ECCO.
An Electron Counting Implementation for
Image Compression and Optimization of
Cryo-EM Data Management.
December 18th, 2019
Room 2.1.5, Politecnico di Milano
Candidate:
Chiara Coletti, matr. 897815
Supervisor:
Prof. Marco Domenico Santambrogio
Co-supervisors:
Prof. Sara Vinco, Politecnico di Torino
Dr. Eleonora D'Arnese
INTRODUCTION
ECCO. An Electron Counting Implementation
for Image Compression and Optimization of
Cryo-EM Data Management.
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 8978153
Cryogenic Electron Microscopy
17 Å
Recorded on film
[Fernández, 2007]
Revolution in cryo-EM biological structure determination resolution:
example of RNA Polymerase III.
2007 2010 2015 2015
Atomic model
Direct e- detectors
[Hoffmann, 2015]
10 Å
Recorded on film
[Fernández, 2010]
3.9 Å
Direct e- detectors
[Hoffmann, 2015]
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 897815
Image Processing Framework
4
Cryogenic electron
microscope
Supercomputing
facility
…
Raw data stream
400 Gb/s network connection
…
4D camera
87000 fps

576 x 576 pixels
4D camera
75000 fps limit

576 x 576 pixels
Noisy data with
very low SNR+
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 897815
Edge Computing Approach
5
ECCO
Raw images
Processed
images
460 Gb/s 8 Gb/s
Cryogenic electron
microscope
Supercomputing
facility
Data compression
before network
transfer
Noise rejection
before further
processing+
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 8978156
Electron Counting
10.4
0.2
0.2
0.1
0.07
0.03
Electron
hits detector
Energy
scattering
Charge
integration
Electron
counting
Improved DQE across all spatial frequencies and 55:1 data compression
DQE = detective quantum efficiency [Booth, 2013]
1 2 3 4
PROBLEM
DEFINITION
ECCO. An Electron Counting Implementation
for Image Compression and Optimization of
Cryo-EM Data Management.
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 8978158
Problem Statement
Counting
efficiency
Super
resolution
Collision
loss reduction
Real-time
implementation
Maximizing true
counts, while
minimizing false
ones.
Improving
precision in
events
localization.
Avoiding the
fusion of
overlapping
electrons.
Considering the
experimental
setting time
constraints.
MATERIALS
AND METHODS
ECCO. An Electron Counting Implementation
for Image Compression and Optimization of
Cryo-EM Data Management.
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781510
Experimental Dataset
2k x 2k pixels images from a 4D Scanning
Transmission Electron Microscope (4D
STEM), acquired under vacuum probe
conditions with K2 direct detection
camera at 400fps.
No ground truth available, electrons
energy deposition is a stochastic
phenomena.
Example of a single noisy
diffraction pattern.
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 897815
Example of a single counted
diffraction pattern.
TP
FP
11
Evaluation Methods
The state-of-the-art implementation is
included in the py4DSTEM tool [https://
github.com/bsavitzky/py4DSTEM].
From the counted diffraction patterns:
- TP : electrons counted inside the
diffraction disk;
- FP : electrons counted outside the
diffraction disk.
precision =
TP
TP + FP
PROPOSED
SOLUTIONS
ECCO. An Electron Counting Implementation
for Image Compression and Optimization of
Cryo-EM Data Management.
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781513
Solution: Denoising
Histogram of diffraction pattern intensities
Density
Pixels intensity values
Gaussian noise
threshold = μ+4σ
The threshold to remove Gaussian noise is chosen as μ+4σ, assuming a
clear distinction between signal and noise distributions.
Gaussian noise
distribution parameters:
μ = 16.60
σ = 2.80
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781514
Solution: Denoising
Histogram of denoised diffraction pattern intensities
Density
Pixels intensity values
Gaussian noise
threshold = μ+4σ
The introduction of a denoising step enhances the signal with respect to
the noise, that is shrunk on a more narrow interval of intensities.
