presentación de powerpoint - seminario mirian andrés › wp... · seminario de informática...
TRANSCRIPT
![Page 1: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/1.jpg)
Detección automática de neuronas usando aproximaciones de Machine Learning
Seminario de Informática “Mirian Andrés”16 de mayo de 2016
Gadea Mata Martínez
![Page 2: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/2.jpg)
Rotterdam (Países Bajos)
![Page 3: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/3.jpg)
Find neurons
Recognize them
Localize them
![Page 4: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/4.jpg)
Image acquired with a Leica SP5 Confocal Microscope (20x lens).
The size is about 10,000x12,000 pixels (about 70mm2).
One mosaic contain on the orderof 40 transfected (with proteinGFP) neurons.
We have studied 8 images(mosaics).
![Page 5: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/5.jpg)
Gold-standard:
A neurobiologist manually marked all regionsthat contained neurons.
The region has a size of 500x500 pixels. For non-neuron regions randomly sampled
1,000 patches from background.
![Page 6: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/6.jpg)
Study of images
We considered two machine-learning approaches.
1. WNDCHRM which computes an extensive list of image features.Over 1,000 features were computed among them:
- different types of polynomial decompositions- high contrast features- pixel statistics- texture descriptors (*)
Not all features are equally relevant. The programme computes a Fisher score for each,which allows building a ranked preferences of features. And it only works with the top-15%features according this score.
Classification of patches was done using a weighted neighbor distances (WND) classifier(it’s a version of kNN algorithm).
(*) An image texture is a set of metrics calculated in image processing designed to quantifythe perceived texture of an image. Image texture gives us information about the spatialarrangement of color or intensities in an image or selected region of an image.Image textures can be artificially created or found in natural scenes captured in an image.Image textures are one way that can be used to help in segmentation or classification ofimages.
Rotterdam 2015
![Page 7: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/7.jpg)
Study of images
We considered two machine-learning approaches.
2. Computing the scale-invariant feature transform (SIFT)This algorithm find the keypoint descriptors of the image that are:
- highly distinctive,- invariant to image scaling, rotation, a range of affine distortions,- robust to noise and illumination change.
![Page 8: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/8.jpg)
Study of images
2. Computing the scale-invariant feature transform (SIFT)
To classify the patches based on these features we used:- a bag-of-words model- Naïve-Bayes (NB) classifier from the WEKA toolkit.
M -keypoints
One point has 128 descriptors
P -keypoints
…
128
Patch 1
Patch 2
Patch nN - patches
xx x
xx
xx x
xxx
x
Patch 1
Patch 2
Simple K Means(SKM) algorithm is
used to get a Bag of Words (with k words)
and to normalizethese data with
histograms
…
N - histograms
(UL)
Naïve-Bayesalgorithm is
used to classify
neuron
no neuron
(SL)
![Page 9: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/9.jpg)
Analysis of images:
25% 25%75% 75%
LocationJ ? Yes: TP / No: FP
training-set
testing-set
About 1000 patches per image (8 mosaics)
WNDCHRM
SIFTneuron no neuron
classification-set
Which are the true-positives?
