Automated Identification of Filaments in Cryoelectron Microscopy Images

1
Automated Identification of Filaments in
Cryo-electron Microscopy Images

• Yuanxin Zhu
• Postdoctoral Research Associate
• Imaging Technology Group, Beckman Institute
• University of Illinois at Urbana-Champaign

2
Overview

• Introduction
• Methods
• Results
• Analysis
• Summary
• Conclusions

3
1. Introduction

• Example images of specimens of tobacco mosaic
  virus (TMV) acquired (66, 000x) to a 1Kx1K CCD
  camera using cryo-EM.

near-to-focus(-0.3µm) , NTF image
further-from-focus (-3µm) , FFF image
4
1. Introduction (Contd)

• Three steps of cryo-EM.
• Preserve specimens in vitreous ice.
• Acquire 2D projection images using a transmission
  electron microscope (TEM).
• Reconstruction of 3D electron density maps from
  2D images.
• Drawbacks.
• Beam induced radiation damage ? extremely
  low-dose of electrons ? low-contrast images ?
  necessity of averaging large number of images for
  high-resolution reconstruction ? tedious and very
  time-consuming for manual data processing.

5
2. Methods
6
2.1 Whats Perceptual Organization?

• Our vision system attempts to group information
  by rules including proximity, similarity,
  collinearity, parallelism, and connectivity.
• In computer vision, perceptual organization takes
  primitive image elements and generate groupings
  that encode the structural interrelationships
  between the elements by exploiting those rules.

This figure illustrates the human ability to
spontaneously detect certain groupings from among
an otherwise random background of similar
elements. This figure contains three non-random
groupings resulting from parallelism,
collinearity, and endpoint proximity
(connectivity).Courtesy of Artificial
Intelligence, 31(3), pp. 371, March 1987.
7
2.2 Detection of filaments in FFF images

• Three-level perceptual organization algorithm
• At the signal (low) level, interesting filament
  boundary points organized into edge chains by
  filtering and exploring proximity.
• At the primitive (middle) level, edges grouped
  into line segments by exploiting collinearity.
• At the structural (high) level, Line segments
  grouped into filaments using parallelism and some
  high-level knowledge.

8
2.1.1 Edge detection

• Noise smoothing, edge enhancement, and edge
  localization.

a b
c
9

• 2.1.2 Grouping of line segments

• Hough transform makes collinear points in the X-Y
  plane map into concurrent curves in the ?-?
  parameter space.

10

• 2.1.2 Grouping of line segments (continued)

• End point detection and merging collinear line
  segments.

11
2.1.3 Detection of filaments

• A filament is an image region enclosed by two
  line segments with a particular structural
  relationship.

12
2.1.4 Separation of end-to-end joins

• Cut a detected image region out of the original
  image.
• Enhance possible end-to-end joins by filtering
  the region
• Sum along columns of the region to generate a 1D
  projection.
• Search for peaks in the 1D projection which
  indicate possible end-to-end joins.

13
2.1.4 Separation of end-to-end joins (contd)
a
c
b
14
2.2 Identification of Filaments in the NTF Image

• Alignment of the NTF image to its corresponding
  FFF image in the defocus pair.
• Extraction of filaments in the NTF image using
  coordinates that are identified in the FFF image
  and shifted according to the result of alignment.

15
2.2.1 Alignment using cross correlation

• Calculate the cross-correlation function (CCF).
• Form the cross power spectrum (the product of
  multiplying the complex conjugate of the FFT of
  the first image by the FFT of the second).
• Inverse-FFT of the cross power spectrum.
• Identify the global peak of the CCF indicating
  the relative displacement of the two images.
• Drawbacks The CCF peak shape deteriorates
  rapidly with increasing defocus difference
  between a pair of images.

16
2.2.2 Alignment using phase correlation

• Calculate the phase-correlation function (PCF).
• Form the cross power spectrum.
• Calculate the phase difference (the cross power
  spectrum divided by its modulus).
• Inverse-FFT of the phase difference.
• Identify the global peak of the PCF indicating
  the relative displacement.
• Advantages the PCF exhibits a well-shaped peak
  and its generally insensitive to narrow
  bandwidth noise and conventional image
  degradations.

17
2.2.3 Illustration of alignment

• a b
• The peak shape on the phase-correlation surface
  (a) is much more well-posed than that on the
  cross-correlation surface (b).

18
2.2.3 Illustration of alignment (continued)
After alignment of the NTF image to the FFF image
using phase correlation, total 9 filaments were
identified in the NTF image. The relative
displacement between the pair of images is (4,-7
) in pixel.
19
3. Results
20
4. Analysis

• False alarms.
• Most of those false alarms are poor quality
  filaments which are not suitable for further
  processing. However, our automated reconstruction
  process currently has the ability to reject them.
• Curved filaments.
• its straightforward to extend the approach to
  identify curved filaments if we consider curved
  filaments as piecewise straight ones.
• Subpixel (non-integer) displacement.

21
5. Summary

• An accurately approach were developed for
  identification of filamentous structures in
  cryo-EM images acquired in defocus pairs and
  evaluated by applying it to images of TMV
  specimens.
• A defocus pair consists of a NTF image captured
  first and a FFF image second.
• A three-level perceptual organization algorithm
  was proposed to detect filaments in the FFF
  image.
• Filaments in the NTF image can be reliably
  extracted through alignment of the NTF image to
  the FFF image using phase correlation technique.

22
6. Conclusions

• Achieved the immediate goal of fully automated
  filament identification in cryo-EM images.
• The automation will facilitate 3D reconstruction
  of helical objects using cryo-EM and expedite the
  development of an integrated molecular
  microscopy.
• Demonstrated that techniques from computer vision
  have proven to be effective and are opening the
  door to rapid advances in experimental structural
  biology.
• Provide valuable insight into how one might
  approach other similar problems identification of
  single particles, microtubules, etc.

23
Acknowledgements

• Bridget Carragher, Clinton Potter, and David
  Kriegman for their advising and leadership.
• Jim Pulokas and the rest of the Imaging
  Technology Group at the Beckman Institute for
  collecting the data and various assistance.
• Ron Milligan at The Scripps Research Institute
  for providing the TMV specimen.
• National Science Foundation and National
  Institute of Health for funding support.