Gaussian noise
distribution parameters:
μ = 16.60
σ = 2.80
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781515
Denoising Algorithms
Approaches for
image denoising
Unsupervised
algorithms
Supervised deep
learning
Self-supervised deep
learning [Batson, 2019]
Block Matching
3D
Non-Local
Means
Denoising
CNN
Baby
UNet
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781516
Denoising Algorithms
Approaches for
image denoising
Unsupervised
algorithms
Supervised deep
learning
Block Matching
3D
Non-Local
Means
Denoising
CNN
Baby
UNet
Self-supervised deep
learning [Batson, 2019]
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781517
Denoising Algorithms
Approaches for
image denoising
Unsupervised
algorithms
Supervised deep
learning
Block Matching
3D
Non-Local
Means
Denoising
CNN
Baby
UNet
Self-supervised deep
learning [Batson, 2019]
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781518
Denoising Algorithms
Approaches for
image denoising
Unsupervised
algorithms
Supervised deep
learning
Block Matching
3D
Non-Local
Means
Denoising
CNN
Baby
UNet
Self-supervised deep
learning [Batson, 2019]
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781519
Solution: Super-Resolution
2
Thresholded
diffraction pattern
Neighbourhoods
identification
Event localisation in
maximal point
1 3
Py4DSTEM implementation localizes the electron event in the maximal point of
a neighbourhood of non-zero pixels, keeping the original spatial resolution.
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 897815
2 3
20
Solution: Super-Resolution
Thresholded
diffraction pattern
Connected pixels
identification
Event localisation at
center of mass
1
ECCO implementation localizes the electron event in the center of mass of a set
of connected non-zero pixels, eventually doubling (or more) the original spatial
resolution.
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781521
Solution: Collision Loss Reduction
2e-
1e-
N°ofevents
Events intensities
3e- …
Histogram of events intensities after counting
The counts' intensities follow the theoretical Landau distribution, defining
several peaks that can be used to identify overlapping electrons and reduce
collision losses.
ECCO. An Electron Counting Implementation
for Image Compression and Optimization of
Cryo-EM Data Management.
RESULTS AND
CONCLUSIONS
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781523
Results: Denoising
The aim is the optimization of the precision, balancing TP and FP,
while keeping a reasonably low execution time for real-time implementation.
NLM BM3D DnCNN Baby UNet
py4D
STEM
- - Superv. Self-sup. Superv. Self-sup. -
Precision [%] 96.75 97.19 93.45 91.86 49.44 19.74 89.64
TP 744 968 599 553 88 30 727
FP 25 28 42 49 90 122 84
Run time [s] 3.98 192.15 9.71 8.59 55.68 49.53 0.40
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781524
Results: GPU Acceleration
284x 94x
GPU NLM
parallel algorithm
acceleration factor
with respect to CPU
GPU DnCNN
testing phase
acceleration factor
with respect to CPU
20x
GPU DnCNN
training phase
acceleration factor
with respect to CPU
GPU execution time:
0.01s per frame
GPU execution time:
0.11s per frame
GPU execution time:
20.83s per frame
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781525
Contributions
The introduction
of NLM for noise
reduction
increases the
precision of
7.11%.
The innovative
events
localization
allows sub-pixel
accuracy in
counting.
The novel
classification
approach
increases images
information
content.
The 284x speed-
up with GPU
implementation
shows potential
for real-time
execution.
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781526
Future Works
FPGA
Extension to different
datasets to test the
solution's robustness.
Implementation of new
denoising algorithms for
alternative solutions.
Design of hardware
implementations on
different platforms.
Thank you for the attention.
Candidate:
Chiara Coletti, matr. 897815
ECCO: An Electron Counting Implementation
for Image Compression and Optimization of
Cryo-EM Data Management.
Questions?
Supporting Materials
Candidate:
Chiara Coletti, matr. 897815
ECCO: An Electron Counting Implementation
for Image Compression and Optimization of
Cryo-EM Data Management.