Gold-standard
neuron no neuron
Supervised:
P: positive | TP: true positive | FP: false positive | FN: false negative | recall = TP/P | precision = TP/(TP+FP)
method recall precision
LocationJ 0.80 0.86
WNDCHRM-A 1 0.62
SIFT-A 0.74 0.46
![Page 10: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/10.jpg)
P: positive | TP: true positive | FP: false positive | FN: false negative | recall = TP/P | precision = TP/(TP+FP)
method recall precision
LocationJ 0.64 0.50
WNDCHRM-A 0.89 0.19
SIFT-A 0.60 0.26
Results:
Gold-standard
neuron no neuron
To improve performance:
WNDCHRM-A
TP
SIFT-A
TP
FP
FP
About 1000 patches per image (8 mosaics)
neuron no neuron
classification-set
method recall precision
LocationJ 0.64 0.50
WNDCHRM-A 0.89 0.19
SIFT-A 0.60 0.26
WNDCHRM-B 0.66 0.44
SIFT-B 0.94 0.57
WNDCHRM
SIFT
FP
Cross-validation25% - 75%
InternationalSymposium
on BiomedicalImaging (ISBI)
Praga 2016
![Page 11: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/11.jpg)
Rotterdam 2016:
neuron no neuron
classification-set
Gold-standard
neuron no neuron
500x500 pixels
Overlapping 50%
To consider neuron, an overlapping 50%
SETTINGS:
922
16441
409
15519
![Page 12: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/12.jpg)
Time: <1’ 1’ 3’ 5’ 5’ 6’ 29’
Over. % grid
25 50 70 75 75 80 90
Over. % pos.
25 50 70 25 75 80 90
Nº Pos patches
298 253 233 2688 236 225 221
Nº patches
750 1665 4650 6660 6660 10304 41216
Mosaic nº 6:
Settings to calculate the patches
• We work with 8 mosaics with different sizes. The time is similar for each one, except to mosaic number 4 that the time to obtain the patches is 16 minutes.
![Page 13: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/13.jpg)
• Colaboración con Carlos Fernández (Universidad de A Coruña)
Obtener las características de los patches de los mosaicos
Estudiar diferentes algoritmos de Machine Learning
Analizar los datos para ver que algoritmo es mejor
![Page 14: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/14.jpg)
neuron no neuron
WNDCHRMGet the features (1059)
SIFTNormalize the features, get histograms
Technical Problems:• Time
• WNDCHRM: all mosaics 48h
• Obtain histograms: Study a mosaic with 100 words takes +24h
m01 6’ m05 6’
m02 6’ m06 8’
m03 3’ m07 8’
m04 44’ m08 6’
![Page 15: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/15.jpg)
SIFT features:
M -keypoints
One point has 128 descriptors
P -keypoints
…
128
Patch 1
Patch 2
Patch nN - patches
xx x
xx
xx x
xxx
x
Patch 1
Patch 2
Simple K Means(SKM) algorithm is
used to get a Bag of Words (with k words)
and to normalizethese data with
histograms
…
N - histograms
(UL)
Parallelize this part This part is calculatedwith Weka
Now, it takes 17h (a mosaic (m06) with 100 words)
![Page 16: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/16.jpg)
Mosaic m05 Nº of words Time
Get histograms 20 48’
40 43’
60 56’
80 65’
Total process 80 3h
The main problem is the time.
Settings: • Size of patch• Overlap % grid• Overlap % to consider positive a patch• Nº of words
This is only with one mosaic. However we need to work with all mosaics to the same time.
The number of sift features in all mosaics is 743.907
![Page 17: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/17.jpg)
SIFT features:
M -keypoints
One point has 128 descriptors
P -keypoints
…
128
Patch 1
Patch 2
Patch nN - patches
xx x
xx
xx x
xxx
x
Patch 1
Patch 2
Simple K Means(SKM) algorithm is
used to get a Bag of Words (with k words)
and to normalizethese data with
histograms
…
N - histograms
(UL)
Parallelize this part Develop k-means algorithmand parallelize
Under construction
![Page 18: Presentación de PowerPoint - Seminario Mirian Andrés › wp... · Seminario de Informática “Mirian Andrés” ... x x x x x x x x x x x Patch 1 Patch2 Simple K Means (SKM) algorithm](https://reader033.vdocuments.us/reader033/viewer/2022053014/5f118603cfe0ab48d2088eff/html5/thumbnails/18.jpg)
Conclusion:
• Have a system to calculate this features (WNDCHRM and SIFT)• Obtain the data and study them (Carlos Fernández)• Analyze the results• Compare them with other methods
Thanks! Dank u wel! Gracias!