You are on page 1of 11

Two-Dimensional

Phase Unwrapping
Using Neural Networks
Wade Schwartzkopf, Thomas E. Milner, Joydeep Ghosh,
Brian L. Evans, and Alan C. Bovik
Laboratory for Image and Video Engineering Laboratory for Image and Video Engineering
Department of Electrical and Computer Engineering Department of Electrical and Computer Engineering
The University of Texas at Austin, Austin, TX 78712-1084 The University of Texas at Austin, Austin, TX 78712-1084
2000 IEEE Southwest Symposium on
Image Analysis and Interpretation
Phase-Based Imaging Methods
Synthetic
Aperture Radar
Mountains
Magnetic
Resonance
Imaging
Knee
Optical Doppler
Tomography
Blood Vessel
Optical Doppler Tomography (ODT)
• Uses lasers to image skin tissue
• Obtains two images
– Structure: proportional to amplitude
– Velocity: proportional to phase
OCT images of
polystyrene microspheres
a b
400 µ m 400 µ m
Structure Velocity
Depth
Surface
y(n)+2π
y(n)
y(n)-2π
y(n-1)+π
y(n-1)-π
1-D Phase Unwrapping
• Use information from neighboring pixels
• Errors propagate
Original Signal
0
π

Wrapped Signal
0
π

Unwrapped Signal
0
π

0.0 0.6π
1.6π 1.2π
2-D Phase Unwrapping
• Possible to use second dimension to recover
• Residues: conflicts in different dimensions
0.0 0.6π
-0.4π 1.2π
0.0 0.6π
1.6π 1.2π
Two of the
possible orderings
Why Neural Networks?
• Fast unwrapping of ODT velocity images
– 10 100 × 100 images per second
– Global optimization methods impractical
– Fast local methods necessary
• Learn features in ODT velocity images of skin
tissue
– Sparse images
– All blood flow follows parabolic law
• Goal: Train neural networks to detect phase
jumps in ODT phase images
Neighboring pixels
Current pixel
Phase Jump Detector
• Feedforward multilayer
perceptron network
• Wrapped values of 5×5
neighborhood of pixels
as inputs to network
• Three outputs for the current
pixel
– Positive jump of 2π
– No jump of 2π
– Negative jump of 2π
• Non-iterative, pixel parallel
computation
Phase Jump Detector
• Neural network parameters
– 25 input neurons (5 × 5 image block)
– 5 hidden neurons
– 3 output neurons
• Training of neural network
– Train using 90 simulated ODT images with conjugate-gradient
method
– Validate with 10 simulated ODT images
– Simulated images have variety of vessel sizes and shapes
• Tested on 100 simulated ODT images
Simulated Image
Simulated flow in a blood
vessel without wrapping
Simulated flow in a blood
vessel after wrapping
Phase Jump Detection
Location of positive (black) and
negative (white) phase jumps
Pixels in the range [π, 3π)
• Unwrap along columns
– Top pixel is always zero since no movement at edge of skin
– At positive (negative) jump, add (subtract) 2π to rest of column
Confusion Matrix
Classified
Positive No Jump Negative
Positive 0.89 0.11 0
No Jump 0.0004 0.9988 0.0008
A
c
t
u
a
l
Negative 0 0.09 0.91

Phase-Based Imaging Methods Synthetic Aperture Radar Magnetic Resonance Imaging Optical Doppler Tomography Mountains Knee Blood Vessel .

Optical Doppler Tomography (ODT) • Uses lasers to image skin tissue • Obtains two images – Structure: proportional to amplitude – Velocity: proportional to phase Surface a b Structure Velocity Depth 400 µm 400 µm OCT images of polystyrene microspheres .

1-D Phase Unwrapping • Use information from neighboring pixels y(n)+2π y(n-1)+π y(n) y(n-1)-π y(n)-2π • Errors propagate Original Signal π 0 -π π 0 -π π 0 Wrapped Signal 3π Unwrapped Signal .

2π .6π -0.4π 1.6π 1.0 0.0 0.2-D Phase Unwrapping • Possible to use second dimension to recover • Residues: conflicts in different dimensions 0.6π 1.2π Two of the possible orderings 0.6π 1.6π 1.0 0.2π 0.

Why Neural Networks? • Fast unwrapping of ODT velocity images – 10 100 × 100 images per second – Global optimization methods impractical – Fast local methods necessary • Learn features in ODT velocity images of skin tissue – Sparse images – All blood flow follows parabolic law • Goal: Train neural networks to detect phase jumps in ODT phase images .

pixel parallel computation .Phase Jump Detector • Feedforward multilayer perceptron network • Wrapped values of 5×5 neighborhood of pixels as inputs to network • Three outputs for the current pixel – Positive jump of 2π – No jump of 2π – Negative jump of 2π Neighboring pixels Current pixel • Non-iterative.

Phase Jump Detector • Neural network parameters – 25 input neurons (5 × 5 image block) – 5 hidden neurons – 3 output neurons • Training of neural network – Train using 90 simulated ODT images with conjugate-gradient method – Validate with 10 simulated ODT images – Simulated images have variety of vessel sizes and shapes • Tested on 100 simulated ODT images .

Simulated Image Simulated flow in a blood vessel without wrapping Simulated flow in a blood vessel after wrapping .

add (subtract) 2π to rest of column Location of positive (black) and negative (white) phase jumps Pixels in the range [π.Phase Jump Detection • Unwrap along columns – Top pixel is always zero since no movement at edge of skin – At positive (negative) jump. 3π) .

91 No Jump 0.89 0.9988 0.0004 Negative 0 .11 0.0008 0.09 0 0.Confusion Matrix Classified Positive No Jump Negative A c t u a l Positive 0.