list to see what you have achieved).
• Typos: page 54: "functional" needs capital.
• page 129: the the arcuate...
• Figure 8.6 caption: the the projection

Conclusion chapter: rough draft

Place for summarising the whole thesis

• critical discussion of tractfinder: strengths and weaknesses
• future directions
• perspective on intraoperative diffusion MRI in general
• ...

Customised workflow(s) for building with and without figures

Would be nice to be able to recognise different types of draft documents w.r.t. figures and run workflow accordingly.

For example:

• No need to clone figures from OneDrive if draft mode enabled which replaces figures with placeholders (might cause file-not-found issues if still tries to read the figure file to get it's size)
• If using endfloat, split output into text-only and float-only PDFs

Look into this further down the line if it becomes an issue

Improvements to inner product and mapping section

• Small paragraph describing KL divergence and other alternative potential metrics
• stronger motivation
• discussion of competing effects to IP value: integral amplitude, and angle correspondence

BTCD dataset [Chapter 4]

Add description of BTCD dataset and analyses

Benchmark validation with ground truth [Chapter 4]

This covers any validation against a "ground truth", where all tested methods have "seen" the same training and test data.

• Atlas trained / tesed on TractSeg train / test data split
  • expand this analysis to all tracts
• FibreCupPhantom??
• Could consider training / testing atlases using the TractoInferno dataset (for generalisability)
  -> has more data than TractSeg (144 vs 105)
  -> more hetrogenous data from different centres / scanners
  -> different tract definitions

Language: tractfinder

"tractfinder" or "Tractfinder" or "tractfinder"?

I personally don't like capitalisation, but "tractfinder" makes it sounds like a generic technique with different implementations, like tractography.

Language: first person voice

first person singular voice is generally avoided in scientific writing, lest we be somehow reminded that there are actual humans behind the work.
Most of the pre-written text lifted from papers is written in first person-plural, and various amounts of passive-voice-gymnastic. This makes sense when writing on behalf of multiple authors, but feels wierd for a thesis.

Research paper declaration

Issue for documenting the various citations and declarations for previously published works.

Guidance: https://www.ucl.ac.uk/doctoral-school/rights-and-responsibilities/research-integrity-and-ethics/guidance-incorporating-published-work-your

• Chapter 3, sections 3.3-3.5.1 (Atlas and mapping): ISMRM abstract, HBM manuscript, CARS manuscript
• Section 3.6 (training data requirements): IEEE abstract
• Chapter 4: CARS manuscript
• Section 4.1, graph comparing linear and nonlinear registration
• Sections 5.1-5.3.4 (excluding 5.1.4-5) (validation): HBM paper
• Section 6.1.1 (data quality and CSD): ISMRM Diffusion abstract
• Section 6.2 (brain shift case study NHNN): CARS manuscript

Quantitative validation overhaul [Chapter 4]

The validation section is garbage. Have concluded that splitting the results analysis based on dataset makes no sense, and is repetitive. Instead I've identified these "themes" around which we can restructure the text:

1. Consistency
   - relative performances between methods are the same accross datasets
   - TF results exhibit low variability in comparisons with tractography (and all other methods for that matter)

2. Tract definitions influence results more than anything else
   - within method vs within tract comparisons: tract matters more
   - projection tracts tend to have higher agreement than association (this is WEAK unless we add IFOF to the analysis)

3. Comments on the comparison metrics
   - signed bundle distance more informative than the unsigned version
   - density correlation highlights the threshold issue

4. Findings are consistent with other publications, namely TractSeg vs RecoBundles

5. tractfinder is good
   - "better" than TractSeg (DKFZ) when compared with tractography in 4/12 comparisons (tracts and metrics) across HCP and tractoinferno datasets
   - "better" than TractSeg (XTRACT) everywhere

What still needs to be worked on:

• Any discussion about just atlas registration
• should we relinquish any sort of value judgement for this section, given that we go on to directly compare TF and TS on the same training data?
• How to confer these messages in graphical form --> come up with some better figures!
• Adding IFOF to this analysis would be strong if pos!!!
• Maybe just ditch the tractoinferno reference streamlines, because they are so variable!! Have a figure that shows this well? And then don't consider them any further

Other notes

Have decided to ditch the generalised DSC as it doesn't add anything and is confusing as noone else uses it and the density correlation gives you basically the same info

Atlas/mapping: first draft [Chapter 2]

Need to flesh out the mapping part of chapter 2.

