Advanced PhaseBased Segmentation of Multiple Cells from Brightfield Microscopy Images

1
Advanced Phase-Based Segmentation of Multiple
Cells from Brightfield Microscopy Images Rehan
Ali Wolfson Medical Vision Laboratory, Department
of Engineering Science, University of Oxford
2
Overview

• The case for brightfield image processing
• Considering different forms of spatially variant
  phase
• Physical Phase
• Image / Local Phase
• Our segmentation framework
• Results
• Limitations and future extensions

3
Anatomical vs Functional Imaging

• Medical image processing relies on multi-modal
  approach for jointly analysing anatomic and
  functional data
• This is rare in microscopy image processing! Most
  work done directly on fluorescence images
• Brightfield offers detailed anatomic information
  but suffers from poor contrast
• Accurate boundary delineation for quantitative
  analysis of intracellular fluorescent data is
  best done on anatomic data

CT (left) and PET (right) images from
www.dana-farber.org
Three cells in a brightfield (left) and
fluorescence (right) image
4
Improving Contrast

• Manual defocus
• Most common as non-enhanced brightfield is most
  widely available form of microscope
• Hardware Enhancement
• Zernike Phase Contrast
• Differential Interference Contrast
• Software Enhancement
• Phase Recovery

Digital phase recovery images from
www.iatia.com.au
Image from www.microscopyu.com
Image from www.microscopyu.com
5
Effects of Physical Phase when Defocusing
?z - 2µm
?z - 8µm
?z -100µm
?z 0µmIn Focus
?z 2µm
?z 8µm
?z 100µm
6
Defocusing Theory

• Transport of Intensity equation can be derived
  from the time-independent wave equation
• Defines laplacian relation between z-axis
  intensity derivative and physical phase

7
Phase Recovery

• Quantitative Phase Microscopy (QPM, Iatia)
  proposes a FFT-based solution to TIE, to recover
  phase (Barty et al, 1998)
• Problems
• Low frequency artefacts
• May not recover very thin cell body regions

Brightfield Image
QPM result
8
Image / Local Phase

• Intensity-independent feature detector used in
  medical imaging
• Basis is the analytic signal used in 1D signal
  processing
• Allows computation of odd and even quadrature
  filters
• These respond to odd and even features in
  bandpassed signals
• Local phase is the ratio of filter response to
  feature symmetry? describes type of feature
• Local energy is the magnitude of filter responses
  ? describes strength of feature
• Monogenic signal representation extends this to
  nD by allowing computation of nD-specific QFs

Odd
Even
Blue signal after DoG bandpass filtering Red
signal after filtering with HT of DoG
9
Monogenic Signal
Quadrature Filters Odd h1(x,y) Ib(x,y) h2(x,y)
Ib(x,y)
Bandpass Filter Even Ib(x,y) I(x,y)b(x,y)
Not useful in low contrast cases
Diff of Gaussians Bandpass, Spatial
Domain Normally zero-DC
10
Monogenic Signal Results w Bandpass Filter
?I/ ?z
Local Phase ? -p to p
Local Orientation F -p to p (points in
direction of max signal variance)
11
Results with Non-Zero DC Lowpass Filter
?I/ ?z
Local Phase ? -p to p
Local Orientation F -p to p (points in
direction of max signal variance)
12
Level Set

• Use narrowband Level Set framework for multiple
  object segmentation
• Use local phase and local orientation based terms
  in speed function

13
Local Orientation Term
14
Segmentation Overview
Initialisation
Inputs
I(x,y,dz)
Local Phase
Initial Class Map
?I/ ?z
I(x,y,-dz)
Local Orient
User clicks on cells
Physical Phase
Requires motorised z-axis stage to capture /- z
images Cells must be fixed / stationary during
defocus image acquisition (e.g. adherent)
Monogenic Signal
Split up cells(region growing)
Level Set
15
Validation vs Manual Segmentation
AlgorithmManual
16
Current Issues

• Automatic Initialisation required for HT analysis
• Use a strongly defocused image
• Use a specific fluorescence image
• Use texture from in-focus image
• Method struggles with ultra-thin cell body
  regions
• Parameter tuning required to obtain good results

17
Summary

• Segmentation of anatomical cell features is
  ideally done on an anatomical image (e.g.
  brightfield)
• Information can be extracted from difficult
  images using one or combination of techniques
• Physical phase models of image formation
• Image signal analysis with local phase
• Code now available for Windows / Linux at
  www.sephace.com

18
www.sephace.com
19
Acknowledgements

• Wolfson Medical Vision Labs
• Professor Sir Michael Brady FRS FREng
• Dr Mark Gooding
• Dr Weiwei Zhang
• Nathan Cahill
• Gray Cancer Institute
• Dr Martin Christlieb
• Life Sciences Interface Doctoral Training
  Centre,University of Oxford
• EPSRC UK for funding