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781529
Py4DSTEM Implementation
Dark reference
subtraction
Gaussian noise
thresholding
X-rays
thresholding
NN maximal
point selection
-Median filter
along strikes
direction
-Filtered pattern
subtraction from
original
-Fitting image 

to a gaussian
distribution:
N(μg,σg2)
-Thresholding
image at μg+kσg
-Computing μ and
σ2 of the image
-Thresholding at
μ+hσ (where h>>k)
-Consider 8x8
neighbourhoods in
binarized image
-Put e- event at
maximal point
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781530
Our Counting Implementation
Dark
reference
subtraction
Gaussian
noise
thresholding
Connected
pixel CoM
-Median filter
along strikes
direction
-Filtered pattern
subtraction
from original
-Fitting image 

to a gaussian
distribution:
N(μg,σg2)
-Thresholding
image at μg+kσg
-Finding
connected
components
-Assigning
integral signal
to their CoM
Image
denoising
-Enhancement
with different
approaches:
- NLM
- BM3D
- DnCNN
Events
classification
-Recognizing
events’ overlap
-Possibility to
assign
probabilistic
labels
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781531
Non Local Means
Variance law in probability theory demonstrates that if pixels are averaged, the noise
standard deviation of the average is divided by ⟹ we replace the intensity of a pixel with
an average of the intensities of “similar” pixels [Buades, 2011].
n
n
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781532
Block Matching 3D
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781533
Denoising CNN
Noisy pattern Noise estimate
Convolutional
Convolutional
Convolutional
ReLU
ReLU
BatchNorm
128 x 128 128 x 128128 x 128 x 64 128 x 128 x 64 128 x 128 x 64
• Convolution + ReLU to gradually separate image structure from noisy
observation;
• Batch normalisation to speed up learning and improve performances;
• Residual learning to learn residuals of the image, easier than learning
the denoised image representation.
[Zhang, 2017]
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 897815
2 x 642 x 32 642 x 32
1282
34
Unet
1282 x 16
Input image
642 x 16 642 x 32
642 x 32 642 x 32
3 x 1282 x 16 1282 x 16 1282
Output image
Legend:
Convolutional
Pooling
Upsampling
Concatenation
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781535
Results: Denoising
Original DP BM3D outputNLM output
DnCNN output UNet outputNLM thresholded DP
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781536
Results: Classification
Binary Discrete Probabilistic py4DSTEM
Precision [%] 96.75 93.92 96.16 89.64
TP 744 1375 3975 727
FP 25 89 159 84
1
1
1 1
2
3 0.8
1.7
3.2
Binary
classification
Discrete
classification
Probabilistic
classification
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781537
Results: GPU Acceleration
NLM DnCNN py4DSTEM
- Testing Training -
CPU run time [s] 3.98 10.44 417.52 0.40
GPU run time [s] 0.01 0.11 20.83 -
284x
94xGPU NLM
acceleration factor
GPU DnCNN
acceleration factor
ECCO - Master Thesis Dissertation Candidate: Chiara Coletti, 89781538
Bibliography
Fernández-Tornero, C., Böttcher, B., Riva, M., Carles, C., Steuerwald, U., Ruigrok, R. W., ... &
Schoehn, G. (2007). Insights into transcription initiation and termination from the electron
microscopy structure of yeast RNA polymerase III. Molecular cell, 25(6), 813-823.
Fernández‐Tornero, C., Böttcher, B., Rashid, U. J., Steuerwald, U., Flörchinger, B., Devos, D. P., ... &
Müller, C. W. (2010). Conformational flexibility of RNA polymerase III during transcriptional
elongation. The EMBO journal, 29(22), 3762-3772.
Hoffmann, N. A., Jakobi, A. J., Moreno-Morcillo, M., Glatt, S., Kosinski, J., Hagen, W. J., ... & Müller, C.
W. (2015). Molecular structures of unbound and transcribing RNA polymerase
III. Nature, 528(7581), 231.
Booth, C., Mooney, P. (2013), Applications of electron-counting direct-detection cameras in high-
resolution cryo-electron microscopy. Microscopy and Analysis, 27(6):13-21 (AM).
Zhang, K., Zuo, W., Chen, Y., Meng, D., & Zhang, L. (2017). Beyond a gaussian denoiser: Residual
learning of deep cnn for image denoising. IEEE Transactions on Image Processing, 26(7),
3142-3155.
