Computer Vision REU Computer Vision REU Week 12 Week 12 Adam Kavanaugh Adam Kavanaugh
Apr 01, 2015
Computer Vision REUComputer Vision REUWeek 12Week 12
Adam KavanaughAdam Kavanaugh
Recap from Last TimeRecap from Last Time
Began work with a blurring method to Began work with a blurring method to improve mask generation and background improve mask generation and background removal accuracyremoval accuracy
Found that only 30/71 slides were having Found that only 30/71 slides were having brain slices cut from them with previous brain slices cut from them with previous “Slicer” implementation“Slicer” implementation
As of Friday, “Phase 2” is on hold pending As of Friday, “Phase 2” is on hold pending Dr. Lobo’s approval.Dr. Lobo’s approval.
Blurring the DatabaseBlurring the Database
In an attempt to improve the ability of the In an attempt to improve the ability of the slicer to extract brains from the slides, I slicer to extract brains from the slides, I applied a blur to the D3 level of all the applied a blur to the D3 level of all the slides.slides.
Blur was applied at Sigma values of 1, 2 Blur was applied at Sigma values of 1, 2 and 3 to find which of the 3 values was and 3 to find which of the 3 values was best suited for this purpose.best suited for this purpose.
Sigma 3: AnalysisSigma 3: Analysis
At Sigma 3, slice cuts improved At Sigma 3, slice cuts improved dramatically on the very noisy slidesdramatically on the very noisy slides
However, most cuts were not usable as However, most cuts were not usable as they merged together multiple brains into a they merged together multiple brains into a single mask.single mask.
Extracted slices from 70/71 slides.Extracted slices from 70/71 slides.
Sigma 2: AnalysisSigma 2: Analysis
At Sigma 2, merging of multiple brains At Sigma 2, merging of multiple brains becomes less of a factor, but this trades becomes less of a factor, but this trades off with the loss some slides all together off with the loss some slides all together as well as a loss of detail around edges.as well as a loss of detail around edges.
However, this resulted in more “useable” However, this resulted in more “useable” brain slices than at the Sigma 3 level.brain slices than at the Sigma 3 level.
Extracted slices from 68/71 slidesExtracted slices from 68/71 slides
Sigma 1: AnalysisSigma 1: Analysis
Best overall sigma value for the D3 image Best overall sigma value for the D3 image level.level.
Least amount of combined brain slices, Least amount of combined brain slices, while still cleaning up the background while still cleaning up the background enough to make removal very smooth.enough to make removal very smooth.
More defined masks than Sigma 2 levelMore defined masks than Sigma 2 levelExtracted slices from 70/71 slides, with the Extracted slices from 70/71 slides, with the
highest overall slice count from most highest overall slice count from most slidesslides
Original D3Original D3
Without BlurWithout Blur
Sigma 3 ExampleSigma 3 Example
Sigma 2 ExampleSigma 2 Example
Sigma 1 ExampleSigma 1 Example
Overall Blur Effect on InputOverall Blur Effect on Input
Half of the data slides had a moderate to Half of the data slides had a moderate to high level of noise and other background high level of noise and other background problems.problems.
Applying a blur to the D3 level smoothed Applying a blur to the D3 level smoothed out much of this noise.out much of this noise.
Sigma 1 worked best due to the Sigma 1 worked best due to the decreased detail at the D3 level which decreased detail at the D3 level which caused many brains to be mergedcaused many brains to be merged
From Friday..From Friday..
Devastation Struck! Devastation Struck! Other efforts put on hold.Other efforts put on hold.Worked with Dan over the weekend and Worked with Dan over the weekend and
on Monday to resolve the issue with the on Monday to resolve the issue with the matcher.matcher.
Current status: Stable Condition, Current status: Stable Condition, Recovered. Recovered.
Some Matching TestsSome Matching Tests
In support of Dan’s efforts I got his full In support of Dan’s efforts I got his full system and modified it for my testing system and modified it for my testing purposes.purposes.
At the time he was testing different At the time he was testing different threshold levels starting from 33% and threshold levels starting from 33% and moving Upmoving Up
I began at the top with 100% and also I began at the top with 100% and also tested at 90%tested at 90%
Full Pixel AllowanceFull Pixel Allowance
Matcher results on test query of slide Matcher results on test query of slide MnSOD:MnSOD:
Top 6 slide matches:Top 6 slide matches:1) MnSOD [1.084780]1) MnSOD [1.084780]2) MnSOD [1.127043]2) MnSOD [1.127043]3) MnSOD [1.138518]3) MnSOD [1.138518]4) MnSOD [1.138938]4) MnSOD [1.138938]5) BAPP-LPSLesion-6-2 [1.234906]5) BAPP-LPSLesion-6-2 [1.234906]6) MnSOD [1.388898]6) MnSOD [1.388898]
90% Pixel Allowance90% Pixel Allowance
Same test using a slice from MnSOD:Same test using a slice from MnSOD:Top 6 results:Top 6 results:
1) MnSOD [1.580257]1) MnSOD [1.580257]2) MnSOD [1.968792]2) MnSOD [1.968792]3) MnSOD [1.981233]3) MnSOD [1.981233]4) MnSOD [2.011915]4) MnSOD [2.011915]5) MnSOD [2.064471]5) MnSOD [2.064471]6) MnSOD [2.094589]*6) MnSOD [2.094589]*
*MnSOD matches continue through top 9*MnSOD matches continue through top 9
Matcher Results and MoreMatcher Results and More
As of today, matcher seems to be back up As of today, matcher seems to be back up to quality and providing the same, if not to quality and providing the same, if not better results than we demonstrated better results than we demonstrated previouslypreviously
Look for a more detailed and in-depth Look for a more detailed and in-depth analysis of these and more matcher analysis of these and more matcher results on Friday when Dan presents his results on Friday when Dan presents his findings.findings.
The Fate of BlurThe Fate of Blur
Currently, the mask growing technique Currently, the mask growing technique that Dan has developed works well that Dan has developed works well enough for our purposes in cutting out enough for our purposes in cutting out brain slices.brain slices.
If another matching criteria is needed, then If another matching criteria is needed, then using the blur to produce the masks might using the blur to produce the masks might become important due to more detailed become important due to more detailed masks.masks.
Moving ForwardMoving Forward
Continue making adjustments to the 2D Continue making adjustments to the 2D system to help bring it closer to system to help bring it closer to finalization.finalization.
Familiarize myself with the 3D system so Familiarize myself with the 3D system so as to begin work sorting out the problem as to begin work sorting out the problem presented to us in Phase 2:presented to us in Phase 2:
“Brain Slicer 2: This Time With Depth!” “Brain Slicer 2: This Time With Depth!”