User:Jadrian Miles/Thesis proposal feedback/dhl notes

4:05 start

as before, nice story

4:07 basser pointing

4:07 research agenda

you are having more "uh"s and "um"s than I think of you doing. I'm not sure how you can best reduce that, but maybe being more cognizant about it will help

4:09 brain anatomy

cute shrinking of sushi into disected corona

4:11

4:12 WM macrostructure

I think the image was the arcuate fasciculus? [but apparently not]

4:14 diffusino MRI

4:16 diffusion-based tissue modeling

nice distinguishing of micro and macro here

Chad's question was more about clustering voxels than about clustering curves. You might want to follow up about that with him, because I think you guys were imagining different topics.

4:20 research agenda

4:20 thesis statement

that's a mouthful!

jfh: what is unambiguous?

I would say that the solution will be better, but not necessarily unambiguous

4:24 talk roadmap

4:25 research context

computer scientists could also be repalced by applied mathematicians or other fields -- your wording could be inferred to mean only computer scientists

4:26 collabs

4:27 pipeline

the negative space of your pipeline is really interesting looking :-)

4:28 pubs

Win's paper isn't "in review" any more, sadly. But maybe he forgot to tell you :-) [I did just check on the pubs page, and it is listed as "in review" not "in press" at this point, fyi.

4:29 MRI protocol development

4:30 curve clustering

4:32 bundle modeling

4:32 roadmap

you should count to yourself when you ask for questions. You are only giving a few seconds -- not enough for almost anyone to ask a question. Subjective presentor time is much longer than clock time...

4:33 research agenda

4:35 anatomical priors

you mention in the doc isertion of WM into GM happens orthogonal to the surface?

4:38 iterative adjustment

wow, they look identical!  :-)

very cute animations -- many made me smile along the way

4:39 piecewise priors

very nice convolution animations :-)

not clear that boundaries are "the most important" places -- important, yes, but be careful about black-and-white statements

4:43 microstructural validation

Are you sure that Ricean is the right distribution for axon diameters? I know that's true for magnitude MR image data, but seems like a strange distribution for axon diameters.

jfh: distribution questions and smoothing questions

4:46 ex-vivo&histology

4:46 research agenda

the red caps and the red parts of the sides of your example fascicle geometry model is a little confusing.

4:48 initializing the combined model

(why are your slide titles blurry? -- everything else looks crisp)

4:48 talk roadmap

jfh: convergence

charniak: why do you get more out of the images after using the current methods?

(nice answer)

4:52 roadmap

4:54 initialization

your undivide-and-conquer in geometry is nifty -- very CS like.

charniak: is that greedy, and if so does it make any difference. Ie, does it matter what order you join things?

4:58 system energy

how will you weight the different parts of the system energy? Product is strange -- shouldn't they be summed? I think that there is a probabilistic derivation that you might be able to use to derive this. My guess is that that will either lead to a product of everything including over the clusters, or a set of sums, including over the different micro-structural elements are.

5:02 timeline

nice echoing of your architecture diagram in your timeline

question: health privacy

ben: how much variation? Can you identify someone from their scan?

chad: what optimization algorithm? is it a standard optimization algorithm?

peter: regularization clever; "there are quite a number of pathways that connect to the outside." [I would have phrased the gray-to-gray constraint as either gray to gray or gray to exit]

jfh: worries that your bundles may be too different from anatomists

jfh: validation -- how do we know it's any good? slice afterwards?

(I stepped out, so didn't note things while I was gone)

peter: jittery boundaries were being discussed when I cam back in

ben: how do you convert image differences (between simulated DWI's and actual) to parameter changes? (this is a deep and good question -- this may be one of the ongoing challenges)

jfh: where is the incentive to avoid too many (100K) "clusters"? I think this comes back to the probabilistic formulation.

chad: unambiguous. I think you probably need to weaken that word... It's an attention-grabber (not in a great way)

chad: re thesis statement, worried about overfitting, parsimony, etc. Probabilistic! Three questions, which I'm afraid I didn't manage to get.

jfh: modify thesis statement. currently involves proving a negative. make a safer statement ("could not be")

peter: will you review kiselev macro+micro approaches (Jadrian note: this is published as "Gibbs Tracking", MRM 2008). spaghetti plate. spin-glass. (not sure that the spin-glass is curve-oriented -- isn't it voxel-oriented?) Peter, I think, is arguing that these methods are still micro combined with macro. Maybe you should say "fasciculus + microstructure" or even something that has more anatomical description of the macro and micro parts as well.

peter: how about a reduced computational phantom that you can try solving for? Even something that stays stable for data created from the model itself.

peter: can check the robustness of solution with a computational phantom where you can add noise.

jfh: bothered that scheme isn't idempotent. ie, brain->images->model->images->model2 -- model should very close to model2, shouldn't it?

chad: suggests presentation could have been inverted. too much high-level stuff too early. may have been selling a little too hard and not giving the detail enough. jfh suggests that a figure showing the micro and macro together, perhaps in 2D to visually show how they would go back and forth.

peter: more demonstration of synthetic data

ben/jfh: recovery from excessive challenges may involve a different thesis statement (ie, if the yellow parts are gone, the thesis may no longer be sufficiently large)

peter: not sure how the whole thing will fit together. worried about putting it into practice and testing it out.

What are the effects on your proposal now to address your issues.

More publications may help with that.

Be prepared for more time -- this work may take longer than you are predicting.