Batson, J., & Royer, L. (2019). Noise2self: Blind denoising by self-supervision. arXiv preprint arXiv:
1901.11365.

ECCO: An Electron Counting Implementation for Image Compression and Optimization of Cryo-EM Data Management

  • 1.
    ECCO. An Electron CountingImplementation for Image Compression and Optimization of Cryo-EM Data Management. December 18th, 2019 Room 2.1.5, Politecnico di Milano Candidate: Chiara Coletti, matr. 897815 Supervisor: Prof. Marco Domenico Santambrogio Co-supervisors: Prof. Sara Vinco, Politecnico di Torino Dr. Eleonora D'Arnese
  • 2.
    INTRODUCTION ECCO. An ElectronCounting Implementation for Image Compression and Optimization of Cryo-EM Data Management.
  • 3.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 8978153 Cryogenic Electron Microscopy 17 Å Recorded on film [Fernández, 2007] Revolution in cryo-EM biological structure determination resolution: example of RNA Polymerase III. 2007 2010 2015 2015 Atomic model Direct e- detectors [Hoffmann, 2015] 10 Å Recorded on film [Fernández, 2010] 3.9 Å Direct e- detectors [Hoffmann, 2015]
  • 4.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 897815 Image Processing Framework 4 Cryogenic electron microscope Supercomputing facility … Raw data stream 400 Gb/s network connection … 4D camera 87000 fps
 576 x 576 pixels 4D camera 75000 fps limit
 576 x 576 pixels Noisy data with very low SNR+
  • 5.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 897815 Edge Computing Approach 5 ECCO Raw images Processed images 460 Gb/s 8 Gb/s Cryogenic electron microscope Supercomputing facility Data compression before network transfer Noise rejection before further processing+
  • 6.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 8978156 Electron Counting 10.4 0.2 0.2 0.1 0.07 0.03 Electron hits detector Energy scattering Charge integration Electron counting Improved DQE across all spatial frequencies and 55:1 data compression DQE = detective quantum efficiency [Booth, 2013] 1 2 3 4
  • 7.
    PROBLEM DEFINITION ECCO. An ElectronCounting Implementation for Image Compression and Optimization of Cryo-EM Data Management.
  • 8.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 8978158 Problem Statement Counting efficiency Super resolution Collision loss reduction Real-time implementation Maximizing true counts, while minimizing false ones. Improving precision in events localization. Avoiding the fusion of overlapping electrons. Considering the experimental setting time constraints.
  • 9.
    MATERIALS AND METHODS ECCO. AnElectron Counting Implementation for Image Compression and Optimization of Cryo-EM Data Management.
  • 10.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781510 Experimental Dataset 2k x 2k pixels images from a 4D Scanning Transmission Electron Microscope (4D STEM), acquired under vacuum probe conditions with K2 direct detection camera at 400fps. No ground truth available, electrons energy deposition is a stochastic phenomena. Example of a single noisy diffraction pattern.
  • 11.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 897815 Example of a single counted diffraction pattern. TP FP 11 Evaluation Methods The state-of-the-art implementation is included in the py4DSTEM tool [https:// github.com/bsavitzky/py4DSTEM]. From the counted diffraction patterns: - TP : electrons counted inside the diffraction disk; - FP : electrons counted outside the diffraction disk. precision = TP TP + FP
  • 12.
    PROPOSED SOLUTIONS ECCO. An ElectronCounting Implementation for Image Compression and Optimization of Cryo-EM Data Management.