This includes adding the fixel-decomposed inner product approach.

Language: tense

When talking about specific experiments, use present or past tense?

"The segmentations were compared using dice scores"

"We consider two possible approches"

etc.

ISMRM diffusion workshop abstract [Chapter 4]

Write up the "stability with decreasing data quality" study with more emphasis on tractfinder vs the other methods.

Consider revisiting this study with the fixel method?

Expand Atlas and mapping chapters

• Stronger and more detailed motivation of the atlas and how it's constructed, including discussions of uncertainty and tract definitions
• literature review for each tract definition

Atlas registration [Chapter 4]

• Complete section on nonlinear registration genearlly, and specifically T1+FA registration
• Expansive discussion on affine registration: considerations from mirroring the atlas construction and practicality for application, FOD reorientation etc.

PDF/SVG figures with LaTeX text

Hacks and solutions

Center-aligning text: https://tex.stackexchange.com/questions/205659/ensure-position-of-latex-code-in-inkscape-graphic#comment482324_206691

Expand tumour deformation modelling section

Increase detail on motivation for model, choice of parameters and implementation.

• More explanation model need for selection of $\lambda$ (maybe with figure?)
• Reverse deformation model: why would you need it
  -> motivate need for general model that can be inverted
• Explain need for normalisation constant better
• details of implementation: simplification of tumour outline (convex / concave), initialisation of $lambda$ etc.
• ...?

Workflow build improvments

Putting this of for now.

• Convert to docx using pandoc. Currently there are issues with setting the working directory / pointing pandoc to included files and figures that aren't in the pandoc working directory
• Add PDF builds to a tagged release. Only upload as workflow artefacts for intermediate tags (vX.x.x) and upload as release for minor tags (vX.x)

LitReview: first draft [Chapter 1]

This is obviously going to take a lot of effort, and I have the issue of avoiding self-plagiarising my MRes.

As a first goal, write an "unsubstantiated claims" draft: write everything from memory full of terrible amounts of gut feeling and subjectivity, without adding any references at this point.

Will need to try and avoid retrofitting citations to support statements though. So try and avoid making to many specific claims at this point.

Part / epigraph formatting

Exact formatting not decided on, but will need to hack \part to get epigraph text on same page and increase vertical spacing before part title.

TeX macro for formatting abbreviations

This may turn out to be unnecessary, but if I have to keep defining abbreviations* in loads of float captions then this will be useful.

The macro would take a list of abbreviations, like \abbrevs{XY, ABC} and format them a certain way, e.g XY: <full definition>, ABC: <full definition> or XY=<full definition>; ABC=<full definition>. This would have to be linked up with some array or set of defined abbreviations.

That way I don't have to worry about consistently formatting abbreviation lists from the start, or repeatedly reformatting them all at once.

Possibly doable with only the acronym package and either \newcommand or xparse, maybe also etoolbox.

---

*abbreviations / acronyms / initialisms

BTCD data analysis

Quantitative validation of BTCD dataset with reference tractography (manual+filtered)

• Split results by ipsi-/contralateral hemisphere
• Specifically compare TF and TS in cases where tumour deformation (virtue) was used

Registration (linear / non-linear) (first draft) [Chapter 3]

• General problem: alignement of atlas with data
• linear registration
  • case for -- allows / mirrors linear registration in atlas creation
  • case against -- can be too broad e.g. tracts that are close together with similar directions -> FPs
• non-linear registration
  • allow for intraoperative brain shift
  • handle patient differences not covered with tumour deformation
  • ...

Appendices

Would be cleanest to put the spherical harmonics section as an appendix, question is only if it's too critical to not have in main text. If so, where to put it? feels awkward to stuff it between mri physics and neuroimaging, but it's relevant to more than just the CSD explainer

Secondly, put tractography ROIs in appendix, like in HBM paper? Or stick to text explainers in atlas and validation chapters?