  • 13.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781513 Solution: Denoising Histogram of diffraction pattern intensities Density Pixels intensity values Gaussian noise threshold = μ+4σ The threshold to remove Gaussian noise is chosen as μ+4σ, assuming a clear distinction between signal and noise distributions. Gaussian noise distribution parameters: μ = 16.60 σ = 2.80
  • 14.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781514 Solution: Denoising Histogram of denoised diffraction pattern intensities Density Pixels intensity values Gaussian noise threshold = μ+4σ The introduction of a denoising step enhances the signal with respect to the noise, that is shrunk on a more narrow interval of intensities. Gaussian noise distribution parameters: μ = 16.60 σ = 2.80
  • 15.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781515 Denoising Algorithms Approaches for image denoising Unsupervised algorithms Supervised deep learning Self-supervised deep learning [Batson, 2019] Block Matching 3D Non-Local Means Denoising CNN Baby UNet
  • 16.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781516 Denoising Algorithms Approaches for image denoising Unsupervised algorithms Supervised deep learning Block Matching 3D Non-Local Means Denoising CNN Baby UNet Self-supervised deep learning [Batson, 2019]
  • 17.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781517 Denoising Algorithms Approaches for image denoising Unsupervised algorithms Supervised deep learning Block Matching 3D Non-Local Means Denoising CNN Baby UNet Self-supervised deep learning [Batson, 2019]
  • 18.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781518 Denoising Algorithms Approaches for image denoising Unsupervised algorithms Supervised deep learning Block Matching 3D Non-Local Means Denoising CNN Baby UNet Self-supervised deep learning [Batson, 2019]
  • 19.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781519 Solution: Super-Resolution 2 Thresholded diffraction pattern Neighbourhoods identification Event localisation in maximal point 1 3 Py4DSTEM implementation localizes the electron event in the maximal point of a neighbourhood of non-zero pixels, keeping the original spatial resolution.
  • 20.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 897815 2 3 20 Solution: Super-Resolution Thresholded diffraction pattern Connected pixels identification Event localisation at center of mass 1 ECCO implementation localizes the electron event in the center of mass of a set of connected non-zero pixels, eventually doubling (or more) the original spatial resolution.
  • 21.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781521 Solution: Collision Loss Reduction 2e- 1e- N°ofevents Events intensities 3e- … Histogram of events intensities after counting The counts' intensities follow the theoretical Landau distribution, defining several peaks that can be used to identify overlapping electrons and reduce collision losses.
  • 22.
    ECCO. An ElectronCounting Implementation for Image Compression and Optimization of Cryo-EM Data Management. RESULTS AND CONCLUSIONS
  • 23.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781523 Results: Denoising The aim is the optimization of the precision, balancing TP and FP, while keeping a reasonably low execution time for real-time implementation. NLM BM3D DnCNN Baby UNet py4D STEM - - Superv. Self-sup. Superv. Self-sup. - Precision [%] 96.75 97.19 93.45 91.86 49.44 19.74 89.64 TP 744 968 599 553 88 30 727 FP 25 28 42 49 90 122 84 Run time [s] 3.98 192.15 9.71 8.59 55.68 49.53 0.40
  • 24.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781524 Results: GPU Acceleration 284x 94x GPU NLM parallel algorithm acceleration factor with respect to CPU GPU DnCNN testing phase acceleration factor with respect to CPU 20x GPU DnCNN training phase acceleration factor with respect to CPU GPU execution time: 0.01s per frame GPU execution time: 0.11s per frame GPU execution time: 20.83s per frame
  • 25.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781525 Contributions The introduction of NLM for noise reduction increases the precision of 7.11%. The innovative events localization allows sub-pixel accuracy in counting. The novel classification approach increases images information content. The 284x speed- up with GPU implementation shows potential for real-time execution.
  • 26.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781526 Future Works FPGA Extension to different datasets to test the solution's robustness. Implementation of new denoising algorithms for alternative solutions. Design of hardware implementations on different platforms.
  • 27.
    Thank you forthe attention. Candidate: Chiara Coletti, matr. 897815 ECCO: An Electron Counting Implementation for Image Compression and Optimization of Cryo-EM Data Management. Questions?
  • 28.
    Supporting Materials Candidate: Chiara Coletti,matr. 897815 ECCO: An Electron Counting Implementation for Image Compression and Optimization of Cryo-EM Data Management.
  • 29.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781529 Py4DSTEM Implementation Dark reference subtraction Gaussian noise thresholding X-rays thresholding NN maximal point selection -Median filter along strikes direction -Filtered pattern subtraction from original -Fitting image 
 to a gaussian distribution: N(μg,σg2) -Thresholding image at μg+kσg -Computing μ and σ2 of the image -Thresholding at μ+hσ (where h>>k) -Consider 8x8 neighbourhoods in binarized image -Put e- event at maximal point
  • 30.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781530 Our Counting Implementation Dark reference subtraction Gaussian noise thresholding Connected pixel CoM -Median filter along strikes direction -Filtered pattern subtraction from original -Fitting image 
 to a gaussian distribution: N(μg,σg2) -Thresholding image at μg+kσg -Finding connected components -Assigning integral signal to their CoM Image denoising -Enhancement with different approaches: - NLM - BM3D - DnCNN Events classification -Recognizing events’ overlap -Possibility to assign probabilistic labels
  • 31.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781531 Non Local Means Variance law in probability theory demonstrates that if pixels are averaged, the noise standard deviation of the average is divided by ⟹ we replace the intensity of a pixel with an average of the intensities of “similar” pixels [Buades, 2011]. n n
  • 32.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781532 Block Matching 3D
  • 33.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781533 Denoising CNN Noisy pattern Noise estimate Convolutional Convolutional Convolutional ReLU ReLU BatchNorm 128 x 128 128 x 128128 x 128 x 64 128 x 128 x 64 128 x 128 x 64 • Convolution + ReLU to gradually separate image structure from noisy observation; • Batch normalisation to speed up learning and improve performances; • Residual learning to learn residuals of the image, easier than learning the denoised image representation. [Zhang, 2017]
  • 34.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 897815 2 x 642 x 32 642 x 32 1282 34 Unet 1282 x 16 Input image 642 x 16 642 x 32 642 x 32 642 x 32 3 x 1282 x 16 1282 x 16 1282 Output image Legend: Convolutional Pooling Upsampling Concatenation
  • 35.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781535 Results: Denoising Original DP BM3D outputNLM output DnCNN output UNet outputNLM thresholded DP
  • 36.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781536 Results: Classification Binary Discrete Probabilistic py4DSTEM Precision [%] 96.75 93.92 96.16 89.64 TP 744 1375 3975 727 FP 25 89 159 84 1 1 1 1 2 3 0.8 1.7 3.2 Binary classification Discrete classification Probabilistic classification
  • 37.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781537 Results: GPU Acceleration NLM DnCNN py4DSTEM - Testing Training - CPU run time [s] 3.98 10.44 417.52 0.40 GPU run time [s] 0.01 0.11 20.83 - 284x 94xGPU NLM acceleration factor GPU DnCNN acceleration factor
  • 38.
    ECCO - MasterThesis Dissertation Candidate: Chiara Coletti, 89781538 Bibliography Fernández-Tornero, C., Böttcher, B., Riva, M., Carles, C., Steuerwald, U., Ruigrok, R. W., ... & Schoehn, G. (2007). Insights into transcription initiation and termination from the electron microscopy structure of yeast RNA polymerase III. Molecular cell, 25(6), 813-823. Fernández‐Tornero, C., Böttcher, B., Rashid, U. J., Steuerwald, U., Flörchinger, B., Devos, D. P., ... & Müller, C. W. (2010). Conformational flexibility of RNA polymerase III during transcriptional elongation. The EMBO journal, 29(22), 3762-3772. Hoffmann, N. A., Jakobi, A. J., Moreno-Morcillo, M., Glatt, S., Kosinski, J., Hagen, W. J., ... & Müller, C. W. (2015). Molecular structures of unbound and transcribing RNA polymerase III. Nature, 528(7581), 231. Booth, C., Mooney, P. (2013), Applications of electron-counting direct-detection cameras in high- resolution cryo-electron microscopy. Microscopy and Analysis, 27(6):13-21 (AM). Zhang, K., Zuo, W., Chen, Y., Meng, D., & Zhang, L. (2017). Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Transactions on Image Processing, 26(7), 3142-3155. Batson, J., & Royer, L. (2019). Noise2self: Blind denoising by self-supervision. arXiv preprint arXiv: 1901.11365.